TRADEMARK CONFUSION REVEALED:

AN EMPIRICAL ANALYSIS

DARYL LIM*

The likelihood of confusion standard defines the scope of trademark

infringement. Likelihood of confusion examines whether there is a substantial
risk that consumers will be confused as to the source, identity, sponsorship, or
origin of the defendants’ goods or services. This Article presents a contemporary
empirical analysis of the various factors and how they interact. Conventional
wisdom teaches us that courts should comprehensively traverse each factor and
that likelihood of confusion cases generally require jury determination. However,
the data reveals that neither is true. Instead, courts provide early off-ramps to
litigants by “economizing,” and analyzing only a handful of factors or by
“folding” factors within each other. The findings also reveal (1) which forums
are pro-defendant and which are pro-plaintiff; (2) the impact of rivalry and fair
use on outcomes; and (3) an apparent Ninth Circuit dominance.
What constitutes “confusion” remains highly subjective and difficult to
evaluate. Proxies like intent, survey evidence, mark strength, and consumer
sophistication fail to incorporate real-world purchasing conditions or are better
considered within omnibus factors. In contrast, actual confusion, mark
similarity, and competitive proximity provide judges with a potent trio of factors
to guide the infringement inquiry. Together with safe harbors for descriptive and
expressive uses, these rules of thumb enable courts to resolve trademark disputes
more coherently, consistently, and expeditiously. This Article concludes with a
blueprint of how these rules of thumb complement artificial intelligence systems

* Professor of Law & Director, Center for Intellectual Property, Information and
Privacy Law, University of Illinois Chicago School of Law. I thank Professor Y. Samuel Wang
from the Department of Statistics and Data Sciences at Cornell University for his valuable
advice on the statistical aspects of this Article. My sincere thanks to Margaret Smiley
Chavez, Steve Fisher, Annemarie Gregoire, Sarah Hampton, Nicole Robinson, Kelly
Welsh, and their colleagues at the American University Law Review who contributed to
this Article. Their courtesy, professionalism, careful editorial work, and thoughtful
comments are exemplary.

1285
1286 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

and how those systems can use empirical studies as training data to inform
future likelihood of confusion analyses.

TABLE OF CONTENTS
Introduction .............................................................................. 1286
I. Studying Confusion ....................................................... 1290
A. Methodology ............................................................ 1291
B. Blends, Triggers, and Polaroid Factors .................... 1296
1. Blends and triggers ............................................ 1297
2. The Polaroid factors ............................................ 1300
C. Missing the Point on Consumer Confusion ........... 1306
1. Intent .................................................................. 1306
2. Surveys................................................................. 1312
3. Mark strength ..................................................... 1316
4. Consumer sophistication ................................... 1319
D. Coherence-Based Reasoning................................... 1320
II. Rules of Thumb ............................................................ 1324
A. Actual Confusion...................................................... 1324
B. Mark Similarity ......................................................... 1328
C. Competitive Proximity ............................................. 1332
D. Summing It Up ........................................................ 1338
E. A Word on Fair Use.................................................. 1340
III. Observations and Implications ................................... 1345
A. Factor Folding .......................................................... 1345
B. Early Off-Ramps ....................................................... 1347
C. Deploying Artificial Intelligence ............................. 1353
1. Conception ......................................................... 1354
2. Execution ............................................................ 1355
3. Limitations .......................................................... 1358
a. “Garbage-in, garbage-out” ........................... 1358
b. Biases ............................................................ 1359
c. Contextualizing the purchasing
conditions ..................................................... 1362
d. Coding challenges ....................................... 1364
Conclusion ................................................................................. 1365

INTRODUCTION
Consumers rely on a consistent commercial lexicon to reduce
mental costs associated with purchasing decisions and in turn
commercial enterprises gain an incentive to invest in quality products
2022] TRADEMARK CONFUSION REVEALED 1287

and services.1 Businesses imbue words, symbols, scents, and sounds

with information about their goods and services.2 In turn, consumers
rely on this imbued information to navigate their decisions by making
snap judgments about the price and quality of products or services they
purchase.3 Thus, electric vehicle enthusiasts may seek out Tesla’s
stylized “T” comprising a stator and rotor, and connoisseurs of Chick-fil-A’s
chicken sandwiches will scout for its distinctive red-on-white text.
When trademark owners assert their rights, courts apply the
likelihood of confusion standard, which seeks to determine whether
the defendants’ use of a similar mark is likely to cause consumer
confusion. The likelihood of confusion standard is the linchpin of
trademark infringement.4
Unfortunately, what constitutes “confusion” remains highly
subjective and difficult to evaluate.5 Additionally, the likelihood of
confusion standard remains poorly theorized and judges applying the
standard often fail to adequately explain their analyses in their
opinions in a way that future courts can easily apply and replicate.6
When defendants counterfeit the trademark outright, liability is
clear. Literal infringement has occurred. However, trademarks protect
their owners beyond literal infringement like patents and copyrights.7
Nonliteral infringement can occur when, for example, there is a
colorable difference in the marks. This exposes parties to uncomfortably

1. Qualitex Co. v. Jacobson Prods. Co., 514 U.S. 159, 163–64 (1995)
(“[T]rademark law . . . reduces the customer’s costs of shopping and making
purchasing decisions, . . . for it quickly and easily assures a potential customer that this
item—the item with this mark—is made by the same producer as other similarly
marked items that he or she liked (or disliked) in the past.”).
2. See 1 J. THOMAS MCCARTHY, MCCARTHY ON TRADEMARKS AND UNFAIR
COMPETITION § 3, at 1 (5th ed., 2010).
3. Qualitex Co., 514 U.S. at 163–64.
4. See infra, Part I. The Lanham Act prohibits the use of a registered mark in a
manner “likely to cause confusion,” 15 U.S.C. § 1114(1)(a), as well as the use of any
term or name in a manner “likely to cause confusion” about the affiliation of the user
with another person. Id. § 1125(a)(1); see also Mil. Ord. of Purple Heart Serv. Found.,
Inc. v. Mil. Ord. of Purple Heart of U.S., Inc., 852 F. App’x 6, 9 (D.C. Cir. 2021).
5. See infra Part I.
6. See infra Part II; see also Barton Beebe, An Empirical Study of the Multifactor Tests
for Trademark Infringement, 94 CALIF. L. REV. 1581, 1582 (2006) (“Its current condition
is Babelian.”).
7. See generally Daryl Lim, Judging Equivalents, 36 SANTA CLARA HIGH TECH. L.J. 223
(2020) (tracing the origins of the doctrine of equivalents and explaining the rationale
behind the doctrine—to protect intellectual property owners from infringers seeking
to evade liability by making insubstantial changes) [hereinafter Lim, Judging
Equivalents].
1288 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

uncertain waters.8 Patent law has claims to give notice of a patent’s

metes and bound.9 Neither trademark nor copyright law has claims,
leaving courts unclear on operationalizing technical similarity or
market substitution considerations.10
In a negative feedback loop, indeterminacy over the likelihood of
confusion standard muddies trademark law’s focal point and scope,
while polluting adjacent disciplines like copyright and patent law.11
Proper notice about the existence and scope of legal rights is critical
to any property system, but especially trademark rights, because
trademarks last indefinitely, meaning their owners obtain a timeless
monopoly without the same limitations and threshold requirements
placed on patent and copyright holders.12 A patchwork of inconsistent
results destabilizes the system for everyone.13 Indeterminacy has many
negative impacts, including causing negotiations to break down, which
harms both brand owners and potential licensees, and acting as a drag
on dispute resolution, compliance, and social equity.14 Indeterminacy
also acts as a drag on dispute resolution, compliance, and social
equity.15 The rational response must be a call for clarity in the law.
The likelihood of confusion standard examines whether consumers
will likely be confused as to the source, identity, sponsorship, or origin
of the goods and requires “a substantial likelihood that the public will

8. Michael Grynberg, Thick Marks, Thin Marks, 67 CASE W. RSRV. L. REV. 13, 15
(2016) (“Many open questions in modern trademark law concern which parts of the
range belong under the trademark holder’s control.”).
9. 35 U.S.C. § 112(b) (requiring patentees to include in their patent “one or
more claims particularly pointing out and distinctly claiming the subject matter which
the inventor . . . regards as the invention”).
10. Jeanne C. Fromer & Mark A. Lemley, The Audience in Intellectual Property
Infringement, 112 MICH. L. REV. 1251, 1296–99 (2014).
11. See Robert G. Bone, Notice Failure and Defenses in Trademark Law, 96 B.U. L. REV.
1245, 1255 (2016) (“[W]hat makes the scope of rights so uncertain is the vagueness of
the likelihood-of-confusion test (“LOC test”) for infringement.”) [hereinafter Bone,
Notice Failure and Defenses]; Amy Adler & Jeanne C. Fromer, Taking Intellectual Property
into Their Own Hands, 107 CALIF. L. REV. 1455, 1523 (2019) (“Trademark law is similarly
complex and unpredictable with regard to important doctrines.”).
12. See William M. Landes & Richard A. Posner, Trademark Law: An Economic
Perspective, 30 J.L. & ECON. 265, 287 (1987) (“The lack of a fixed term for trademarks
is one of the striking differences between trademarks, on the one hand, and copyrights
and patents, on the other.”). See generally 15 U.S.C. § 1114 (discussing infringement
and remedies).
13. Thomas H. Watson, Pay Per Click: Keyword Advertising and the Search for
Limitations of Online Trademark Infringement Liability, 2 CASE W. RSRV. J.L. TECH. &
INTERNET 101, 122 (2011).
14. See, e.g., Bone, supra note 11, at 1258.
15. Daryl Lim, AI, IP, Algorithms, and Inequality, SMU L. REV. (forthcoming 2022).
2022] TRADEMARK CONFUSION REVEALED 1289

be confused.”16 Each circuit court has enumerated factors relevant in

analyzing likelihood of confusion. Courts use proxies for consumer
confusion like intent, survey evidence, mark strength, and consumer
sophistication.17 However, these either fail to incorporate real-world
purchasing conditions as a doctrinal matter or are better considered
part of a streamlined likelihood of confusion test.18 The trio of actual
confusion, mark similarity, and competitive proximity provides judges
a small but potent cluster of factors.19 Together with safe harbors for
descriptive and expressive uses, these enable a more coherent,
consistent, and expedient resolution of trademark disputes.
This Article presents a contemporary empirical analysis of each
likelihood of confusion factor and how they interact with one another.
Conventional wisdom teaches us that courts should comprehensively
traverse each factor and that likelihood of confusion cases generally
require jury determination.20 The data reveals that neither is true.
Instead, courts provide early off-ramps to litigants by “economizing”
and applying just a handful of factors or by “folding” factors into each
other in grouped layers.21 The findings also reveal the Ninth Circuit’s
dominance, pro-defendant and pro-plaintiff fora, and the impact of
rivalry and fair use on case outcomes.22 This Article concludes with a
blueprint of how artificial intelligence (AI) systems can use empirical
studies as training data to help stakeholders make and predict
confusion analyses.23
Part I introduces this empirical study’s methodology, goals, and
limitations before elaborating on this Article’s doctrinal and policy
impetus. The discussion charts how blending technical trademarks and
trade names along with the expansion of triggers ensnares defendants
in trademark liability. It then shifts to make critical observations
gleaned from the data, including the impact of rivalry on modern case
outcomes, the dominance of the Ninth Circuit in federal trademark
litigation, and the most pro-defendant and pro-plaintiff circuits today.

16. Fisher Stoves, Inc. v. All Nighter Stove Works, Inc., 626 F.2d 193, 194 (1st Cir.
1980).
17. See infra Section I.C.
18. See generally 4 MCCARTHY, supra note 2, § 23:1 (criticizing the “amorphous and
indefinite” nature of the multi-factor likelihood of confusion test for creating legal
unpredictability, increasing litigation costs, and chilling socially valuable uses).
19. See infra Part II.
20. See infra Part II.
21. See infra Part III.
22. See infra Part I.
23. See infra Part III.
1290 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

Part I also presents a detailed doctrinal and empirical analysis of

prominent features in the likelihood of confusion analysis—the intent
of litigating parties, consumer surveys, mark strength, and consumer
sophistication—and explains why each factor leads courts tasked with
ascertaining consumer confusion astray. Finally, the Part introduces
coherence-based reasoning and argues that a more compact nucleus
of factors would better serve courts analyzing likelihood of confusion.
Part II explains why actual confusion, mark similarity, and
competitive proximity offer courts that compact troika in simplifying
likelihood of confusion analysis. The data reveals how most courts rely
on these three factors while either paying lip service to or completely
ignoring the other factors. It also shows appellate courts are complicit
in this “wink-and-nod” practice, affirming lower courts in over 80% of
cases on appeal. Finally, Part II explains why this trio of factors plus the
fair use safe harbors of descriptive and expressive uses should form
trademark law’s rules of thumb for infringement.
Part III examines the implications of the empirical study on
trademark doctrine and practice. First, it observes that while the
likelihood of confusion factors may present themselves as discrete
categories, the dataset reveals that courts do not regard them as such.
Instead, courts combine factors and analyze them both creatively and
rationally. Second, courts rely on a small number of factors to
economize their decisions and give parties an early off-ramp. Third,
Part III explains how the empirical analysis provides a blueprint for
algorithmic adjudication using AI, taking the reader from conception
to execution while identifying and addressing its limitations.

I. STUDYING CONFUSION
Over the years, the jurisprudential roots of trademark law became
unruly and tangled. Unfair competition intermingled with consumer
protection as the Lanham Act of 194624 (“the Act”) blended trade
names and technical trademarks.25 A later legislative revision untied
likelihood of confusion from source confusion, and courts introduced
idiosyncratic rules of affiliation and sponsorship as triggers for
consumer confusion.26
Within the likelihood of confusion tests, factors such as defendants’
intent, survey evidence, and consumer sophistication provided a
convenient but misguided attempt to get a handle on the arduous task

24. 15 U.S.C. § 1051–1127.

25. See infra Section I.B.
26. See infra Section I.B.
2022] TRADEMARK CONFUSION REVEALED 1291

of determining what trademark infringement had become.27 Judges

often resorted to coherence-based reasoning. Once a judge
determined the satisfaction of the selected factors, the judge would
decide that all the other factors were present.28 Though it made their
work easier, it muddied the waters for everyone else. This Part presents
an empirical analysis of these issues and their implications. It begins by
discussing the empirical methodology.

A. Methodology
This empirical study draws upon the well-developed method of case
content analysis.29 The method systematically dissects a sample of
judicial opinions to record consistent features, draw inferences, and
uncover trends.30 This social science approach to the law complements
and augments traditional legal analysis.31 As a testament to its outsized
contribution to the literature, case content analysis generates an
average of seventy-seven citations per article in a milieu where 40% of
law review articles receive no citations at all.32
The value of case content analysis lies in uncovering patterns in
judging. Judges routinely rely on a remarkably limited number of
factors in reaching their conclusions.33 Instead, they employ

27. See infra Section I.B.

28. See infra Section I.D.
29. Mark A. Hall & Ronald F. Wright, Systematic Content Analysis of Judicial Opinions,
96 CALIF. L. REV. 63, 73 (2008) (“Content analysis has proven useful for studying a
broad range of legal subject areas.”); id. (listing “areas as far-ranging as administrative
law, constitutional law, corporate and securities law, criminal law and procedure,
contracts, employment discrimination, health law, and torts”).
30. See KLAUS KRIPPENDORFF, CONTENT ANALYSIS: AN INTRODUCTION TO ITS
METHODOLOGY 18 (2d ed., 2004) (defining content analysis as “a research technique
for making replicable and valid inferences from texts (or other meaningful matter) to
the contexts of their use”).
31. Hall & Wright, supra note 29, at 74; see also id. at 65 (“The method also helps a
researcher to sort out the interaction of multiple factors that bear on an outcome in
the legal system.”); id. at 78 (“Its strength is to provide an objective understanding of
a large number of decisions where each decision has roughly the same value.”); Alan
L. Tyree, Fact Content Analysis of Case Law: Methods and Limitations, 22 JURIMETRICS J. 1,
23 (1981) (explaining that these methods have “considerable power for the discovery
of anomalies which may escape the naked eye”).
32. Hall & Wright, supra note 29, at 74 (“[C]ontent analysis projects appear
somewhat more likely to generate discussion and citation than law review articles more
generally.”).
33. See, e.g., Beebe, supra note 6, at 1602 (“The data collected for this study support
the general hypothesis that decision makers, even when making complex decisions,
consider only a small number of factors and the more specific hypothesis that, in doing
so, decision makers use a core attributes heuristic.”).
1292 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

heuristics—such as the likelihood of confusion factors—to cut through

what would otherwise be a morass of information that could paralyze
decision-making entirely.34 This approach, however, makes it difficult
to draw broader conclusions to inform future cases coherently.35
Scholars employ case content analysis to parse through court decisions
and study how judges and juries apply rules to facts in practice to
address this limitation.36 Thus, while the interpretive method evaluates
legal principles, case content analysis “combines the analytical skills of
the lawyer with the power of science that comes from articulated and
replicable methods of reading and counting cases.”37 In so doing, case
content analysis yields useful information that moves the discussion
toward a greater understanding of the bigger policy questions and
helps uncover areas for further research.38 It also avoids selection bias
issues, which hamper the representativeness of other methods.39
This Article relied on an expansive pool of 188 cases covering nearly
5,000 datapoints based on a Westlaw search for all trademark
infringement cases discussing likelihood of confusion over five years
between September 30, 2016, and October 1, 2021.40 The study

34. See, e.g., Daryl Lim, Retooling the Patent-Antitrust Intersection: Insights from
Behavioral Economics, 69 BAYLOR L. REV. 124 (2017).
35. See, e.g., Daryl Lim, The (Unnoticed) Revitalization of the Doctrine of Equivalents, ST.
JOHN’S L. REV. (forthcoming 2022) [hereinafter Lim, Doctrine of Equivalents].
36. See, e.g., Lee Petherbridge, On the Decline of the Doctrine of Equivalents, 31
CARDOZO L. REV. 1371, 1380 (2010) [hereinafter Petherbridge, Decline of the Doctrine of
Equivalents] (“Content analysis is capable of helping scholars verify, analyze, or refute
empirical claims about case law, and it is to that purpose the approach is put in this
study.”). For earlier studies where I employed a similar methodology, see Lee
Petherbridge, Jason Rantanen, & Ali Majibi, The Federal Circuit and Inequitable Conduct:
An Empirical Assessment, 84 S. CALIF. L. REV. 1293, 1304 (2011); Hall & Wright, supra
note 29, at 77; DARYL LIM, PATENT MISUSE & ANTITRUST: EMPIRICAL, DOCTRINAL & POLICY
PERSPECTIVES 8–9 (2013); Lim, Judging Equivalents, supra note 7.
37. Hall & Wright, supra note 29, at 100.
38. Karen A. Jordan, Empirical Studies of Judicial Decisions Serve an Important Role in
the Cumulative Process of Policy Making, 31 IND. L. REV. 81, 88 (1998); see also Lon L.
Fuller, An Afterward: Science and the Judicial Process, 79 HARV. L. REV. 1604, 1622 (1965)
(“[P]ossible gain from researches of this kind lies in the realm of serendipity. A
puzzling correlation that violates normal anticipations may set our minds going along
new paths and yield unexpected insights.”).
39. Hall & Wright, supra note 29, at 102 (“All of these were universal samples
restricted only by date, subject matter, jurisdiction, and/or source. In short, empirical
researchers studying case law are usually able to avoid the selection bias issues that
plague most other areas of social science.”).
40. For an example of another recent empirical study on trademarks that starting
and ending during the calendar year, see Lisa Larrimore Ouellette, The Google Shortcut
to Trademark Law, 102 CALIF. L. REV. 351, 373 (2014) (“New U.S. federal court decisions
2022] TRADEMARK CONFUSION REVEALED 1293

omitted thirty-two cases captured by Westlaw’s search results because

they did not specifically discuss the likelihood of confusion factors.41
For accuracy, the dataset distinguishes between procedural wins (for
instance, defeating a motion for summary judgment) and substantive
wins on the merits (which result in a finding of infringement or non-
infringement).42 This Article initially used Excel to hand-code the data
before using IBM SPSS Statistics 28.0 to generate the graphs and
crosstabs data.43
The dataset of hand-coded cases included variables such as (1) the
decision’s date; (2) the judicial circuit; (3) whether a district or
appellate court decided the case; (4) the parties’ relationship as rivals
(or not); (5) the procedural posture; (6) the type of mark; (7) the test
employed; (8) whether the opinion discussed survey evidence; (9)
which party a likelihood of confusion factor favored; (10) whether
courts “folded” factors together; (11) case outcomes; and (12) whether
the court discussed fair use.44
Like all empirical studies, this one has its caveats. There are several
well-recognized limitations to case content study databases. First,
coding may result in incomplete or inaccurate coding, despite cross-
coding and verification using a population sample.45 Given that the
focus is on features of written decisions, the data remains valid as long
as it is recognized to refer to a specific population rather than a sample
of all cases in all possible worlds.46 Second, cases from legal databases
such as Westlaw are known to underreport jury decisions.47 To some

related to trademark or service mark distinctiveness or likelihood of confusion were

tracked from December 2011 to November 2012.”); see also John R. Allison & Mark A.
Lemley, The (Unnoticed) Demise of the Doctrine of Equivalents, 59 STAN. L. REV. 955, 963
(2007) (“[W]e collected every district court and court of appeals decision on the
doctrine of equivalents that appeared in Westlaw . . . .”). The number of cases this
Article employs compares favorably with the norm, which ranges from less than 100
cases to 300 cases. See Hall & Wright, supra note 29, at 102 (“Of these 114 universal
samples, only 11 coded more than 1000 cases, and 21 coded from 500 to 1000. Twenty-
six of these projects coded fewer than 100 cases (with 13 of these fewer than 51), and
39 coded between 100 and 300.”).
41. See Daryl Lim, Trademark Case Statistics (Dec. 17, 2021) (on file with the
American University Law Review).
42. Id.
43. Id.
44. Id.
45. See, e.g., Lim, Doctrine of Equivalents, supra note 35.
46. Id.
47. Allison & Lemley, supra note 40, at 963–64 (“The universe of all decisions is of
course different from the universe of those reported in Westlaw, and in particular our
study underreports jury decisions. But our focus on written decisions (both reported
1294 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

degree, comparing it to other studies that employ similar methods to

control for that feature can mitigate the effects of the underreporting.
Courts have found that “simple counts and percents are sufficient to
document” a claim about case law trends, challenge conventional
wisdom, or suggest further study issues.48 The case-counting method
codes the entire population of relevant cases.49 Statistics are
unnecessary to prove that sample cases are representative of a larger
population.50 Because the outcome of each case—the dependent
variable of interest—has five possible categories, the most appropriate
regression model would be a multinomial regression.51 However, each
of the seven relevant factors also has five possible categories. Hence,
the sample size required for a multivariate regression that would
simultaneously test each category’s effect of each relevant factor is
much larger than that of the current dataset.52 This Article refrained
from presenting regression results and instead presented descriptive
statistics. Specifically, it describes the observed distribution of case
outcomes when conditioned on one or two relevant factors. Although
this does not attest to a specific level of statistical significance, these
values are still instructive for the reader.
Third, analysis of judicial opinions has well-known limitations.53
Statistics fail to account for extralegal factors influencing outcomes,
such as the state of the case record on appeal and judicial deliberations
discussed in the opinion.54 In addition, litigants may consider the
expertise and reputation of the district court judge in deciding
whether to appeal, introducing selection bias effects into the appellate

and unreported) allows us to parse the grounds for decision and the reasoning of the
opinions.”).
48. Hall & Wright, supra note 29, at 118.
49. Id.
50. Id.
51. ALAN AGRESTI, CATEGORICAL DATA ANALYSIS 293 (3d ed. 2013).
52. Id.
53. See R. Polk Wagner & Lee Petherbridge, Is the Federal Circuit Succeeding? An
Empirical Assessment of Judicial Performance, 152 U. PA. L. REV. 1105, 1128–29 (2004)
(discussing unobserved reasoning, strategic behavior, and selection bias).
54. Harry T. Edwards & Michael A. Livermore, Pitfalls of Empirical Studies that
Attempt to Understand the Factors Affecting Appellate Decisionmaking, 58 DUKE L.J. 1895,
1899 (2009).
2022] TRADEMARK CONFUSION REVEALED 1295

data.55 Moreover, most cases settle, so decided cases are a nonrandom

subset of all cases.56
Fourth, case content analysis trades depth for breadth. The
complexity of trademark litigation also makes it difficult to generalize
even from a study covering hundreds of cases.57 Numbers do not reflect
judicial rhetoric or more subtle clues about a judicial opinion’s
precedential value.58 Case coding documents what judges do rather
than draw normative implications from the observations.59
Fifth, parties are not randomly distributed throughout the judicial
districts.60 Some district courts may hear more cases that eventually
settle. Other courts may hear more cases where the parties file based
on domicile. District court judges are therefore not assigned a random
sample of patent lawsuits since they are assigned cases from the judicial
district where they sit.61 Circumstances such as a particular judge or
jury may cause a case to settle where the same case before another
judge or jury could proceed to an appeal.62 This Article focuses on how
appellate and lower courts interpret precedent. Those interpretations
are not uniform and can never be so.63

55. David Schwartz, Practice Makes Perfect? An Empirical Study of Claim Construction
Reversal Rates in Patent Cases, 107 MICH. L. REV. 223, 243 (2008).
56. Jay P. Kesan & Gwendolyn G. Ball, How Are Patent Cases Resolved? An Empirical
Examination of the Adjudication and Settlement of Patent Disputes, 84 WASH. U. L. REV. 237,
273–74 (2006) (finding that between 65% and 68% of all patent cases filed in three
particular years were resolved via settlement or a probable settlement).
57. See David L. Schwartz, Explaining the Demise of the Doctrine of Equivalents, 26
BERKELEY TECH. L.J. 1157, 1188 (2011) (“Because patent litigation as a whole is so
complex, it is incredibly complicated to develop and test empirical models.”);
Petherbridge, Decline of the Doctrine of Equivalents, supra note 36, at 1380 (noting biases
inherent in this approach such as “unobserved reasoning, selection bias, and strategic
behavior”).
58. Hall & Wright, supra note 29, at 87.
59. Id. at 97 (“Still, imperfect data must suffice because observing actual behaviors
and gauging true attitudes would be impossible or cost-prohibitive. Similarly, even
though judge-reported facts may not ‘purport to be the real facts,’ they are ‘near
enough so that the savings in labor justifies the approximation.’”).
60. Kimberly A. Moore, Forum Shopping in Patent Cases: Does Geographic Choice Affect
Innovation?, 79 N.C. L. REV. 889, 924–25 (2001).
61. Schwartz, supra note 55, at 242.
62. See id. at 242 n.119 (“‘[I]f [judges] have a really tough case, they can put
tremendous pressure on the parties to settle so there won’t be an appealable order.’”
(alteration in original) (quoting Judge Richard A. Posner)).
63. See Jeffrey J. Rachlinski, Does Empirical Evidence on the Civil Justice System Produce
or Resolve Conflict?, 65 DEPAUL L. REV. 635 (2016) (“[E]ven when the empirical scholars
completely agree on the underlying facts, interpretation of the results can dramatically
1296 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

Sixth, case outcomes are impacted by parties’ factor-based

calculation of a successful outcome. The Priest-Klein “selection
hypothesis” predicts that, given various conditions, plaintiff win rates
at trials should approach fifty percent because only the close cases
survive settlement—or summary adjudication.64 The hypothesis
assumes parties have equal stakes in the litigation.65 More recent
studies cast the fifty-percent hypothesis in doubt, including those
dealing specifically with intellectual property law.66 Notably, “win rate”
means the percentage of time one party (plaintiff or defendant) wins
when a factor is decided in that party’s favor, not the percentage of
time that party wins when the factor is relevant. With these caveats in
mind, this discussion turns to the theory underlying likelihood of
confusion and the points of departure from conventional wisdom in
practice.

B. Blends, Triggers, and Polaroid Factors

This Section opens by discussing the impact of blending technical
trademarks and trade names in modern trademark law. It proceeds to
introduce the likelihood of confusion factors before presenting a
doctrinal and empirical analysis of intent, surveys, mark strength, and
consumer sophistication, arguing that each, in turn, detracts from an
accurate likelihood of confusion analysis. Finally, the Section closes by

differ. Empirical legal scholarship is still worth conducting, but the hope that it will
resolve partisan debates in law is unrealistic.”).
64. See George L. Priest & Benjamin Klein, The Selection of Disputes for Litigation, 13
J. LEGAL STUD. 1, 4–6, 17–18 (1984). Priest and Klein’s fifty percent has been modified
when there are different stakes involved. For example, if the plaintiff has more to win
than the defendant has to lose. See Yoon-Ho Alex Lee & Daniel M. Klerman, Updating
Priest and Klein 2 (Univ. S. Cal. Ctr. L. & Soc. Sci., Research Paper No. 15-21, https://
papers.ssrn.com/sol3/papers.cfm?abstract_id=2619856.
65. Priest & Klein, supra note 64, at 24–29.
66. See e.g., Theodore Eisenberg, Testing the Selection Effect: A New Theoretical
Framework with Empirical Tests, 19 J. LEGAL STUD. 337, 338–39 (1990) (testing the fifty-
percent hypothesis and rejecting it as a description of all civil litigation); see also Mark
A. Lemley & Colleen V. Chien, Are the U.S. Patent Priority Rules Really Necessary?, 54
HASTINGS L.J. 1299 (2003) (arguing that the Priest-Klein hypothesis is not borne out
by the data in patent cases); Jason Rantanen, Why Priest-Klein Cannot Apply to Individual
Issues in Patent Cases, (Univ. Iowa Legal Stud., Research Paper No. 12-15, 2012),
http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2132810 (“At best, the Priest-
Klein hypothesis only applies to the selection of disputes, not the selection of
individual issues. Due to the presence of multiple issues in patent cases, there is
axiomatically no basis for inferring that a patentee would expect a 50 percent chance
of winning on each one.”).
2022] TRADEMARK CONFUSION REVEALED 1297

explaining the impact of coherence-based reasoning on the likelihood

of confusion factors.

1. Blends and triggers

Early trademark common law distinguished between trade names
and technical trademarks.67 Most trade name disputes involved rivals.68
Unfair competition law governed these disputes and focused on
directly competing uses diverting trade,69 taking the form of passing
off and reverse passing off business names.70 In the twentieth century,
courts blurred the distinction between technical trademarks and trade
names, blending the most expansive aspects in favor of trademark
owners.
In 1946, Congress “federalized” common law protection of
trademarks used in interstate commerce with the Lanham Act.71 The
Act codified this blended standard, requiring only that the
unauthorized use be connected with goods or services.72 Trade names
enjoyed the protection offered to technical trademarks as long as
owners could show “secondary meaning.”73 Cases interpreted this as an
association by consumers with the source of the product that imbued
trade names with an acquired distinctiveness.74 The Act subsequently
welded the two concepts, allowing all signs to acquire distinctiveness
through secondary meaning.75

67. See FRANK I. SCHECHTER, THE HISTORICAL FOUNDATIONS OF THE LAW RELATING
TO TRADE-MARKS 161 (1925).
68. Edward S. Rogers, The Lanham Act and the Social Function of Trademarks, 14 LAW
& CONTEMP. PROB. 173, 178–80 (1949).
69. See Mark P. McKenna, The Normative Foundations of Trademark Law, 82 NOTRE
DAME L. REV. 1839, 1904 (2007) (noting “that courts only developed the likelihood of
confusion factors after jettisoning the requirement of direct competition”).
70. See CORPORATE COUNSEL’S GUIDE TO UNFAIR COMPETITION § 3:31 (2021).
“Passing off” occurs when defendants sell its goods with the plaintiff’s mark, whereas
in “reverse passing off,” defendants sell plaintiff’s goods with the defendant’s
trademark. Id.
71. CAE, Inc. v. Clean Air Eng’g, Inc., 267 F.3d 660, 672 (7th Cir. 2001).
72. See Lanham Act, Pub. L. No. 79-489, § 32(1)(a), 60 Stat. 427, 437 (1946)
(specifying that only goods and services fell under the purview of the Act).
73. See, e.g., Milton Handler & Charles Pickett, Trade-Marks and Trade Names—An
Analysis and Synthesis, 30 COLUM. L. REV. 168, 168–69 (1930).
74. See E. Columbia, Inc. v. Waldman, 181 P.2d 865, 867 (Cal. 1947) (en banc).
75. Compare 15 U.S.C. § 1052(f) (2006) (allowing registration of a mark “which has
become distinctive of the applicant’s goods in commerce”), with Trade-Mark Act of
1905, Pub. L. No. 58-84, § 5(b), 33 Stat. 724, 725–26 (providing that no mark that is
distinguishable “from other goods of the same class shall be refused” trademark
registration because of the nature of the mark).
1298 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

As sellers expanded into adjacent product markets in the post-World

War Two era, courts expanded the scope of protection to include
complementary products and services.76 Trademark scope could protect
virtually anything that functions as an identifier of source—shapes, colors,
smells, and sounds.77 Congress included new types of protectable
subject matter from technical trademarks to “anything . . . capable of
carrying [source] meaning” as a potential trademark.78 As a result, the
likelihood of confusion standard became more complex. Although
courts previously compared the marks, cases now require courts to
consider a much broader range of information, including advertising
slogans, product packaging, and product designs.

Figure 1: Trade Names, Technical Trademarks, and Modern Trademarks

Trade Technical Modern

Names Trademarks Trademarks
Sufficient if Requires Sufficient if
Distinctiveness
descriptive distinctiveness descriptive

Intent Intent required Strict Liability Intent optional

Actual harm Likelihood of Likelihood of
Harm
required harm sufficient harm sufficient
Comparison No Yes Optional

Injunction Narrow Broad Broad

Contemporary empirical evidence from this Article’s dataset

indicates that defendants win on the merits 26% of the time when the
parties are rivals. Defendants win on the merits only 46% when they
are not rivals, with rivals winning twice as often. These numbers show
the impact of unfair competition in shaping modern trademark
doctrine.
Congress subsequently amended the Act to remove the restriction
on source confusion, allowing courts to consider other forms of

76. See PAMELA WALKER LAIRD, ADVERTISING PROGRESS: AMERICAN BUSINESS AND THE
RISE OF CONSUMER MARKETING 31 (1998) (discussing post-war expansion of consumer
products).
77. See Bone, Notice Failure and Defenses, supra note 11, at 1268.
78. Qualitex Co. v. Jacobson Prods. Co., 514 U.S. 159, 162 (1995).
2022] TRADEMARK CONFUSION REVEALED 1299

confusion in the infringement analysis.79 Courts dutifully expanded

the scope of confusion from purchasers to include non-purchasers
(“post-sale confusion”) and allowed businesses to prohibit confusion
over sponsorship or endorsement of goods and services.80
Law and economics scholarship prompted this expansion, driven by
a belief that stronger protection maximized wealth and, in turn,
promoted economic efficiency.81 The resulting fusion infused unfair
competition into trademark law and invited courts to find defendants’
marks infringing well before consumers purchased a product or
service with the allegedly infringing mark, based on the idea that
defendants misappropriated the plaintiff’s goodwill to appeal to
consumers.82
With new triggers, confusion can manifest itself in various ways. For
instance, “forward confusion occurs when ‘the junior user attempts to
trade on the senior[] user’s goodwill and reputation,’” misleading
consumers to believe that the junior and senior user’s goods or services
are related.83 Similarly, reverse confusion occurs where consumers
believe the junior user is the source of the senior user’s goods.84
Whereas protection previously stopped at the shores of adjacent
products, trademark law now allows even a pancake chain restaurant
to attempt to prohibit an evangelical Christian organization from using

79. See S. REP. NO. 87-2107, at 4 (1962), reprinted in 1962 U.S.C.C.A.N. 2844, 2847.
80. Act of Oct. 9, 1962, Pub. L. No. 87-772, § 2, 76 Stat. 769, 769 (deleting the
requirement that confusion be of purchasers as to the source of origin of such goods
or services).
81. See e.g., W. T. Rogers Co. v. Keene, 778 F.2d 334, 339 (7th Cir. 1985)
(explaining that “competition is not impaired by giving each manufacturer a perpetual
‘monopoly’ of his identifying mark” if he has chosen a “distinctive” trademark where
the available names are “for all practical purposes infinite”); see Landes & Posner, supra
note 12, at 270–79 (advancing Chicago School economic theory within trademark law
scope).
82. Gibson Guitar Corp. v. Paul Reed Smith Guitars, 423 F.3d 539, 549 (6th Cir.
2005).
83. Fuel Clothing Co. v. Nike, Inc., 7 F. Supp. 3d 594, 610 (D.S.C. 2014) (quoting
Coryn Grp. II, LLC v. O.C. Seacrets, Inc., CIV WDQ-08-2764, 2010 WL 1375301, at *4
(D. Md. Mar. 30, 2010)).
84. Kelly-Brown v. Winfrey, 717 F.3d 295, 304–05 (2d Cir. 2013); J.T. Colby & Co.
v. Apple Inc., 586 F. App’x 8, 9 (2d Cir. 2014) (“The Lanham Act guards against this
‘reverse confusion’ to prevent ‘a larger, more powerful company [from] usurping the
business identity of a smaller senior [trademark] user.’”).
1300 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

a similar mark.85 This development caused a jurisprudential disjuncture

to occur.
While the statute had changed, earlier courts did not update the
likelihood of confusion test, which had been designed to capture more
than just source confusion.86 As will be shown below, factors like
consumer sophistication, the likelihood of expansion, and the
marketing channels are of little assistance in evaluating whether a
company’s claim that it is the exclusive soda for sporting events in the
minds of the consumers is true.87 Worse, the multiple targets that the
likelihood of confusion standard now addresses makes applying it even
more unwieldy and unpredictable.88

2. The Polaroid factors

The multifactor test for the likelihood of confusion attempts to
provide analytical rigor to the complicated question of how consumers
perceive different marks. Barton Beebe’s 2006 empirical study
revealed courts most frequently deployed the Second Circuit’s test in
Polaroid Corp. v. Polaroid Electronics Corp.89 In that case, Judge Friendly
articulated what became known as the eight Polaroid factors:
(1) strength of the plaintiff’s mark;
(2) similarity of plaintiff’s and defendant’s marks;
(3) competitive proximity of the products;
(4) likelihood that plaintiff will “bridge the gap” and offer a product
like a defendant’s;
(5) actual confusion between products;
(6) good faith on the defendant’s part;

85. First-Amended Complaint for Injunctive Relief for Federal Trademark

Infringement and Dilution at 6, IHOP IP, LLC v. Int’l House of Prayer, No. CV10-6622
2010 WL 3775268 (C.D. Cal. Sept. 9, 2010).
86. See, e.g., King of the Mountain Sports, Inc. v. Chrysler Corp., 185 F.3d 1084,
1090 (10th Cir. 1999) (discussing the three-part test to determine similarity between
marks).
87. Supreme Assembly, Ord. of Rainbow for Girls v. J.H. Ray Jewelry Co., 676 F.2d
1079, 1082 (5th Cir. 1982) (“other association”); Caterpillar Inc. v. Walt Disney Co.,
287 F. Supp. 2d 913, 918 (C.D. Ill. 2003) (“otherwise affiliated”).
88. James Gibson, Risk Aversion and Rights Accretion in Intellectual Property Law, 116
YALE L.J. 882, 908 (2007) (“The case law on sponsorship and approval, however, is so
ambiguous as to make it almost impossible to know ex ante whether a given use will be
infringing.”).
89. Beebe, supra note 6, at 1593 (“This is especially true in the Second Circuit
where the multifactor test is most often applied and where appellate panels have
repeatedly emphasized that the multifactor analysis must be exhaustive and explicit.”);
Polaroid Corp. v. Polarad Elecs. Corp., 287 F.2d 492, 495 (2d Cir. 1961).
2022] TRADEMARK CONFUSION REVEALED 1301

(7) quality of defendant’s product; and

(8) sophistication of the buyers.90
Confusion is more likely when an accused product contains multiple
indicia of similarity.91 For instance, house brands typically include
house marks, product-specific marks, product packaging, and color or
configuration.92 Conversely, consumers are less likely to be confused
when defendants copy only a few elements.93 However, no single factor
in the likelihood of confusion inquiry is determinative. Conventional
wisdom teaches that courts need to undertake “a highly fact-intensive
inquiry both as to the assessment of the evidence concerning each
factor and as to the overall synthesis of factors and the evidence.”94
Trademark litigation is inherently impressionistic, particularly
because actual confusion is rare. Sometimes, each side claims a
numerically equal number of factors in their favor, leaving courts to
assign weights.95 Courts caught up in the swirl sloppily pepper their
judgments with different operative terms to describe the same thing,
including affiliation,96 endorsement,97 connection,98 and whether the
use produced confusion “of any kind.”99 As the Fifth Circuit bluntly put
it, “Congress adopted an open-ended concept of confusion. . . . Any kind
of confusion will now support an action for trademark infringement.”100
Unfortunately, courts in subsequent cases as well as businesses and
their legal advisors struggle to determine the appropriate strength of

90. Polaroid, 287 F.2d at 495.

91. See Beebe, supra note 6, at 1625 (noting that judges “emphasize similarities over
differences,” but finding that “the degree of similarity of [] marks does not appear to
significantly affect the outcome of the test”).
92. See, e.g., Bristol-Myers Squibb Co. v. McNeil-P.P.C., Inc., 973 F.2d 1033, 1042
(2d Cir. 1992).
93. George Miaoulis & Nancy D’Amato, Consumer Confusion & Trademark
Infringement, 42 J. MKTG. 48, 54 (1978) (finding, in the context of competing goods,
that the “primary cue for [] association [between two brands] was not the name but
the visual appearance”).
94. Select Comfort Corp. v. Baxter, 996 F.3d 925, 933–34 (8th Cir. 2021) (“We
have repeatedly emphasized that no one factor is controlling and different factors will
carry more weight in different settings.”).
95. Equitable Nat’l Life Ins. Co. v. AXA Equitable Life Ins. Co., 434 F. Supp. 3d
1227, 1252 (D. Utah 2020) (“Ultimately, while each side can claim three factors weigh
in its favor, they do not weigh equally.”).
96. Pebble Beach Co. v. Tour 18 I Ltd., 155 F.3d 526, 544 (5th Cir. 1998).
97. Id.
98. Id. at 543.
99. Syntex Lab’ys, Inc. v. Norwich Pharmacal Co., 437 F.2d 566, 568 (2d Cir. 1971).
100. Armstrong Cork Co. v. World Carpets, Inc., 597 F.2d 496, 500 n.5 (5th Cir.
1979).
1302 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

each factor, either alone or relative to other factors.101 Judges

themselves admit the distinctions they make are often done on an
“intuitive basis” rather than through “logical analysis.”102 Reporting on
his dataset of cases, Beebe observed that “scattered among the circuits
are factors that are clearly obsolete, redundant, or irrelevant, or, in the
hands of an experienced judge or litigator, notoriously pliable.”103
Like an untended garden, the likelihood of confusion standard has
grown wild, with different circuit courts spinning off anywhere
between six and thirteen factors.104 Some circuits favor factors others
ignore, and courts have called nearly every factor or factor
combination the most important.105 The reason for this may be
divergent conceptions of trademark policy, with some courts focusing
on unfair competition while others are concentrating on consumer
confusion.106
This Article reveals for the first time in the trademark literature that
the Second Circuit’s Polaroid factors no longer dominate modern
trademark jurisprudence. Instead, the Ninth Circuit’s Sleekcraft factors
have edged out the Polaroid factors as those most frequently applied as
the Ninth Circuit now has the most trademark infringement cases.107
As a result, the Second Circuit is now the second most dominant circuit.

101. Joseph P. Liu, Two-Factor Fair Use?, 31 COLUM. J.L. & ARTS 571, 579 (2008)
(“Under a multi-factor balancing test, it is difficult to register the relative strength of
the factors.”); Eric Goldman, Online Word of Mouth and its Implications for Trademark
Law, in TRADEMARK LAW AND THEORY: A HANDBOOK OF CONTEMPORARY RESEARCH 404,
415–16, 424 (Graeme B. Dinwoodie & Mark D. Janis eds., 2008) (“Assessing consumer
confusion about product source is an inherently inexact process.”).
102. Union Carbide Corp. v. Ever-Ready Inc., 531 F.2d 366, 379 (7th Cir. 1976).
103. Beebe, supra note 6, at 1583–84.
104. See infra, Section I.B.
105. Beebe, supra note 6, at 1583.
106. Alejandro Mejías, The Multifactor Test for Trademark Infringement from a European
Perspective: A Path to Reform, 54 IDEA 285, 314 (2014) (“[T]here is also divergence on
how the factors are treated and employed.”); see Beebe, supra note 6, at 1596–97
(summarizing in chart form the different factors each circuit considers and reporting
“substantial intercircuit variation in plaintiff multifactor test win rates.”).
107. AMF Inc. v. Sleekcraft Boats, 599 F.2d 341, 350–51 (9th Cir. 1979).
2022] TRADEMARK CONFUSION REVEALED 1303

Figure 2: Circuit Variances over Time

This shift may be significant for litigation strategy, especially because

the Second Circuit was the most defendant-friendly circuit where
plaintiffs’ win rate was 31% and the defendant’s win rate was 48%. By
comparison, both plaintiff and defendant win rates were 38% at the
Ninth Circuit. In contrast to the Second and Ninth Circuits, the
Federal Circuit was the most plaintiff-friendly, with an 83% plaintiff
win rate. Defendants there fare comparatively poorly, winning a mere
8% of cases.108 The figure below shows the distribution of cases and
outcomes across circuits.

108. It is possible that defendants fare so poorly at the Federal Circuit because they
are likely on appeal from the Trademark Trial and Appeal Board. However, the data
is inconclusive on this point and invites further study.
1304 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

Figure 3: Circuit by Outcome

As a doctrinal matter, the difference may be less material. As it turns

out, the Ninth Circuit’s factors mirror those of the Second Circuit. The
only difference between the two sets of factors is in linguistics as the
Ninth Circuit considers marketing channels used to promote the
products which is the same as the Second Circuit’s consideration of the
competitive proximity of the products and services.109 Similarly, this
Article shows that the Second Circuit’s “quality of defendant’s
product” factor can be subsumed into the competitive proximity
factor.110 As an empirical matter, both factors rarely appear in case
reports, with the Ninth Circuit’s “marketing channels” factor discussed
in 13% of cases and the “quality” factor appearing in 11% of cases.111

109. See e.g., Las Vegas Sands Corp. v. Fan Yu Ming, 360 F. Supp. 3d 1072, 1077 (D.
Nev. 2019) (folding the two factors together).
110. See infra Section II.C.
111. Compare, for example, with mark similarity, which appeared in 85% of cases,
competitive proximity in 73% of cases, and actual confusion in 74% of cases. See infra
Appendix.
2022] TRADEMARK CONFUSION REVEALED 1305

Figure 4: Comparing the Polaroid and Sleekcraft Factors

Polaroid Factors Sleekcraft Factors

Strength of the plaintiff’s mark Strength of the plaintiff’s mark
Similarity of plaintiff’s and Similarity of plaintiff’s and
defendant’s marks defendant’s marks
Competitive proximity of products Competitive proximity of products
or services or services
Marketing channels used to
-
promote the products
Likelihood that plaintiff will
Likelihood that either party will
“bridge the gap” and offer a
expand into new markets.
product like a defendant’s
Actual confusion Actual confusion
Defendant’s good faith Defendant’s intent
Quality of defendant’s product -
Type of goods and the carefulness
Buyer sophistication
of likely consumers

Substitution bias within each circuit’s set of factors is particularly

virulent when open-ended wording gives courts cover, as the Act does
here.112 Courts applying the Act took that opportunity and leaned into
the likelihood of confusion factors like defendants’ intent, survey
evidence, and trademark strength, which were malleable and easy to
wield to reach their desired outcomes.113 Savvy trademark attorneys
also saw the opportunity to leverage more clever lawyering and focus
less on the case’s merits.114 Strikingly, Beebe’s empirical study reported
that intent and surveys were so heavily weighted that courts stampeded
over other factors.115 Overreliance on these proxies results in a
fundamentally flawed analysis. The next Section explains why.

112. See e.g., SNA, Inc. v. Array, 51 F. Supp. 2d 554, 562–63 (E.D. Pa. 1999), aff’d sub
nom. Silva v. Karlsen, 259 F.3d 717 (3d Cir. 2001) (concluding that defendants’ attempt
to use metatags to “lure internet users to their site” was in bad faith).
113. See infra Section I.C.
114. See Beebe, supra note 6, at 1581 (suggesting this problem exists even with true
source confusion cases because outcomes tend to be driven by the court’s focus on
intent).
115. Id. at 1607.
1306 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

C. Missing the Point on Consumer Confusion

What constitutes “confusion” is highly subjective and difficult to
evaluate. Proxies like intent, survey evidence, mark strength, and
consumer sophistication fail to incorporate real-world purchasing
conditions or are better considered within other factors. Trademark
infringement is fundamentally flawed if the likelihood of confusion
turns on these proxies.

1. Intent
Likelihood of confusion’s good faith or intent factor examines
whether defendants sought to benefit from plaintiffs’ goodwill.116 All
circuits but the Federal Circuit recognized this as a major factor in
finding liability.117 “In analyzing whether a defendant has acted in bad
faith, the question is whether the defendant attempted ‘to exploit the
good will and reputation of a senior user by adopting the mark with
the intent to sow confusion between the two companies’ products.’”118
Courts recognize that intentional copying may not indicate that the
defendant attempted to capitalize on the plaintiff’s trademark or trade
dress.119 However, there may be legitimate reasons to copy or imitate
the primary features of another company’s product. These include
functional features that have economic benefits without any secondary
meaning.120 In doing so, courts “want competitors to be inspired by—and
to improve on—the findings of their predecessors.”121 Therefore, it is a
“nefarious variety of passing off—the kind that confuses consumers
and exploits a competitor’s established goodwill—that trademark law
is prepared to prevent.”122
Stating the distinction is easy in theory, hard in practice. Cases in the
dataset reveal divergent views on when defendants cross the line. Some
courts are prepared to exculpate defendants if they had no intent to
confuse consumers.123 Indeed, one court commended “upcycling,” or

116. Sicilia Di R. Beibow & Co. v. Cox, 732 F.2d 417, 431 (5th Cir. 1984) (discussing
how the proper test focuses mainly on intent).
117. See Beebe, supra note 6, at 1589–90.
118. Tiffany & Co. v. Costco Wholesale Corp., 971 F.3d 74, 88 (2d Cir. 2020).
119. See Beebe, supra note 6, at 1630.
120. Fuddruckers, Inc. v. Doc’s B.R. Others, Inc., 826 F.2d 837, 844–45 (9th Cir.
1987).
121. Vital Pharms., Inc. v. Monster Energy Co., No. 19-60809-CIV, 2021 WL
3371942, at *44 (S.D. Fla. Aug. 3, 2021).
122. Id.
123. Id. at *53 (“If a defendant intentionally copies an aspect of the plaintiff’s
product, but not with intent to confuse consumers, then the defendant’s intent has
2022] TRADEMARK CONFUSION REVEALED 1307

“restoring previously nonfunctional antique watch movements and

parts,” as good faith.124 The court used this reasoning despite the
defendant’s intent to benefit from displaying the plaintiff’s mark,
though intending to capitalize on its historical significance rather than
its modern-day reputation.125
Others courts stand ready to pin the defendant down on a lower
negligence standard for failure to exercise due diligence.126 Yet some
will find against the defendant on an attempt standard, even without
proof that actual confusion resulted from it,127 and “some courts find
evidence of bad faith even where they conclude the defendant did not
choose its mark purposely to promote confusion.”128 Unsurprisingly,
these courts emphatically state that a defendant’s lack of intent is
generally not relevant to the likelihood of consumer confusion.129 Yet,
surprisingly, the presence of intent may not be decisive either. For
example, in one case, the court expressed that even when there is
explicit evidence of bad faith, that factor alone should not determine
the outcome of a case.130 Instead, a defendant’s “[b]ad faith and intent
to deceive are relevant to the extent that they add to the likelihood

little bearing on the ultimate question: whether the allegedly infringing product is
likely to confuse consumers.”); QuikTrip W., Inc. v. Weigel Stores, Inc., 984 F.3d 1031,
1036 (Fed. Cir. 2021) (“[T]he ‘only relevant intent is intent to confuse. There is a
considerable difference between an intent to copy and an intent to deceive.’”).
124. Hamilton Int’l Ltd. v. Vortic LLC, 486 F. Supp. 3d 657, 667–68 (S.D.N.Y. 2020),
aff’d, 13 F.4th 264, 268 (2d Cir. 2021) (“The Court credits this testimony, concluding
that he did not intend to cause consumer confusion but rather sought to ‘preserve
American history’ by salvaging and restoring the hearts of antique pocket watches.”).
125. Id. at 668; see also Champion Spark Plug Co. v. Sanders, 331 U.S. 125, 130
(1947) (noting that it can be “wholly permissible” that the “second-hand dealer gets
some advantage from the trademark”).
126. AWGI, LLC v. Atlas Trucking Co., 998 F.3d 258, 268 (6th Cir. 2021) (finding
the intent factor irrelevant); see also Ironhawk Techs., Inc. v. Dropbox, Inc., 994 F.3d
1107, 1124 (9th Cir. 2021), amended by 2 F.4th 1150 (9th Cir. 2021) (“This factor ‘favors
the plaintiff where the alleged infringer adopted his mark with knowledge, actual or
constructive, that it was another’s trademark.’”).
127. ServPro Intell. Prop., Inc. v. Blanton, 451 F. Supp. 3d 710, 727 (W.D. Ky. 2020)
(“Courts have held that ‘[i]f a party chooses a mark with the intent of causing
confusion, that fact alone may be sufficient to justify an inference of confusing
similarity.’” (emphasis added) (quoting Homeowners Grp., Inc. v. Home Mktg.
Specialists, Inc., 931 F.2d 1100, 1111 (6th Cir. 1991))).
128. Equitable Nat’l Life Ins. Co. v. AXA Equitable Life Ins. Co., 434 F. Supp. 3d
1227, 1248 (D. Utah 2020).
129. ServPro Intell. Prop., Inc, 451 F. Supp. 3d, at 727; Variety Stores, Inc. v. Wal-Mart
Stores, Inc., 888 F.3d 651, 665 (4th Cir. 2021) (“Bad faith may be inferred from
[Walmart’s] actual or constructive knowledge of [Variety’s] mark.”).
130. Car-Freshner Corp. v. Am. Covers, LLC, 980 F.3d 314, 333 (2d Cir. 2020).
1308 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

that the accused infringer will achieve its objective of consumer

confusion.”131
Beverly Pattishall suggested that factoring in intent makes outcomes
more predictable.132 This inference makes it easier to determine the
state of mind of one person, the defendant, than to forecast the
perceptions of the consumer group.133 Predictability is good, but the
result may not be. Intent inherently focuses on the wrong goalpost.
Merely because the defendant’s mental state is easier to discern than
the perception of the consuming public does not make that factor
more relevant to the inquiry. As Kelly Collins warned, “[t]his is
dangerous because mere ‘copying’ is not always impermissible.”134 The
law encourages reusing generic or functional marks “as a part of our
competitive economic system.”135 For this reason, she argues that the
relevant intent is the one to confuse and not merely to copy.136
Another reason to abandon intent is that it muddies jurisprudential
waters caused by further fusion of trade name and technical trademark
jurisprudence. Courts typically require intent when dealing with
non-inherently distinctive marks.137 Courts have either presumed
intent or dispensed with it for inherently distinctive marks.138
Alejandro Mejías explained that intent is irrelevant because the
focus “is not what the defendant intended to do, but whether his mark
is likely to be confusingly similar for the relevant public.”139 Very few
courts acknowledge this much.140 Judges may like intent because it
makes their job easier, and the outcome feels more just. However,

131. Id.
132. See Beverly W. Pattishall, The Impact of Intent in Trade Identity Cases, 60
TRADEMARK REP. 575, 579–80 (1970).
133. Id. at 577.
134. Kelly Collins, Comment, Intending to Confuse: Why Preponderance Is the Proper
Burden of Proof for Intentional Trademark Infringements Under the Lanham Act, 67 OKLA. L.
REV. 73, 87 (2014).
135. Id.
136. Id. at 87–88 (“This would better serve the purposes of the Lanham Act and
safeguard innocent conduct from triggering liability.”).
137. But see 4 MCCARTHY, supra note 2, § 23:106 (explaining that proof of intent is
merely evidence relevant to whether confusion is likely).
138. See Cashmere & Camel Hair Mfrs. Inst. v. Saks Fifth Ave., 284 F.3d 302, 317 (1st
Cir. 2002) (explaining that where there is evidence of intentional deceit the
presumption is clear); Res. Devs., Inc. v. Statue of Liberty-Ellis Island Found., Inc., 926
F.2d 134, 140 (2d Cir. 1991).
139. Mejías, supra note 106, at 349.
140. See, e.g., Virgin Enters. Ltd. v. Nawab, 335 F.3d 141, 151 (2d Cir. 2003)
(explaining that intent is not “of high relevance to the issue of likelihood of confusion”
because “[i]t does not bear directly on whether consumers are likely to be confused”).
2022] TRADEMARK CONFUSION REVEALED 1309

intent is irrelevant to technical trademark infringement. Recall from

Section I.B that technical trademark infringement focuses on the
consequences of the defendant’s act and not on their intent.141 In
contrast, trade name infringement focuses on defendants’ desired
outcomes, irrespective of consumer confusion.142
Pinning the likelihood of confusion on free-riding becomes
problematic because free-riding is ultimately a concept searching for
meaning.143 The Act does not require proof of intent. Trademark
infringement is, after all, a strict liability offense.144 As the Sixth Circuit
opined, the better view is to consider intent only after other likelihood
of confusion factors indicate liability.145 Intent may go to aggravated
remedies, but it should be irrelevant to the question of guilt. As Beebe
put it, “if trademark law seeks to prevent commercial immorality, then
it should do so explicitly. An injunction should issue, and damages be
granted on that basis alone, and not on the basis of possibly distorted
findings of fact as to the likelihood of consumer confusion.”146
Beebe found that despite the disconnect between the defendant’s
intent and consumer confusion, it stampedes the other factors.147 The
effect is powerful—a “nearly un-rebuttable presumption of a
likelihood of confusion”148 roughly 97% of the time,149 making it
“arguably the single most important confusion factor in use today.”150

141. Rogers, supra note 68, at 178 (explaining the origins of trademark law).
142. See, e.g., Visible Sys. Corp. v. Unisys Corp., 551 F.3d 65, 73 (1st Cir. 2008)
(explaining that sometimes there is a likelihood of confusion in industries regardless
of intent).
143. See, e.g., Ty Inc. v. Perryman, 306 F.3d 509, 512 (7th Cir. 2002) (rejecting
sponsorship dilution claim because “in that attenuated sense of free riding, almost
everyone in business is free riding”).
144. See Taubman Co. v. Webfeats, 319 F.3d 770, 775 (6th Cir. 2003) (recognizing
that the Lanham Act is a “strict liability statute”); see also Rebecca Tushnet, Running the
Gamut From A to B: Federal Trademark and Federal False Advertising Law, 159 U. PA. L. REV.
1305, 1310 (2011) (noting that federal courts have interpreted trademark as a strict
liability offense); Robert G. Bone, Enforcement Costs and Trademark Puzzles, 90 VA. L. REV.
2099, 2109 (2004) (referring to trademark infringement as a form of strict liability).
145. See, e.g., Taubman Co., 319 F.3d at 775 (“[T]he proper inquiry is not one of
intent. In that sense, the Lanham Act is a strict liability statute. If consumers are
confused by an infringing mark, the offender’s motives are largely irrelevant.”).
146. Beebe, supra note 6, at 1631.
147. Id. at 1621.
148. Id. at 1628.
149. Id. (“[The data] suggest that a finding of bad faith intent creates, if not in
doctrine, then at least in practice, a nearly un-rebuttable presumption of a likelihood
of confusion.”).
150. Blake Tierney, Missing the Mark: The Misplaced Reliance on Intent in Modern
Trademark Law, 19 TEX. INTELL. PROP. L.J. 229, 236 (2011).
1310 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

This Article reports that intent appeared in two-thirds of the cases

studied and was deemed neutral 19% of the time. In 27% of all cases,
courts favored plaintiffs on the intent factor. When they did, plaintiffs
won 52% of the time. In 14% of all cases, the courts favored defendants
on the intent factor. When the court favored defendants, they won
65% of the time.

Figure 5: Intent by Outcome

Qualitatively, cases in the dataset show that intent bears a minimal

impact on results.151 The reason is that “an intent to confuse customers
is not required for a finding of trademark infringement.”152
Nonetheless, the intensely fact-specific nature of intent can trip up
parties seeking speedy resolution of the dispute. As one court in the
dataset put it, “[i]ssues of bad or good faith ‘are generally ill-suited for
disposition on summary judgment.’”153 The practice is longstanding,

151. Equitable Nat’l Life Ins. Co. v. AXA Equitable Life Ins. Co., 434 F. Supp. 3d
1227, 1248 (D. Utah 2020) (“Although this factor weighs in AXA’s favor, its impact is
minimal.”); GoTo.com, Inc. v. Walt Disney Co., 202 F.3d 1199, 1208 (9th Cir. 2000)
(emphasizing “the minimal importance of the intent factor”).
152. GoTo.com, 202 F.3d at 1208 (citation omitted).
153. RVC Floor Decor, Ltd. v. Floor & Decor Outlets of Am., Inc., 527 F. Supp. 3d
305, 327 (E.D.N.Y. 2021); see also Zamfir v. Casperlabs, LLC, 528 F. Supp. 3d 1136, 1145
(S.D. Cal. 2021) (“These unresolved factual questions complicate the issue of
Defendant’s intent in choosing the mark.”).
2022] TRADEMARK CONFUSION REVEALED 1311

with courts preferring to leave it to juries to settle the matter.154

Surprisingly, at least one court insisted on a jury trial even when the
marks in question were identical due to the inherently subjective
nature of the inquiry.155
The dataset shows that 6% of cases expressly precluded summary
judgment based on the intent factor. That figure may seem low, but it
is considerably higher than any of the other factors: mark similarity
(4%), buyer sophistication (3%), actual confusion (2%), mark
strength (1%), competitive proximity (1%), “bridging the gap” (0%),
and quality (0%).
Eliminating intent allows a more focused inquiry into the likelihood
of confusion rather than the commercial immorality of defendants. As
a practical matter, it frees parties from costly discovery and allows the
court to grant summary judgment more frequently.156 Judges can also
dispose of cases more easily without trial, and it is less likely that
defendants will be subject to vexatious suits based on the nebulous
aspersions of intent.157
While the “ordinary consumer” is central to the infringement
analysis, it remains poorly theorized.158 In patent cases, courts benefit
from expert testimony.159 Perhaps this is because the subject matter of
patent disputes is by nature technologically challenging, defaulting
those involved to accept, even expect, expert assistance. But trademark
courts must investigate confusion without evidence that any consumers

154. See EMI Catalogue P’ship v. Hill, Holliday, Connors, Cosmopulos Inc., 228 F.3d
56, 67–68 (2d Cir. 2000) (“Because the issue goes to defendants’ intent, it ‘is best left
in the hands of the trier of fact.’”).
155. Tiffany & Co. v. Costco Wholesale Corp., 971 F.3d 74, 88 (2d Cir. 2020) (“And
as we have consistently observed, ‘subjective issues such as good faith are singularly
inappropriate for determination on summary judgment.’”).
156. 10B CHARLES ALAN WRIGHT ET AL., FEDERAL PRACTICE AND PROCEDURE (CIVIL))
§ 2730 n.3 (3d ed. 2015) (“Questions of intent, which involve intangible factors
including witness credibility, are matters for the consideration of the fact finder after
a full trial and are not for resolution by summary judgment.”).
157. Thomas L. Casagrande, A Verdict for Your Thoughts? Why an Accused Trademark
Infringer’s Intent Has No Place in Likelihood of Confusion Analysis, 101 TRADEMARK REP.
1447, 1455 (2011) (proposing an elimination of intent as a factor to be considered in
determining trademark infringement).
158. See e.g., Thomas R. Lee et al., Trademarks, Consumer Psychology, and the
Sophisticated Consumer, 57 EMORY L.J. 575, 575 (2008) (“[N]either courts nor
commentators have made any serious attempt to develop a framework for
understanding the conditions that may affect the attention that can be expected to be
given to a particular purchase.”).
159. Douglas G. Smith, The Increasing Use of Challenges to Expert Evidence Under
Daubert and Rule 702 in Patent Litigation, 22 J. INTELL. PROP. L. 345, 354 (2015).
1312 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

were confused, imagining consumers’ likely experience as filtered

through the parties’ competing interests. This notional consumer is
“neither savant nor dolt.”160 One who “lacks special competency with
reference to the matter at hand but has and exercises a normal
measure of the layman’s common sense and judgment.”161 Instead,
courts rely on surveys, mark strength, and consumer sophistication to
determine the likelihood of confusion. But, like intent, none of these
factors provide a good proxy. The Sections below explain why.

2. Surveys
Surveys attempt to measure whether consumers believe that the
plaintiff’s mark is the source of the alleged infringer’s product or
whether it sponsors or approves it.162 Plaintiffs may provide survey
evidence that an appreciable number of relevant consumers are likely
to be confused.163 According to a case in the dataset, survey evidence is
not a prerequisite for establishing public recognition, but “it is the
most persuasive evidence of it.”164
Surveys present respondents with defendants’ marks and measure
consumers’ reactions in the context that consumers encounter the
mark in question.165 Proof of marketing supports broad public
recognition.166 They typically involve control groups to show causality
between the defendants’ mark and consumer confusion.167
In theory, a survey needs to pass muster under the Federal Rules of
Evidence, which requires considering the “validity of the techniques
employed.”168 Courts can bar significantly flawed surveys as evidence

160. Kraft Foods Grp. Brands LLC v. Cracker Barrel Old Country Store, Inc., 735
F.3d 735, 743 (7th Cir. 2013).
161. United States v. 88 Cases, More or Less, Containing Bireley’s Orange Beverage,
187 F.2d 967, 971 (3d Cir. 1951).
162. 3 ANNE GILSON LALONDE, GILSON ON TRADEMARKS § 8.03 (2021).
163. See 6 MCCARTHY, supra note 2, § 32:158.
164. Kibler v. Hall, 843 F.3d 1068, 1074 (6th Cir. 2016).
165. Shari Seidman Diamond & David J. Franklyn, Trademark Surveys: An Undulating
Path, 92 TEX. L. REV. 2029, 2037 (2014).
166. Therma-Scan, Inc. v. Thermoscan, Inc., 295 F.3d 623, 632 (6th Cir. 2002).
167. See, e.g., Bracco Diagnostics, Inc. v. Amersham Health, Inc., 627 F. Supp. 2d
384, 448 (D.N.J. 2009) (criticizing a survey’s design for failure to use “a control
mechanism”).
168. Shari Seidman Diamond, Reference Guide on Survey Research, in REFERENCE
MANUAL ON SCI. EVIDENCE 359, 364 (3d ed. 2011).
2022] TRADEMARK CONFUSION REVEALED 1313

when they are more prejudicial than probative169 or deemed

unreliable.170
The problem is that commentators and courts alike acknowledge
that surveys are often unreliable and expensive, costing hundreds of
thousands of dollars.171 Courts routinely attack the representativeness
of the survey from a parade of cherry-picked witnesses and extrapolate
a standard of what consumers generally believe.172 The inexact science
of assessing trademark strength causes judges to rely upon or reject
surveys based on whether the results agree with their subjective
impressions.173 As a result, judicial unease with surveys sometimes
bubbles to the surface, with Judge Richard Posner remarking that “no
doubt there are other tricks of the survey researcher’s black arts that
we have missed.”174
Constructing a robust survey is dauntingly hard. Surveys need to
employ a control175 and calculate noise.176 As an indication of the

169. Citizens Fin. Group, Inc. v. Citizens Nat’l Bank, 383 F.3d 110, 120 (3d Cir.
2004).
170. Id.
171. Nabisco, Inc. v. PF Brands, Inc., 191 F.3d 208, 224 (2d Cir. 1999)
(“[C]onsumer surveys . . . are expensive, time-consuming and not immune to
manipulation.”); Johnson v. Revenue Mgmt. Corp., 169 F.3d 1057, 1063 (7th Cir.
1999) (“[S]urvey evidence in trademark and trade dress cases can be very costly.”); see
also Robert H. Thornburg, Trademark Survey Evidence: Review of Current Trends in the
Ninth Circuit, 21 SANTA CLARA COMPUT. & HIGH TECH. L.J. 715, 717 (2005) (“[T]he most
basic of surveys cost[s] in the hundreds of thousands of dollars.”).
172. But cf. Citizens Fin. Grp., 383 F.3d at 122 (“In general, ‘actual confusion’
evidence collected by employees of a party in a trademark action must be viewed with
skepticism because it tends to be biased or self-serving.”).
173. 6 MCCARTHY, supra note 2, § 32:196 (“Since an estimation of the probable
mental reactions and associations of the buying public is not a science, there is always
the temptation to decide on the basis of a ‘hunch.’ That is, the trier of fact (or any
human being) would rather extrapolate from his or her own subjective impressions
than extrapolate from some hard evidence of other persons’ subjective impressions--
especially if the two do not agree.”); see also Peter Weiss, The Use of Survey Evidence in
Trademark Litigation: Science, Art or Confidence Game?, 80 TRADEMARK REP. 71, 83 (1990)
(“[A] reading of the many cases in which either great weight or little weight was given
to survey evidence will, I feel reasonably certain, lead most objective analysts to the
conclusion that, while some surveys went down because they were indeed ‘seriously
flawed,’ many others either stayed up or went down depending on the result which the
judges wanted to reach.”).
174. Indianapolis Colts, Inc. v. Metro. Balt. Football Club Ltd. P’ship, 34 F.3d 410,
416 (7th Cir. 1994).
175. 6 MCCARTHY, supra note 2, § 32:187 (“Courts have held that a survey that fails
to use a control may be given less weight or even excluded from evidence altogether.”).
176. DANIEL KAHNEMAN, OLIVER SIBONY & CASS SUNSTEIN, NOISE: A FLAW IN HUMAN
JUDGMENT 488 (2021) (“Noise is variability in judgments that should be identical.”).
1314 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

treacherousness of this task, one court dismissed an expert witness who

authored a book on the very subject of conducting trademark surveys
for producing a “useless” survey.177 As a matter of justice between the
parties, the staggering costs of surveys put defendants at a
disadvantage. Robert Bone explained that “[p]roving a high
[likelihood of confusion] puts a premium on surveys and expert
testimony and is likely to require extensive discovery, all of which will
increase direct litigation costs and strengthen a trademark owner’s
ability to leverage cease-and-desist threats in frivolous and weak
cases.”178
Qualitatively, cases in the dataset warn that surveys only represent
circumstantial evidence of actual confusion, providing an experimental
environment, not real consumers making mistaken purchases.179 As one
court put it, “[a]necdotal evidence can be more direct evidence of
actual confusion and so is ‘both relevant and probative.’”180 Another
court “noted a trend away from according great weight to survey
evidence,” and afforded the survey no weight.181 Unlike actual
confusion, that court explained that “survey evidence is circumstantial,
not direct, evidence of the likelihood of confusion. Surveys do not
measure the degree of actual confusion by real consumers making
mistaken purchases.”182 Accordingly, the court faulted the survey for
“depart[ing] from real-market conditions in a way that was both biased
and misleading.”183

177. Vital Pharms., Inc. v. Monster Energy Co., No. 19-60809-CIV, 2021 WL
3371942, at *65 (S.D. Fla. Aug. 3, 2021).
178. See Bone, Notice Failure and Defenses, supra note 11, at 1269 n.110.
179. Vital Pharms., Inc., 2021 WL 3371942, at *55 (“[S]urvey evidence is not direct
evidence of customer confusion in the real marketplace.”).
180. Juul Labs, Inc. v. 4X PODS, 509 F. Supp. 3d 52, 66 (D.N.J. 2020).
181. Vital Pharms., Inc., 2021 WL 3371942, at *55.
182. Id. at *64.
183. Id. at *65 (“[E]vidence at trial confirmed the obvious: that the artificial coolers
Mr. Berger showed his survey participants looked nothing like the coolers consumers
would encounter in real stores.”); see, e.g., Citizens Banking Corp. v. Citizens Fin. Grp.,
Inc., 320 F. App’x 341, 348 n.4 (6th Cir. 2009) (minimizing the weight of a confusion
survey because it “failed to mimic the purchase conditions”); Coherent, Inc. v.
Coherent Techs., Inc., 935 F.2d 1122, 1126 (10th Cir. 1991) (affirming the district
court’s finding that the “survey did not show actual confusion because it failed to
simulate decisions in the marketplace”); Am. Footwear Corp. v. Gen. Footwear Co.,
609 F.2d 655, 661 (2d Cir. 1979) (finding that “the critical defect in this survey was the
failure to conduct it under actual marketing conditions”—and so the “district court’s
rejection of this survey evidence was not clearly erroneous”); 4 MCCARTHY, supra note
2, § 23:2.50 (stating that a survey is only evidence of confusion if “the survey mirrors
the real world setting which can create an instance of actual confusion”).
2022] TRADEMARK CONFUSION REVEALED 1315

Scholars also warn against placing a premium on surveys. According

to Beebe, “the conventional view of the utility of survey evidence may
be incorrect”: only 20% of the cases he reviewed addressed survey
evidence, 10% credited survey evidence, and 7% ruled in favor of the
outcome that the credited survey evidence favored.184 This dataset
shows a near-identical result fifteen years later. Of the 20% of cases that
addressed survey evidence, 12% credited survey evidence, and 6% ruled
in favor of the outcome that the credited survey evidence favored.

Figure 6: Outcome by Survey Evidence

As with intent, there is a certain circular irony to the whole exercise

regarding surveys. Courts rely on surveys only to support conclusions
that judges reach using other factors. The analysis also works
backward—faced with survey evidence showing a likelihood of
confusion, judges may regard the marks as more similar than they
might have appeared in the absence of the survey.185 As Peter Weiss
remarked, “[o]ne might sum it all up by saying that the function of
surveys in trademark litigation is to plumb the minds of the public to
make up the minds of the judges.”186 Dispensing with surveys and

184. Beebe, supra note 6, at 1641. A more recent study that expanded Beebe’s
dataset found that only about 17% of cases addressed survey evidence. See Robert C.
Bird & Joel H. Steckel, The Role of Consumer Surveys in Trademark Infringement: Empirical
Evidence from the Federal Courts, 14 U. PA. J. BUS. L. 1013, 1035 (2012).
185. Diamond & Franklyn, supra note 165, at 2043.
186. Weiss, supra note 173, at 86.
1316 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

relying on the court’s judgment would not only be cheaper and

simpler, but it would also be the intellectually honest thing to do.
Surveys sometimes overlap with trademark strength since parties
may use the former to measure the potency of a mark’s goodwill and
its worthiness of protection.187 Known as the Abercrombie spectrum,
generic and descriptive marks are not distinctive, suggestive marks are
marginally distinctive, while arbitrary or fanciful marks are inherently
distinctive.188 Trademark strength is usually the first factor courts
consider.189

3. Mark strength
A mark’s distinctiveness is its uniqueness in denoting a product.
Marks may be fanciful, arbitrary, suggestive, descriptive, or generic
from most to least distinctive.190 Generic terms are unprotectable and
descriptive ones are protectable only when buyers view them as
distinctive of a unique source.191 Evaluating the strength of a mark
requires the fact finder to evaluate several factors: its degree of
inherent distinctiveness, its “conceptual strength,” its distinctiveness in
the marketplace, and its “commercial strength.”192 Unlike conceptual
strength, commercial strength considers advertising expenditures,
consumer studies linking the mark to a source, sales success,
unsolicited media coverage of the product, attempts to plagiarize the
mark, and the length and exclusivity of the mark’s use.193
The dataset reveals that mark strength comes up in 70% of the cases.
In 47% of all cases, the courts favored plaintiffs on the mark strength
factor. When they did, plaintiffs won 46% of the time. In 15% of all

187. Beebe, supra note 6, at 1646 (“In trademark law, the question is always of
consumer perception in the marketplace rather than judicial perception in the
courtroom.”).
188. Abercrombie & Fitch Co. v. Hunting World, Inc., 537 F.2d 4, 9 (2d Cir. 1976).
189. See, e.g., Welding Servs., Inc. v. Forman, 509 F.3d 1351, 1361 (11th Cir. 2007)
(“The stronger or more distinctive a trademark or service mark, the greater the
likelihood of confusion . . . .”); Barton Beebe & C. Scott Hemphill, The Scope of Strong
Marks: Should Trademark Law Protect the Strong More Than the Weak?, 92 N.Y.U. L. REV.
1339, 1349 n.40 (2017) (“Strength is the first factor in the Second, Fourth, Fifth, Sixth,
Eighth, Ninth, and Eleventh Circuits, the second factor in the Third Circuit, and the
last factor in the First and Tenth Circuits.”). Courts consider design marks under the
Seabrook factors. See Seabrook Foods, Inc. v. Bar-Well Foods, Ltd., 568 F.2d 1342
(C.C.P.A. 1977).
190. U.S. Pat. & Trademark Off. v. Booking.com B. V., 140 S. Ct. 2298, 2302 (2020).
191. Id. at 2303.
192. Ouellette, supra note 40, at 353.
193. Variety Stores, Inc. v. Walmart Inc., 852 F. App’x 711, 719 (4th Cir. 2021).
2022] TRADEMARK CONFUSION REVEALED 1317

cases, the courts favored defendants on the mark strength factor.

When the court favored defendants, they won 71% of the time.

Figure 7: Mark Strength by Outcome

In an empirical study on mark strength, Lisa Ouellette observed that

“courts often have difficulty applying these tests.”194 According to her,
[t]he complex doctrine that has evolved around trademark strength
and the likelihood of confusion appears to be a (largely
unsuccessful) attempt to provide some analytical rigor to the
essential questions of how strongly a mark identifies goods or
services and how well it distinguishes those products from others in
the marketplace.195
Determining the bounds of an owner’s trademark requires more
than just looking at the mark; it requires assessing what protection the
trademark owner should be entitled to for that mark.196 Distinctive
marks are memorable as source indicators and possess greater
conceptual strength to consumers.197 Courts equate distinctiveness
with a greater breadth of protection, are more willing to find confusing

194. Ouellette, supra note 40, at 353.

195. Id. at 360.
196. See Virgin Enters. Ltd. v. Nawab, 335 F.3d 141, 148 (2d Cir. 2003).
197. See id. (noting that consumers are more likely to attribute two products with
more unique names to the same source versus two products with more generic names).
1318 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

similarities,198 and usually find that the strongest marks merit the
widest range of protection.199
Like the “black arts” of surveys, empirical studies confirm that courts
judge mark strength intuitively.200 For instance, Beebe reported how
courts failed to categorize the plaintiff’s mark in a specific category of
distinctiveness in half of the cases he studied.201 He observed that
“considerations such as the comparative quality of the parties’ goods
or the inherent distinctiveness of the plaintiff’s mark rarely aid in this
inquiry.”202 Others have variously criticized trademark strength as
“needlessly open-ended”203 and “inconsistent.”204 One court
acknowledged distinctiveness “is far from an exact science and that the
differences between the classes, which is not always readily
apparent . . . makes placing a mark in its proper context . . . tricky
business at best.”205
As with survey evidence, Thomas McCarthy notes, that
a cynic would say that . . . when the court wants to find no
infringement, it says that the average buyer is cautious and
careful . . . . But if the judge thinks there is infringement, the judge
sets the standard lower and says the average buyer is gullible and not
so discerning.206

198. See, e.g., First Sav. Bank, F.S.B. v. First Bank Sys., Inc., 101 F.3d 645, 655 (10th
Cir. 1996) (“When the primary term is weakly protected to begin with, minor
alterations may effectively negate any confusing similarity between the two marks.”).
199. See, e.g., Ford Motor Co. v. Money Makers Auto. Surplus, Inc., No. 03CV493,
2005 WL 2464715, at *1, *3 (D. Neb. Sept. 14, 2005) (finding that the various Ford
Motor Company marks at issue “are among the most famous marks in the world” and
are “therefore entitled to the widest scope of protection”).
200. See, e.g., Thomas R. Lee, Eric D. DeRosia & Glenn L. Christensen, Sophistication,
Bridging the Gap, and the Likelihood of Confusion: An Empirical and Theoretical Analysis, 98
TRADEMARK REP. 913, 913 (2008) (analyzing how courts rely on “precedent built on
‘personal intuition and subjective, internalized, stereotypes.’”); see also Beebe, supra
note 6, at 1581 (describing the variation among circuits in their application of
multifactor tests for likelihood of confusion).
201. Beebe, supra note 6, at 1633–35 (stating that some use of the spectrum was
made in only 193 out of 331 cases and that the mark was placed in a specific category
in only 164 cases).
202. Id. at 1645.
203. Timothy Denny Greene & Jeff Wilkerson, Understanding Trademark Strength, 16
STAN. TECH. L. REV. 535, 582 (2013).
204. Beebe, supra note 6, at 1633.
205. Banff, Ltd. v. Federated Dep’t Stores, Inc., 841 F.2d 486, 489 (2d Cir. 1988).
206. 4 MCCARTHY, supra note 2, § 23:92; see also Ann Bartow, Likelihood of Confusion,
41 SAN DIEGO L. REV. 721, 747 (2004) (noting that judges give meaning to terms on a
“case-by-case” basis).
2022] TRADEMARK CONFUSION REVEALED 1319

The courts themselves regard likelihood of confusion merely “as a

heuristic device to assist in determining whether confusion exists.”207
There is no requirement for the likelihood of confusion to consider
survey evidence or mark strength. Eliminating both would simplify the
likelihood of confusion and make it less prone to error.

4. Consumer sophistication
Consumer sophistication provides context to the consumer
information available and the ability of consumers to discern between
the marks.208 Courts analyze the degree of care reasonably expected of
potential customers from the perspective of “the ordinary purchaser,
buying under the normally prevalent conditions of the market and
giving the attention such purchasers usually give in buying that class of
goods.”209 More expensive products or services mean consumers take
more time and effort when making decisions, and therefore, the
likelihood of confusion decreases.210 However, the defendant’s
distribution methods may affect consumers’ degree of care, even when
an individual product is not expensive.211
Scholars criticized the artificiality of consumer sophistication,
likening it to expecting judges to perform a “Vulcan mind-meld” with
consumers in the marketplace.212 Courts may easily project their
normative view of how careful a consumer should be or their view of a

207. Sullivan v. CBS Corp., 385 F.3d 772, 778 (7th Cir. 2004).
208. Andrew Martineau, Imagined Consumers: How Judicial Assumptions About the
American Consumer Impact Trademark Rights, for Better and for Worse, 22 DEPAUL J. ART,
TECH. & INTELL. PROP. L. 337, 352 (2012) (“This would seem to be a crucial part of the
test, given that the standard for infringement is whether consumers are likely to be
confused.”).
209. Gen. Mills, Inc. v. Kellogg Co., 824 F.2d 622, 627 (8th Cir. 1987).
210. Kibler v. Hall, 843 F.3d 1068, 1080 (6th Cir. 2016) (observing that when
consumers exercise caution in purchasing items, they are less likely to confuse their
origins, such as “when consumers have expertise in the items and when the items are
particularly expensive”).
211. See, e.g., ZW USA, Inc. v. PWD Sys., LLC, 889 F.3d 441, 447–48 (8th Cir. 2018)
(finding fact that parties sold their respective low-cost products on different websites
under different trade names strongly cut against a likelihood of confusion).
212. See William E. Gallagher & Ronald C. Goodstein, Inference Versus Speculation in
Trademark Infringement Litigation: Abandoning the Fiction of the Vulcan Mind Meld, 94
TRADEMARK REP. 1229, 1230 (2004) (criticizing the lack of empirical evidence required
to validate an inference of likelihood of confusion, such as no requirement for
consumer surveys or evidence showing actual confusion).
1320 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

defendant’s conduct.213 But, like intent, surveys, and mark strength,

consumer sophistication suffers from inherent capriciousness.
The dataset reveals that consumer sophistication comes up in 46%
of the cases, among the lowest of all the Polaroid factors. In 18% of all
cases, the courts favored plaintiffs on the consumer sophistication
factor. When they did, plaintiffs won 49% of the time. In 14% of all
cases, the courts favored defendants on the consumer sophistication
factor. When this factor favored defendants, they won 63% of the time.

Figure 8: Buyer Sophistication by Outcome

Three irrelevant factors are plenty, but there is one final culprit.
That is, the sheer multitude of factors courts must consider. The total
number of factors makes the likelihood of confusion analysis difficult
to deploy, bogging down courts and encouraging selective application.
Instead, judges and juries rely on coherence-based reasoning to make
sense of their findings to cope with the sheer number of factors.

D. Coherence-Based Reasoning
Over the past century, trademark law ossified the likelihood of
confusion standard from pragmatic judge-made rules of thumb into a
rigid and formalistic standard. The Restatement (First) of Torts merely

213. August Storck K.G. v. Nabisco, Inc., 59 F.3d 616, 618 (7th Cir. 1995) (“Many
consumers are ignorant or inattentive, so some are bound to misunderstand no matter
how careful a producer is.”).
2022] TRADEMARK CONFUSION REVEALED 1321

mentioned “the following factors are important,” and the early cases
applied the factors loosely.214 However, appeals courts chastised lower
courts for failing to address each factor, with orders to reverse and
remand.215 We can deduce this formalism ended up burdening courts
with an unwieldy craft, forcing judges to pay lip service to all the factors
while systemically relying on only a few. At the same time, their
opinions recite disclaimers that the likelihood of confusion factors act
only as a guide and that no single factor is dispositive.
Studies show that experts do not integrate multifactor test (“MFT”)
factors well.216 Even using stringent tests to aid in decision-making can
lead to consistent and predictable mistakes.217 It may occur early in the
decision-making process, and a single attribute can trigger coherence-based
reasoning.218
Trademark law expects courts to decipher between six and thirteen
likelihood of confusion factors, which often point in opposite
directions, yet still reach a coherent conclusion in every case.219 Worse,
the likelihood of confusion factors in each circuit are not exhaustive,
with courts occasionally considering other factors such as geographical
proximity.220
Courts are divided on whether “it is incumbent upon the district judge
to engage in a deliberate review of each factor.”221 Some emphatically state

214. RESTATEMENT (FIRST) OF TORTS § 729 (AM. L. INST. 1939).

215. Barton Beebe, An Empirical Study of the Multifactor Tests for Trademark
Infringement, 94 CALIF. L. REV. 1581, 1593 (2006).
216. See, e.g., Robyn M. Dawes, The Robust Beauty of Improper Linear Models in Decision
Making, 34 AM. PSYCH. 571, 573 (1979) (positing that experts in a field are better at
selecting and coding information than integrating it).
217. See Chris Guthrie et al., Inside the Judicial Mind, 86 CORNELL L. REV. 777, 779–
80 (2001).
218. See Dan Simon, Daniel C. Krawczyk, & Keith J. Holyoak, Construction of
Preferences by Constraint Satisfaction, 15 PSYCH. SCI. 331, 331 (2004) (suggesting that a
single variable can initiate spreading coherence).
219. See Tana v. Dantanna’s, 611 F.3d 767, 775 n.7 (11th Cir. 2010) (explaining that
this test “presupposes that various factors will point in opposing directions”; it is the
job of the Court to determine the relative importance of the evidence probative of
each factor in an effort to decide whether, “in light of the evidence as a whole, there
is sufficient proof of a likelihood of confusion to warrant a trial of the issue”).
220. See id. at 781 (holding that “[t]he district court did not err in considering the
geographic proximity of use as an eighth factor demonstrating the unlikelihood of
confusion”); see also J-B Weld Co. v. Gorilla Glue Co., 978 F.3d 778, 789 (11th Cir. 2020)
(“While all seven factors must be considered, they are not necessarily exhaustive if
other evidence is probative of a likelihood of confusion.”).
221. Compare Arrow Fastener Co. v. Stanley Works, 59 F.3d 384, 400 (2d Cir. 1995)
(discussing each factor), with Bumble Bee Seafoods LLC v. UFS Indus., Inc., No. 04
1322 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

that “the factors are not truly independent—depending on the context, a

strong showing as to one factor may serve to make a different factor more
or less important.”222 Yet others rule only on a few key factors, allowing
them to resolve the dispute without needing a trial.223
Without meaningful guidance, courts weigh those factors
impressionistically. Beebe’s study confirms that judges in the
likelihood of confusion cases employ “‘fast and frugal’ heuristics to
short-circuit the multifactor analysis.”224 Coherence-based reasoning
operates bidirectionally to fit together how a judge decides the
factors,225 both preceding the decision and in forming its basis.226 In
other words, fact-finders assessing a likelihood of confusion test will
look at the evidence as non-independently relative to the final
decision.227 Consequently, the resulting decision is biased because, as
Dan Simon explains, “the hard case morphs into an easy one” in the
mind of the fact-finder.228
Formulating optimal legal rules requires judges to balance factors
while taking account of “false positive” errors (i.e., prohibiting
beneficial conduct) versus “false negative” errors (i.e., permitting
harmful conduct).229 This task requires judges to access information
on the frequency and impact of the error, the likelihood of deterrence,

Civ. 2105, 2004 WL 1637017, at *7 (S.D.N.Y. July 20, 2004) (discussing only the relevant
factors).
222. Select Comfort Corp. v. Baxter, 996 F.3d 925, 934 (8th Cir. 2021).
223. See, e.g., Hamilton Int’l Ltd. v. Vortic LLC, 486 F. Supp. 3d 657, 666 (S.D.N.Y.
2020) (“As the Court noted in its summary judgment opinion, a number of the Polaroid
factors are not helpful to this case.”), aff’d, 13 F.4th 264 (2d Cir. 2021).
224. Beebe, supra note 6, at 1581; Tierney, supra note 150, at 235–36 (“[M]uch of
the time spent going through the list of factors in any given case is in reality just an
attempt to justify a predetermined conclusion about the likelihood of
confusion . . . .”).
225. Dan Simon, A Third View of the Black Box: Cognitive Coherence in Legal Decision
Making, 71 U. CHI. L. REV. 511, 515–16 (2004) [hereinafter Simon, Third View of the
Black Box].
226. See, e.g., Dan Simon et al., The Redux of Cognitive Consistency Theories: Evidence
Judgments by Constraint Satisfaction, 86 J. PERSONALITY & SOC. PSYCH. 814, 816 (2004).
227. Dan Simon, The Limited Diagnosticity of Criminal Trials, 64 VAND. L. REV. 143, 195
(2011).
228. Simon, Third View of the Black Box, supra note 225, at 517 (describing studies
where coherence-based reasoning caused subjects who found for the defendant and
those who found for the plaintiff to be more confident the evidence supported their
view after they had issued their verdict).
229. Andrew I. Gavil & Steven C. Salop, Probability, Presumptions and Evidentiary
Burdens in Antitrust Analysis: Revitalizing the Rule of Reason for Exclusionary Conduct, 168
U. PA. L. REV. 2107, 2119 (2020).
2022] TRADEMARK CONFUSION REVEALED 1323

and the cost to the administrative process.230 On occasion, courts

themselves express frustration with the likelihood of confusion tests,
acknowledging that “[a]lthough our test for a likelihood of confusion
is well-developed, some uncertainty remains as to when confusion must
exist in order to support a trademark infringement claim.”231
The takeaway is that an overload of factors demands too much from
judges and forces them to stampede over those they deem less
significant. In the absence of direct evidence of confusion, courts must
ascertain it through a host of proxy factors.232 Under these trying
circumstances, Beebe empirically observed intent and actual confusion
playing an outsized role in coloring how courts treated the other
likelihood of confusion factors, confirming their perniciousness.233
As Michael Grynberg noted, “[e]ven if judges do no more than
applying heuristics of questionable quality to the disposition of
trademark claims, channeling the process through a consistent
framework aids litigants in identifying and accommodating the factors
that guide fact finding.”234 The question then is, how many factors
should we retain?
Only a few, argued Beebe, pointing out that judges in the likelihood
of confusion cases find only a few factors probative anyway.235 Indeed,
cases in the dataset recognize that courts can short circuit the process
and focus on just a few factors.236 Beebe recommended three or four

230. See id. at 2119–20 (asserting that formulating an optimal legal standard involves
considering error costs, deterrence, and administrative costs).
231. See, e.g., Select Comfort Corp. v. Baxter, 996 F.3d 925, 934 (8th Cir. 2021).
232. Laura A. Heymann, The Reasonable Person in Trademark Law, 52 ST. LOUIS U. L.J.
781, 783 (2008).
233. Beebe, supra note 6, at 1620–21 (“Intent and, to a lesser degree, actual
confusion appear to exert such a coherence-shifting influence when they favor a
likelihood of confusion. Indeed, in the forty-nine opinions in which both findings were
made, thirty-four (69%) of them found that all the factors favored a likelihood of
confusion.”).
234. Michael Grynberg, The Judicial Role in Trademark Law, 52 B.C. L. REV. 1283, 1305
(2011) [hereinafter Grynberg, Judicial Role in Trademark Law].
235. Beebe, supra note 6, at 1614 (“Like any human decision makers, district judges
attempt to decide both efficiently and accurately. In pursuit of efficiency, they consider
only a few factors. In pursuit of accuracy, they consider the most decisive factors.”).
236. Eclipse Assocs. Ltd. v. Data Gen. Corp., 894 F.2d 1114, 1118 (9th Cir. 1990)
(“These tests were not meant to be requirements or hoops that a district court need
jump through to make the determination.”); see also R.H. Donnelley Inc. v. USA
Northland Directories, Inc., No. Civ.04-4144, 2004 WL 2713248, at *6 (D. Minn. Nov.
19, 2004) (folding similarity and intent); Ironhawk Techs. v. Dropbox, Inc., 994 F.3d
1107, 1123–24 (9th Cir. 2021) (folding actual confusion and sophistication); CDOC,
1324 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

“core factors” informing “consumer perception in the marketplace

rather than judicial perception in the courtroom.”237 Alejandro Mejías
went further, recommending just two—similarity of marks and
proximity of goods, as “adding any other relevant factors, instead of
using unmanageable and misguiding large lists of factors that are
extremely difficult to balance, seems to be more in line with the thesis
of scientific research on decision-making.”238 The next Part explains
why actual confusion, mark similarity, and competitive proximity
should form the core factors and why these factors, together with fair
use safe harbors for expressive and descriptive uses, should form the
rules of thumb in trademark law.

II. RULES OF THUMB

Mark similarity, goods and services, and evidence of actual confusion
anchor the likelihood of confusion test as the most relevant factors.239
Jurisprudence supports that view. In one case from the dataset, the
Ninth Circuit has described a “trinity [that] constitutes the most
crucial body of the Sleekcraft analysis”—mark similarity, goods/services
similarity, and marketing and advertising channels.240 Safe harbors
protect core policies most in danger of being invaded by trademark
expansionism while making it simpler and cheaper for businesses to
do their due diligence and comply with the law.241 This Part explains
why.

A. Actual Confusion
Actual confusion is the most direct and decisive evidence of
confusion.242 Courts explain that where confusion occurred, it “is of

Inc. v. Liberty Bankers Life Ins. Co., 844 F. App’x 357, 362 (Fed. Cir. 2021) (folding
competitive proximity and sophistication).
237. Beebe, supra note 6, at 1646.
238. Mejías, supra note 106, at 348 (concentrating the analysis on the main two
factors).
239. See infra Part III.
240. GoTo.com, Inc. v. Walt Disney Co., 202 F.3d 1199, 1207 (9th Cir. 2000); see also
Stone Creek, Inc. v. Omnia Italian Design, Inc., 862 F.3d 1131, 1136 (9th Cir. 2017)
(“Two particularly probative factors are the similarity of the marks and the proximity
of the goods.” (citing Lindy Pen Co. v. Bic Pen Corp., 796 F.2d 254, 256–57 (9th Cir.
1986))).
241. See infra Part II.
242. Beebe, supra note 6, at 1608 (finding a 92% plaintiff success rate where the
court found actual confusion); Tana v. Dantanna’s, 611 F.3d 767, 779 (11th Cir. 2010)
(“The last factor, actual confusion in the consuming public, is the most persuasive
evidence in assessing likelihood of confusion.”); see also Groeneveld Transp. Efficiency,
2022] TRADEMARK CONFUSION REVEALED 1325

course convincing evidence that confusion is likely to occur.”243 As a

policy lever, actual confusion gives courts the ability to anchor their
analysis in real-world characteristics. In addition, the evidence is
pre-existing, does not depend on the vagaries of survey design, and
should make it easier for courts to dispose of cases pretrial.244
The dataset reveals that actual confusion comes up in 74% of the
cases. In 32% of all cases, the courts favored plaintiffs on the actual
confusion factor. When they did, plaintiffs won 54% of the time. In
18% of all cases, the courts favored defendants on the consumer
sophistication factor. When the court favored defendants, they won
77% of the time.

Inc. v. Lubecore Int’l, Inc., 730 F.3d 494, 517 (6th Cir. 2013) (“Nothing shows the
likelihood of confusion more than the fact of actual confusion.”); Variety Stores, Inc.
v. Walmart Inc., 852 F. App’x 711, 720 (4th Cir. 2021) (“[A]ctual confusion, is the
‘most important factor’ . . . .”); John Benton Russell, New Tenth Circuit’s Standards:
Competitive Keyword Advertising and Initial Interest Confusion in 1-800 Contacts v.
Lens.com, 30 BERKELEY TECH. L.J. 993, 1000 (2015) (“[C]ourts across several circuits
view this as the strongest evidence a plaintiff can present in a trademark infringement
case.”); Mark D. Robins, Actual Confusion in Trademark Infringement Litigation:
Restraining Subjectivity Through a Factor-Based Approach to Valuing Evidence, 2 NW. J. TECH.
& INTELL. PROP. 117, 117 (2004) (“In a case where all other circumstances point to a
finding of non-infringement, significant evidence of actual confusion dramatically
alters the equation.”).
243. Morningside Grp. Ltd. v. Morningside Cap. Grp., L.L.C., 182 F.3d 133, 141 (2d
Cir. 1999).
244. I am grateful to Jon Lee for this insight.
1326 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

Figure 9: Actual Confusion by Outcome

At the bottom, the inquiry concerns whether there was confusion

that could lead to “a diversion of sales, damage to goodwill, or loss of
control over reputation.”245 For that reason, courts look for actual
confusion among “prospective purchasers of [plaintiff’s] products.”246
Relevant circumstances include the extent of the parties’ advertising,
the length of time the allegedly infringing product has been
advertised, or any other factor that might influence the likelihood that
actual confusion would be reported.247
Courts accept both anecdotal and survey evidence indicating actual
confusion.248 This Article explained in Section I.C.2 that courts should
avoid survey evidence in its current manifestation. As to anecdotal
evidence, there is no absolute number of instances of actual confusion
that must be met to win in a likelihood of confusion analysis. Rather,
courts look to the totality of the circumstances in evaluating the
evidence of actual confusion.249 For example, “[i]nquiries about the
relationship between an owner of a mark and an alleged infringer do

245. Reply All Corp. v. Gimlet Media, LLC, 843 F. App’x 392, 397 (2d Cir. 2021)
(quoting Lang v. Ret. Living Publ’g Co., 949 F.2d 576, 583 (2d Cir. 1991)).
246. Lang, 949 F.2d at 583; SLY Mag., LLC v. Weider Publ’ns L.L.C., 529 F. Supp.
2d 425, 441 (S.D.N.Y. 2007), aff’d, 346 F. App’x 721 (2d Cir. 2009).
247. Jellibeans, Inc. v. Skating Clubs Ga., Inc., 716 F.2d 833, 844 (11th Cir. 1983).
248. George & Co. v. Imagination Ent. Ltd., 575 F.3d 383, 398 (4th Cir. 2009).
249. AmBrit, Inc. v. Kraft, Inc., 812 F.2d 1531, 1543 (11th Cir. 1986).
2022] TRADEMARK CONFUSION REVEALED 1327

not amount to actual confusion.”250 Testimony from one customer—the

mark’s owner, and its employee—arguing customers mistakenly visited
the defendant’s store when intending to visit the owner’s store
constitutes de minimis evidence of actual confusion.251 While
“[i]solated instances of [actual] confusion are insufficient to support a
finding of likely confusion,”252 courts have found confusion by five
people,253 or even one person increases the likelihood of confusion.254
At the same time, “it is well established that no actual confusion is
required to prove a case of trademark infringement.”255 Courts have
justified this conclusion “[b]ecause of the difficulty in garnering such
evidence.”256
Confusion must be by the “actual consuming public” and therefore
anchored in a real-world context.257 The absence of actual confusion
“over a substantial period . . . creates a strong inference that there is
no likelihood of confusion.”258 In combining the two ideas, “[s]hort-
lived confusion or confusion of individuals casually acquainted with a
business is worthy of little weight, while confusion of actual customers
of a business is worthy of substantial weight.”259 Where a large volume
of contacts or transactions could give rise to confusion, and only
limited instances of confusion present themselves, courts give evidence
of actual confusion little weight.260

250. Reply All Corp., 843 F. App’x at 398 (alteration in original).

251. RVC Floor Decor, Ltd. v. Floor & Decor Outlets of Am., Inc., 527 F. Supp. 3d
305, 326 (E.D.N.Y. 2021).
252. Progressive Distrib. Servs., Inc. v. United Parcel Serv., Inc., 856 F.3d 416, 433
(6th Cir. 2017).
253. AWGI, LLC v. Atlas Trucking Co., 998 F.3d 258, 267 (6th Cir. 2021).
254. Innovation Ventures, LLC v. N2G Distrib., Inc., 763 F.3d 524, 536 (6th Cir.
2014) (“[A] single instance of actual confusion can, in some cases, ‘increase the
likelihood of confusion.’”).
255. Louis Vuitton Malletier S.A. v. Haute Diggity Dog, LLC, 507 F.3d 252, 263 (4th
Cir. 2007).
256. Monster Energy Co. v. BeastUp LLC, 395 F. Supp. 3d 1334, 1358 (E.D. Cal.
2019) (alteration in original).
257. Rearden LLC v. Rearden Com., Inc., 683 F.3d 1190, 1210 (9th Cir. 2012).
258. CareFirst of Md., Inc. v. First Care, P.C., 434 F.3d 263, 269 (4th Cir. 2006)
(finding an inference of no likelihood of confusion where there was no evidence of
confusion for nine years).
259. Safeway Stores, Inc. v. Safeway Disc. Drugs, Inc., 675 F.2d 1160, 1167 (11th Cir.
1982).
260. George & Co. v. Imagination Ent. Ltd., 575 F.3d 383, 399 (4th Cir. 2009)
(explaining that “the company’s failure to uncover more than a few instances of actual
confusion creates a presumption against likelihood of confusion in the future” when
there are so many opportunities for confusion to occur).
1328 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

Some courts hold that a lack of evidence of actual confusion does

not create a presumption of no confusion but is “simply a factor in the
court’s analysis.”261 However, as a policy lever, it gives courts the ability
to anchor their analysis in real-world characteristics. In addition, the
evidence is pre-existing, does not depend on the vagaries of survey
design, and should make it easier for courts to dispose of cases
pretrial.262 If found, it is worth its weight in gold, tipping the balance
in the plaintiff’s favor more than any other factor.263

B. Mark Similarity
Three axioms apply to the “similarity” analysis: (1) marks should be
considered in their entirety and as they appear in the marketplace; (2)
similarity is judged by appearance, sound, and meaning; (3) and,
similarities weigh more heavily than differences.264 Courts determine
whether a mark confuses the public when viewed alone to account for
the possibility that similar marks “may confuse consumers who do not
have both marks before them but who may have a general, vague, or
even hazy, impression or recollection of the other party’s mark.”265
At the most basic level, marks are confusingly similar if “ordinary
consumers would likely conclude that [the two products] share a
common source, affiliation, connection or sponsorship.”266 Identical,
even dominant, features do not “automatically mean that two marks
are similar.”267 Courts look to “the overall impression created by the
marks, not merely compare individual features,” and “may consider
the marks’ visual, aural, and definitional attributes and compare the
trade dress of the products in determining whether the total effect
conveyed by the two marks is confusingly similar.”268

261. See, e.g., Am. Dairy Queen Corp. v. W.B. Mason Co., 543 F. Supp. 3d 695, 718–
19 (D. Minn. 2021).
262. I am grateful to Jon Lee for this insight.
263. “Bridging the gap” reported a 71% win rate but given its relative infrequency
(25% versus 74% for actual) and large overlap with competitive proximity (which
could explain why it is even at 25%, for that matter 71%), the better view is to discount
it. Lim, supra note 41.
264. Entrepreneur Media, Inc. v. Smith, 279 F.3d 1135, 1144 (9th Cir. 2002).
265. Maker’s Mark Distillery, Inc. v. Diageo N. Am., Inc., 679 F.3d 410, 421 (6th Cir.
2012).
266. Fisons Horticulture, Inc. v. Vigoro Indus., Inc., 30 F.3d 466, 477 (3d Cir. 1994).
267. Sensient Techs. Corp. v. SensoryEffects Flavor Co., 613 F.3d 754, 764 (8th Cir.
2010) (quoting Gen. Mills, Inc. v. Kellogg Co., 824 F.2d 622, 627 (8th Cir. 1987)).
268. Luigino’s, Inc. v. Stouffer Corp., 170 F.3d 827, 830 (8th Cir. 1999); see also
Leelanau Wine Cellars, Ltd. v. Black & Red, Inc., 502 F.3d 504, 516–17 (6th Cir. 2007)
(explaining that courts give similarity considerable weight and consider the
2022] TRADEMARK CONFUSION REVEALED 1329

As the Eleventh Circuit noted, “[i]f a trademark operates in a

crowded field of similar marks on similar goods or services, slight
differences in names may be meaningful because consumers will not
likely be confused between any two of the crowd and may have learned
to carefully pick out one from the other.”269 Similarly, mark similarity
takes prominence when the goods are direct competitors in the
marketplace.270 The goods or services will likely be virtually identical.
However, the marks need not be as similar for there to be a likelihood
of confusion.271
Courts extol the importance of mark similarity.272 Beebe found it was
“by far the most important factor.”273 In injunction cases, 83% of
plaintiffs who won the similarity factor prevailed in the likelihood of
confusion analysis, as did 90% in plaintiff summary judgment
motions.274 The dataset reveals that only 23% of plaintiffs in injunction
cases who won the similarity factor prevailed in likelihood of confusion
analysis, a far lower number than before. However, 90% in plaintiff
summary judgment motions succeeded just as before.
Fifteen years later, the dataset also reveals that mark similarity still
comes up in 85% of the cases, the most frequently invoked factor of
them all. In 61% of all cases, the courts favored plaintiffs on the mark
similarity factor. When they did, plaintiffs won 47% of the time on the
merits. In 18% of all cases, the courts favored defendants on the mark
similarity factor. When the court favored defendants, they won 88% of
the time on the merits.

“pronunciation, appearance, and verbal translation of conflicting marks,” and courts

must “view marks in their entirety and focus on their overall impressions, not
individual features” (quoting AutoZone, Inc. v. Tandy Corp., 373 F.3d 786, 795–96
(6th Cir. 2004))).
269. Fla. Int’l Univ. Bd. Trs. v. Fla. Nat’l Univ., Inc., 830 F.3d 1242, 1260 (11th Cir.
2016) (internal quotation marks omitted).
270. Bos. Duck Tours, LP v. Super Duck Tours, LLC, 531 F.3d 1, 30 (1st Cir. 2008).
271. Bridgestone Ams. Tire Operations, LLC v. Fed. Corp., 673 F.3d 1330, 1337
(Fed. Cir. 2012) (likelihood of confusion between “Potenza” and “Turanza” marks was
greater because both referred to tires).
272. Equitable Nat’l Life Ins. Co. v. AXA Equitable Life Ins. Co., 434 F. Supp. 3d
1227, 1246 (D. Utah 2020) (“The similarity of the marks is the ‘first and most
important factor.’”); Juul Labs, Inc. v. 4X PODS., 439 F. Supp. 3d 341, 353 (D.N.J.
2020) (calling it “[t]he single most important factor in determining likelihood of
confusion”); GoTo.com, Inc. v. Walt Disney Co., 202 F.3d 1199, 1205 (9th Cir. 2000)
(noting that mark similarity “has always been considered a critical question in the
likelihood-of-confusion analysis”).
273. Beebe, supra note 6, at 1623.
274. Id. at 1625.
1330 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

Figure 10: Mark Similarity by Outcome

The dataset reveals interesting dynamics between rivalry and mark

similarity. When the parties’ marks were similar and the parties were
rivals, plaintiffs won 47% of the time. Similarly, when the parties were
non-rivals, plaintiffs also won 47% of the time. However, there is a
difference in how often the defendant wins. Defendants only win 10%
of the time when they are rivals and 24% when they are not rivals,
underscoring the expectedly powerful role rivalry plays, but in an
asymmetrical way.
One possible explanation is that similarity between the marks makes
consumers more likely to become confused about the source.
Extremely similar marks or goods may suggest counterfeiting and free
riding. Parodies, comparative advertising, and nominative use make
consumers less likely to be confused, even if the third party uses the
identical term.
Courts even dispensed entirely with the likelihood of confusion test
when parties’ marks were identical, a conclusion with implications for
early off-ramping parties, as discussed in Part III.275 Where a defendant
uses a counterfeit mark, such use is deemed inherently confusing to a

275. See Phillip Morris USA Inc. v. Shalabi, 352 F. Supp. 2d 1067, 1073 (C.D. Cal.
2004) (“[I]n cases involving counterfeit marks, it is unnecessary to perform the step-
by-step examination . . . because counterfeit marks are inherently confusing.”);
Daimler AG v. A-Z Wheels LLC, 334 F. Supp. 3d 1087, 1096 (S.D. Cal. 2018) (“It is not
necessary for the Court to analyze the likelihood of confusion test here considering
Defendants’ use the identical MERCEDES-BENZ mark.”).
2022] TRADEMARK CONFUSION REVEALED 1331

customer.276 As McCarthy explained “[c]ases where a defendant uses

an identical mark on competitive goods hardly ever find their way into
the appellate reports. Such cases are ‘open and shut’ and do not
involve protracted litigation to determine liability for trademark
infringement.”277 This is because “confusing the customer is the whole
purpose of creating counterfeit goods.”278 Such cases create a
presumption of harm such that the factor may stampede the likelihood
of confusion analysis entirely.279
Aside from the simplest forms of counterfeiting, the threshold
triggering confusion, and more so likely confusion, exists only as a
relative measure where reasonable minds may differ. Unlike real
property, there are no metes and bounds. This lack of boundaries
presents interpretive challenges that Michael Grynberg and Graeme
Austin independently attributed to the likelihood of confusion’s
current uncertainty.280 The problem is common to other areas of the
law as well. For instance, copyright law’s substantial similarity standard
suffers many of the same ills as the likelihood of confusion and
demands reconsideration.281

276. EAT BBQ LLC v. Walters, 47 F. Supp. 3d 521, 530 (E.D. Ky. 2014) (“[T]here is
almost never a dispute regarding confusion.”).
277. Wynn Oil Co. v. Thomas, 839 F.2d 1183, 1191 (6th Cir. 1988) (quoting 4
MCCARTHY, supra note 2, § 23:20).
278. Gucci Am., Inc. v. Duty Free Apparel, Ltd., 286 F. Supp. 2d 284, 287 (S.D.N.Y.
2003).
279. See Threeline Imps., Inc. v. Vernikov, 239 F. Supp. 3d 542, 561 (E.D.N.Y. 2017);
see also Dish Network L.L.C. v. Siddiqi, No. 18 CV 4397, 2019 WL 5781945, at *4
(S.D.N.Y. 2019) (“‘[I]t is not necessary to perform the step-by-step examination of each
Polaroid factor’ when a counterfeit mark is at issue.”); Halo Optical Prods., Inc. v.
Liberty Sport, Inc., No. 14-cv-00282, 2017 WL 1082443, at *11 (N.D.N.Y. Mar. 22, 2017)
(“[W]hen dealing with an identical mark . . . courts are not necessarily required to
analyze the Polaroid factors.”); Gucci Am., 286 F. Supp. 2d at 287 (“[T]he Court need
not undertake a factor-by-factor analysis under Polaroid because counterfeits, by their
very nature, cause confusion.”).
280. Grynberg, Judicial Role in Trademark Law, supra note 234, at 1303 (“Trademark’s
fundamental inquiry, whether a likelihood of confusion exists, invites judicial
lawmaking in no small part because the term ‘likelihood of confusion’ presents an
interpretive problem.”); Graeme W. Austin, Tolerating Confusion About Confusion:
Trademark Policies and Fair Use, 50 ARIZ. L. REV. 157, 160 (2008) (“There is considerable
uncertainty about some of the key questions that are germane to the factual inquiry at
the heart of the likelihood of confusion analysis.”).
281. See, e.g., Daryl Lim, Saving Substantial Similarity, 73 FLA. L. REV. 591, 640–41
(2021) [hereinafter Lim, Saving Substantial Similarity] (explaining how the substantial
similarity standard generates “capricious and wrong results”).
1332 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

Courts use sights, sounds, and meaning to make snap judgments

about mark similarity.282 These heuristics allow judges to rely on “a
small set of cheap and reliable factors that are close enough to the
ideal.”283 Adam Samaha approves of it since “[p]rioritizing the judge’s
impressions about the similarity of marks, therefore, tends toward the
high values of trademark law at bargain basement prices.”284
Defendants can easily compare visual or aural elements in context,
making this a useful factor to encourage due diligence.285
The key takeaway is that the commercial context matters.286 Marks
should not be compared side-by-side as they might be shown in the
courtroom.287 Instead, courts determine whether the public would
confuse the marks when viewed alone because some highly similar
marks can confuse consumers that view them without appropriate
commercial context.288 For this reason, courts cannot dissect marks
since consumers encounter them in their entirety in those settings.
Instead, courts focus on their overall impressions rather than on their
features.289 That which qualifies as mark similarity also disqualifies
intent, survey evidence, mark strength, and consumer sophistication.

C. Competitive Proximity
Competitive proximity tells courts how likely consumers are to
assume an association between the marks used on related products.290
For example, “[t]he similarities between the parties’ distribution
channels and marketing strategies suggest an overlapping general class

282. Adam M. Samaha, Looking Over a Crowd—Do More Interpretive Sources Mean More
Discretion?, 92 N.Y.U. L. REV. 554, 614 (2017) (“[A]ccurately estimating the probability
of consumer confusion can require a snap judgment, which often is how consumers
actually formulate impressions and make purchasing decisions.”).
283. Id.
284. Id.
285. See 4 MCCARTHY, supra note 2, § 23:21 (discussing the “sound, sight, and
meaning” test for mark similarity).
286. Flower Mfg., LLC v. CareCo, LLC, 466 F. Supp. 3d 797, 814 (N.D. Ohio 2020)
(“I am to judge the marks’ similarity as they appear in their commercial context.”); see
also Homeowners Grp., Inc. v. Home Mktg. Specialists, Inc., 931 F.2d 1100, 1109 (6th
Cir. 1991) (“[A] court must determine, in the light of what occurs in the marketplace,
whether the mark will be confusing to the public when singly presented.” (internal
quotation marks omitted) (quoting Wynn Oil Co. v. Thomas, 839 F.2d 1183, 1187 (6th
Cir. 1988))).
287. Homeowners Grp., Inc., 931 F.2d at 1106.
288. Daddy’s Junky Music Stores, Inc. v. Big Daddy’s Fam. Music Ctr., 109 F.3d 275,
283 (6th Cir. 1997) (quoting Wynn Oil, 839 F.2d at 1188).
289. AutoZone, Inc. v. Tandy Corp., 373 F.3d 786, 795 (6th Cir. 2004).
290. Virgin Enters. Ltd. v. Nawab, 335 F.3d 141, 150 (2d Cir. 2003).
2022] TRADEMARK CONFUSION REVEALED 1333

of consumers of the parties’ products.”291 However, two products or

services within the same general field do not automatically trigger a
likelihood of confusion.292 Similarly, a high percentage of overlap in
“an extremely small subset of products does not demonstrate a high
degree of relatedness.”293
Services and goods within the same broad industry are not
necessarily related. Rather, related services are marketed and
consumed such that buyers are likely to believe that the services come
from a common company.294 Courts examine “how and to whom the
respective goods or services of the parties are sold.”295 Less likelihood
of confusion exists where the goods are sold through different
avenues, “parties have different customers[,] and [they] market their
goods or services in different ways.”296 “[I]f the parties compete
directly, confusion is likely” between sufficiently similar marks.297 “[I]f
the goods and services are somewhat related, but not competitive, then
the likelihood of confusion will turn on other factors[.]298 [F]inally, if
the products are unrelated, confusion is highly unlikely.”299
The dataset reveals that competitive proximity comes up in 74% of
the cases. In 51% of all cases, the courts favored plaintiffs on the
competitive proximity factor. When they did, plaintiffs won 45% of the
time. In 14% of all cases, the courts favored defendants on the
competitive proximity factor. When the court favored defendants, they
won 85% of the time.

291. Monster Energy Co. v. BeastUp LLC, 395 F. Supp. 3d 1334, 1359 (E.D. Cal.
2019). Other circuits use similar formulations. See e.g., Therma-Scan, Inc. v.
Thermoscan, Inc., 295 F.3d 623, 632 (6th Cir. 2002) (noting that direct rivalry through
similar goods or services is likely confusing).
292. Matrix Motor Co. v. Toyota Jidosha Kabushiki Kaisha, 290 F. Supp. 2d 1083,
1092 (C.D. Cal. 2003).
293. AutoZone, Inc., 373 F.3d at 798 (“[I]f [the defendant] stocked only five types of
batteries all of which were also sold by [the plaintiff], the overlap would be 100%, even
though in reality [the defendant] and [the plaintiff] would share only five products of
the approximately 55,000 offered by [the plaintiff].”).
294. Daddy’s Junky Music Stores, Inc. v. Big Daddy’s Fam. Music Ctr., 109 F.3d 275,
282–83 (6th Cir. 1997) (quoting Homeowners Grp., Inc. v. Home Mktg. Specialists,
Inc., 931 F.2d 1100, 1109 (6th Cir. 1991)).
295. Leelanau Wine Cellars, Ltd. v. Black & Red, Inc., 502 F.3d 504, 519 (6th Cir.
2007).
296. Therma-Scan, 295 F.3d at 636.
297. Kellogg Co. v. Toucan Golf, Inc., 337 F.3d 616, 624 (6th Cir. 2003).
298. Id.
299. Id.
1334 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

Figure 11: Competitive Proximity by Outcome

Competitive proximity encompasses adjacent Polaroid factors. One

example is the likelihood plaintiffs or defendants will expand into each
other’s market or the “bridging the gap” factor.300 The likelihood that
consumers will confuse the sources of parties’ products increases when
there is a “strong possibility that either party will expand its business to
compete with the other’s.”301 This confusion may happen when goods
and services are complementary, sold to the same class of purchasers,
or similar in use and function.302
Courts examine the two concepts in tandem with each other.303
Consider Kohler Co. v. Bold International FZCO,304 where the court noted
that “‘[b]ridging the gap’ refers to the likelihood that the senior

300. Disney Enters., Inc. v. Sarelli, 322 F. Supp. 3d 413, 434 (S.D.N.Y. 2018) (“The
third and fourth Polaroid factors, respectively, address the proximity of the goods or
services at issue and the possibility that the senior user will ‘bridge the gap,’ or expand
the scope of its business and enter the market of the junior user. Thus, these two
distinct but related factors ‘focus on the degree to which the [parties’] products
currently compete with each other or are likely to compete with each other in the
future.’” (alteration in original) (quoting Medici Classics Prods., LLC v. Medici Grp.,
LLC, 683 F. Supp. 2d 304, 311–12 (S.D.N.Y. 2010))).
301. Kibler v. Hall, 843 F.3d 1068, 1082 (6th Cir. 2016).
302. AMF Inc. v. Sleekcraft Boats, 599 F.2d 341, 350 (9th Cir. 1979).
303. RVC Floor Decor, Ltd. v. Floor & Decor Outlets Am., Inc., 527 F. Supp. 3d 305,
325 (E.D.N.Y. 2021) (“For the same reasons explained [in] the ‘competitive proximity’
analysis, the parties serve the same market and any gap has already been bridged.”).
304. 422 F. Supp. 3d 681 (E.D.N.Y. 2018).
2022] TRADEMARK CONFUSION REVEALED 1335

user . . . will enter into the same market as that of the junior user . . .
where the goods are not yet in close competitive proximity.”305 When
the parties’ goods are the same, courts simply fold this factor into
competitive proximity as there is no gap to bridge.306 In this case, a
consumer seeing the goods or services would likely be confused about
their source.307
Another example is the degree of care the consumer might exercise
in purchasing the parties’ goods, as mentioned in Section I.C.4. Courts
look both to the “relative sophistication of the relevant consumer”308
and the cost of the item309 in determining the degree of care likely to
be exercised by the purchaser. The “reasonably prudent consumer” is
expected “to be more discerning—and less easily confused—when
[they are] purchasing expensive items.”310 Conversely, customers may
be less careful when purchasing inexpensive products, thus making
confusion more likely.311
“Bridging the gap” rarely arose, only in 25% of the cases. In 7% of
all cases, the courts favored plaintiffs on the “bridging the gap” factor.
When they did, plaintiffs won 71% of the time. In 6% of all cases, the
courts favored defendants on the “bridging the gap” factor. When the
court favored defendants, they won 83% of the time.

305. Id. at 725.

306. Star Indus., Inc. v. Bacardi & Co., 412 F.3d 373, 387 (2d Cir. 2005)
(“Because . . . [the parties’] products are already in competitive proximity, there is
really no gap to bridge, and this factor is irrelevant to the Polaroid analysis in this
case.”).
307. DJ Direct, Inc. v. Margaliot, 512 F. Supp. 3d 396, 411 (E.D.N.Y. 2021).
308. Fortune Dynamic, Inc. v. Victoria’s Secret Stores Brand Mgmt., Inc., 618 F.3d
1025, 1038 (9th Cir. 2010).
309. Brookfield Commc’ns, Inc. v. W. Coast Ent. Corp., 174 F.3d 1036, 1060 (9th
Cir. 1999).
310. Id.
311. Id.
1336 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

Figure 12: “Bridging the Gap” by Outcome

Similarly, the quality of the defendant’s goods is closely related to

competitive proximity. The quality factor featured even more
infrequently, and 11% less than “bridging the gap.” In 1% of all cases,
the courts favored plaintiffs on the quality factor. When they did,
plaintiffs won 50% of the time. In 1% of all cases, the courts favored
defendants on the “bridging the gap” factor. When the court favored
defendants, they won 100% of the time. Of course, these figures should
be seen in the context of the very small sample size.
2022] TRADEMARK CONFUSION REVEALED 1337

Figure 13: Quality of D’s Product by Outcome

Courts look to the product, the relevant market, and potential

consumers.312 Product proximity overlaps substantially with marketing
and advertising channels and should be subsumed within those
channels. For this reason, product proximity can serve as an omnibus
factor for other factors such as the relative quality of goods sold,
“bridging the gap” from the perspective of the relevant public (rather
than from the legitimate aspirations of the trademark owner), and
similarity of distribution channels.
One court in the dataset was exemplary in defining the relevant
consumer market.313 To determine whether that market included
potential commercial and government customers, it examined the
trademark owner’s revenue sources, proposals it sent to two potential
customers, and the defendant’s exploratory acquisition of the
trademark owner to conclude that both parties targeted similar
customers.314 Despite this fact-intensive inquiry, the court notably

312. Best Cellars, Inc. v. Grape Finds at Dupont, Inc., 90 F. Supp. 2d 431, 456
(S.D.N.Y. 2000).
313. Ironhawk Techs., Inc. v. Dropbox, Inc., 994 F.3d 1107, 1117 (9th Cir. 2021)
(“Before addressing the Sleekcraft factors, we must define the relevant consumer
market because ‘a court conducting a trademark analysis should focus its attention on
the relevant consuming public.’” (quoting Rearden LLC v. Rearden Com., Inc., 683
F.3d 1190, 1214 (9th Cir. 2012))).
314. Id. at 1117–18.
1338 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

concluded that “a reasonable jury could find that [plaintiff’s] potential

consumers include commercial customers.”315
Market definition was an interesting issue that arose in a few cases in
the dataset. Though similar, courts distinguish between market
definition under trademark and antitrust law. For example, one case
in the dataset reported a plaintiff asserting that “courts have looked to
antitrust law . . . to find goods competitive where they are ‘either
identical or available substitutes for each other.’”316 Disagreeing with
this assessment, the court responded that “the question . . . is not
whether [the defendant’s] conduct impair[ed] competition in the
marketplace[,] but whether it . . . infringed” upon a protected interest
in the plaintiff’s trademark.317

D. Summing It Up
In sum, the eight Polaroid Factors can be efficiently subsumed into a
troika of actual confusion, mark similarity, and competitive proximity.
The table below shows the troika being the most prominent factors.
They also deliver consistent win rates to plaintiffs if the particular
mark favors them, at between 45% to 54%, mapping almost exactly
to Priest-Klein’s 50% figure discussed in Section I.A.

315. Id. at 1118.

316. Flower Mfg., LLC v. CareCo, LLC, 466 F. Supp. 3d 797, 812 (N.D. Ohio 2020)
(emphasis omitted).
317. Id. (“It is fair to say that trademark laws were enacted for the protection of the
competitor who owns a mark and not for protection of competition in the marketplace
in general.”).
2022] TRADEMARK CONFUSION REVEALED 1339

Figure 14: Revised Likelihood of Confusion Factors

Rules of Frequency Plaintiff’s Win

Polaroid Factors
Thumb (%) Rate (%)
Strength of the
Discarded 70 46
plaintiff’s mark
Similarity of plaintiff’s
Retained 85 47
and defendant’s marks
Competitive proximity
Retained 73 45
of products or services
Likelihood that Covered by
plaintiff will “bridge competitive
the gap” and offer a proximity of 25 71
product like a products or
defendant’s services
Actual confusion Retained 74 54
Defendant’s good faith Discarded 66 52
Covered by
competitive
Quality of defendant’s
proximity of 11 50
product
products or
services
Buyer sophistication Discarded 46 49

More importantly, the troika moves trademark doctrine a step in the

right direction by limiting ad hoc fact-finding. However, the troika
alone is incomplete. Mark McKenna and Mark Lemley warn that unless
we can “identify more specifically the types of relationships that could
give rise to actionable confusion, there is no logical stopping point for
trademark protection.”318 The converse is also true—we also need to
identify safe harbors. It is difficult even for savvy parties to predict the
outcome in advance and resolve disputes early in proceedings, placing
swathes of activity at significant risk.319

318. See Mark A. Lemley & Mark McKenna, Irrelevant Confusion, 62 STAN. L. REV. 413,
439 (2010).
319. David S. Welkowitz, The Virtues and Vices of Clarity in Trademark Law, 81 TENN.
L. REV. 145, 148 (2013) (“Because the level and even the existence of confusion is
difficult to predict in advance, partly due to the uncertainties built into trademark
law’s test for confusion, those who would engage in valued activity must do so at
significant risk.”).
1340 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

Simplifying confusion benefits other aspects of trademark law. For

example, trademark law’s first sale doctrine also permits some
marketplace confusion by letting others sell used or reconditioned
goods bearing the mark.320 Nominative fair use may likewise fold the
likelihood of confusion standard into its analysis.321 What is “fair”
implicates the confusion arising from using the offending mark—whether
the defendant only used as much as necessary of the plaintiff’s
mark—which in turn impacts the vagueness of the likelihood of
confusion standard.322 The same issue arises with expressive trademark
uses323 and the legality of keyword advertising.324 Fair use is the focus
of the next Section.

E. A Word on Fair Use

As trademarks expand beyond source identification, they seed
public discourse with their communicative value.325 Trademark owners
obtain rights with inchoate boundaries. When the public interacts with
a trademark, the mark may imbue with collective meaning. This
collective meaning has social value, and in appropriate instances, the
law should offer them categorical protection from lawsuits.326

320. See Nitro Leisure Prods., L.L.C. v. Acushnet Co., 341 F.3d 1356, 1362–64 (Fed.
Cir. 2003) (consumer confusion as benchmark for applying the first sale doctrine).
321. Toyota Motor Sales, U.S.A., Inc. v. Tabari, 610 F.3d 1171, 1175–76 (9th Cir.
2010) (asking whether (1) the product was readily identifiable without use of the mark;
(2) defendant used more of the mark than necessary; or (3) defendant falsely
suggested he was sponsored or endorsed by the trademark holder).
322. E.g., KP Permanent Make-Up, Inc. v. Lasting Impression I, Inc., 543 U.S. 111,
123 (2004) (confusion relevant to whether descriptive use is “fair”).
323. See Rogers v. Grimaldi, 875 F.2d 994, 999 (2d Cir. 1989) (adopting balancing
test that asks whether the use of a trademark as the title of an expressive work is
artistically relevant to the underlying work and, if so, whether “the title explicitly
misleads as to the source or the content of the work”).
324. See Network Automation, Inc. v. Advanced Sys. Concepts, Inc., 638 F.3d 1137,
1154 (9th Cir. 2011) (noting that keyword advertisements could be “confusingly
labeled or not labeled at all” making how advertisements appear on the results page
must be considered).
325. See Alex Kozinski, Trademarks Unplugged, 68 N.Y.U. L. REV. 960, 973–74 (1993)
(noting how businesses inject the “effervescent qualities” of trademarks “into the
stream of communication with the pressure of a firehose by means of mass media
campaigns”).
326. See, e.g., William McGeveran & Mark P. McKenna, Confusion Isn’t Everything, 89
NOTRE DAME L. REV. 253, 301–06 (2013) (proposing categorical exclusions for some
favored uses).
2022] TRADEMARK CONFUSION REVEALED 1341

Communication relies on a plethora of legally protected words,

graphics, sounds, and smells.327 Beyond computers or smartphones,
APPLE may represent a nonconformist hip lifestyle compared with
users of LENOVO’s more staid business offerings. Trademarks become
tools of communication and expression, and the public helps shape
their boundaries as they become symbols that embody culture itself.328
Trademark owners may be anxious to protect themselves from uses
that dilute the value of a household logo or name even when
consumers are not confused. Between 2019 and 2021, Apple filed 215
trademark oppositions, targeting small companies and nonprofits that
have nothing to do with providing technology products or services,
including an Indian food blog and a public school. 329 In these
mass-produced, boilerplate-worded oppositions, Apple has argued
that “Apple marks are so famous and instantly recognizable” that other
trademarks will weaken the strength of its brand or cause the “ordinary
consumer to believe that applicant is related to, affiliated with or
endorsed by Apple.”330 While Apple protested that this is simply what
the law “requires,” Professor Christine Farley has called them “bullying
tactics.”331 Whatever the case, the impact is clear; Apple has been
successful in preventing registration of a wide variety of marks. When
faced with an opposition by Apple, applicants expressly withdrew their
applications 17% of the time and failed to respond and subsequently
defaulted 59% of the time.332 The Tech Transparency Report noted
only one win against Apple—by the U.S. government:

327. Diamond & Franklyn, supra note 165, at 2031.

328. Barton Beebe, The Semiotic Analysis of Trademark Law, 51 UCLA L. REV. 621, 624
(2004) (arguing trademark law is both an economic doctrine and “a semiotic doctrine
elaborating the principles of sign systems, of language”).
329. See Apple’s Trademark ‘Bullying’ Targets Small Businesses, Nonprofits TECH
TRANSPARENCY PROJECT (Mar. 11, 2022) [hereinafter TECH TRANSPARENCY PROJECT],
https://www.techtransparencyproject.org/articles/apples-trademark-bullying-targets-
small-businesses-nonprofits [https://perma.cc/96QR-UZBT] (“Many choose to
simply give up rather than take on a mega-corporation with a market value of $2.5
trillion, but in doing so, they lose whatever funds they invested in designing their logo
and hiring a lawyer to deal with the trademark application.”).
330. Ryan Mac & Kellen Browning, Apps and Oranges: Behind Apple’s ‘Bullying’ on
Trademarks, N.Y. TIMES (Mar. 11, 2022) https://www.nytimes.com/2022/03/11/
technology/apple-trademarks.html [https://perma.cc/LEV8-RPEE].
331. Id. For an interesting discussion on how to prevent “bullying” in trademark
law, see Stacey Dogan, Bullying and Opportunism in Trademark and Right-of-Publicity Law,
96 B.U. L. REV. 1293 (2016) (examining various instances of trademark bullying,
including by Monster Energy, one of the most infamous trademark bullies).
332. See TECH TRANSPARENCY PROJECT, supra note 329 (“Of the 118 North American
cases analyzed by TTP, 76 have been decided in Apple’s favor, with a complete defeat
1342 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

Apple appears to have been stopped in only one instance: when it

challenged the U.S. government. The Department of Energy
registered an online research service called Pages in 2020. Apple
owns the trademark Pages for its word processor and opposed the
agency’s trademark on those grounds. The Department’s response
was brief, and its main defense took the form of a one-sentence
declaration: “There is no likelihood of confusion, mistake, or
deception between [Apple’s] marks and DOE’s PAGES mark.” After
more than a year of negotiations, Apple agreed to withdraw its
objections, without DOE making any change to its application. In
TTP’s dataset, no other applicant was able to completely beat back
Apple.333
While the costs of invading free speech and other interests are high,
the costs of being overly permissive in expressive use cases cause only
minimal harm.334 Research on brand extensions shows owners are
rarely “harmed by consumers’ mistaken association of unrelated
products.”335 Consumers rarely alter how they view the brand quality
when they encounter criticism about other products offered under
that mark.336 The negative impact stays with the related products but
does not corrupt a positive view of the owner’s line of products.337
For this reason, Mark McKenna has warned against reflexively
prohibiting every form of confusion.338 Instead, trademark law should
only be concerned with confusion that influences consumer
decision-making. Bone also cautions that economic concerns over
confusion should be distinguished from penalizing intentional
deception without evidence of consumer confusion.339 Similarly, Lisa
Ramsey flags the need to safeguard free speech interests in the face of

of the other parties’ proposed logos or trademarks. None of these cases saw a full trial
before TTAB but were withdrawn or abandoned by the applicants amid Apple’s
pressure.”).
333. Id.
334. William McGeveran, The Trademark Fair Use Reform Act, 90 B.U. L. REV. 2267,
2286 (2010).
335. Lemley & McKenna, supra note 318, at 429.
336. See id. at 429.
337. See id. at 430 (“Consumers, in other words, are smart enough to distinguish
different products and hold different impressions of them.”).
338. Mark P. McKenna, A Consumer Decision-Making Theory of Trademark Law, 98 VA.
L. REV. 67, 73 (2012).
339. Robert G. Bone, Taking the Confusion out of “Likelihood of Confusion”: Toward a
More Sensible Approach to Trademark Infringement, 106 NW. L. REV. 1307, 1377 (2012).
2022] TRADEMARK CONFUSION REVEALED 1343

encroaching trademark enforcement.340 One way to do this is by

fortifying fair use.
Fair use is currently regarded as an affirmative defense.341 That
generally precludes the pretrial disposition of the case.342 In the
interest of early off-ramping cases, it is perhaps fortunate then that fair
use rarely arose in the dataset (6%), with equal probability that a court
would eventually find in favor of either plaintiff (42%) or defendant
(50%). It also provides empirical evidence that converting fair use from
an affirmative defense to a safe harbor would create a powerful tool to
fend off trademark trolls without appreciably disrupting day-to-day
trademark practice.
Safe harbors offer advantages over attempts to prescribe clear rules.
These include improving predictability and ease of determination,
allowing courts to resolve issues sooner in the litigation process. Here,
Gideon Parchomovsky and Alex Stein make a more general point that
“[r]eplacing these criteria with rules that will lay down irrebuttable
presumptions of consumer confusion, or lack thereof, could make
litigation over trademarks cheaper than it presently is.”343 The case is
over as soon as the defendants demonstrate a basic fact.344
Safe harbors exist within trademark law, specifically the likelihood
of confusion tests. For instance, the law does not protect functional
product designs to avoid giving plaintiffs an advantage against rivals
unrelated to the plaintiff’s reputation.345 Similarly, the law keeps
plaintiffs on a leash to not monopolize trademarks with descriptive
words and receive protection for generic terms.346 Expressive uses for

340. See Lisa P. Ramsey, Increasing First Amendment Scrutiny of Trademark Law, 61 SMU
L. REV. 381, 384–85 (2008) (advocating rigorous First Amendment rights of trademark
laws to protect free speech).
341. See, e.g., KP Permanent Make-Up, Inc. v. Lasting Impression I, Inc., 543 U.S.
111, 123 (2004) (finding that confusion is relevant to whether descriptive use is “fair”);
New Kids on the Block v. News Am. Publ’g, Inc., 971 F.2d 302, 308 (9th Cir. 1992)
(noting that confusion is relevant to nominative fair use).
342. Kelly-Brown v. Winfrey, 717 F.3d 295, 308 (2d Cir. 2013) (“fair use . . . requires
consideration of facts outside of the complaint and thus is inappropriate to resolve on
a motion to dismiss.”); see also 2 MCCARTHY, supra note 2, § 11:49 (“Because classic fair
use is an affirmative defense, it is normally not appropriate for consideration on a . . .
motion to dismiss for failure to state a claim.”).
343. Gideon Parchomovsky & Alex Stein, Catalogs, 115 COLUM. L. REV. 165, 178
(2015).
344. See Welkowitz, supra note 319, at 168 (referencing Fed. R. Evid. 301).
345. 15 U.S.C. § 1052(e)(5).
346. See Abercrombie & Fitch Co. v. Hunting World, Inc., 537 F.2d 4, 9–10 (2d Cir.
1976) (explaining the limitations on generic and descriptive marks).
1344 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

commentary, parody, or education should fall within safe harbors.347

Critiquing products or corporate behavior requires us to use them.348
Therefore, the first safe harbor should be expressive uses of protected
trademarks.
The second safe harbor is referential uses of trademarks. Nominative
fair use (referring to the trademark holder or its products) should not
trigger liability.349 For example, rivals and repair services need to make
referential uses to compete and advertise their services to the public.350
The law currently recognizes comparative use as a defense, but it
should go further and offer a safe harbor to these uses.351
Recognizing that the Polaroid factors are a “bad fit” in nominative
fair use cases, one court instructed that future courts should consider
fair use alongside the Polaroid factors when considering a claim of
nominative fair use.352 Likelihood of confusion is relevant to
determining whether the use is objectively fair and whether defendants
use the term “other[] than as a mark.”353 Likewise, nominative fair use
folds confusion into determining whether an expressive use “explicitly
misleads” consumers or whether the use falsely suggests a source or
sponsorship.354

347. See Andy Greene, Nathan Fielder Talks ‘Dumb Starbucks’ and Pranking Instagram,
ROLLING STONE (July 24, 2014), http://www.rollingstone.com/movies/news/nathan-
fielder-talks-dumb-starbucks-and-pranking-instagram-20140724 [https://perma.cc/
Q3BL-R96R] (noting that parody laws allowed comedian Nath Fielder to “open up a
near perfect replicate of a Starbucks just as long as he put the word ‘dumb’ before
everything in the store, down to CDs labeled ‘Dumb Nora Jones Duets’”).
348. See New Kids on the Block v. News Am. Publ’g, Inc., 971 F.2d 302, 307 (9th Cir.
1992) (“Much useful social and commercial discourse would be all but impossible if
speakers were under threat of an infringement lawsuit every time they made reference
to a person, company or product by using its trademark.”).
349. See e.g., Zatarains, Inc. v. Oak Grove Smokehouse, Inc., 698 F.2d 786, 788 (5th
Cir. 1983) (rivals allowed to use “fish fry” to describe their own batter mixes even when
doing so creates some likelihood of confusion with owners’ FISH-FRI trademark).
350. See, e.g., Toyota Motor Sales, U.S.A., Inc. v. Tabari, 610 F.3d 1171, 1180–82 (9th
Cir. 2010) (allowing automobile broker specializing in facilitating Lexus purchases to
use LEXUS mark as part of domain name).
351. See, e.g., Smith v. Chanel, Inc., 402 F.2d 562, 563 (9th Cir. 1968) (holding that
truthful comparative advertising is not trademark infringement).
352. Int’l Info. Sys. Sec. Certification Consortium, Inc. v. Sec. Univ., LLC, 823 F.3d
153, 168 (2d Cir. 2016); see also Coty Inc. v. Excell Brands, LLC, 277 F. Supp. 3d 425,
457 (S.D.N.Y. 2017) (“Acourt considering a claim of nominative fair [use] should
consider three factors in addition to the standard Polaroid factors.”).
353. See, e.g., KP Permanent Make-Up, Inc. v. Lasting Impression I, Inc., 543 U.S.
111, 118 (2004).
354. William McGeveran, Rethinking Trademark Fair Use, 94 IOWA L. REV. 49, 95–97,
100–04 (2008).
2022] TRADEMARK CONFUSION REVEALED 1345

Safe harbors like those for expressive and descriptive uses allow
courts to dispose of the likelihood of confusion cases more simply and
justly. For example, uses that mirror the conventional way descriptive
terms are used in ordinary language give prospective users an
advantage in establishing the protected use and exiting litigation early,
thereby avoiding high litigation costs. In addition, they help carve out
pockets of strong protection and guide the development of trademark
rights in other areas such as merchandising rights, without giving
owners the right to rely upon the likelihood of confusion to justify its
approval. Within this framework, it is also worth considering a safe
harbor beyond descriptive or expressive fair uses that provide small
businesses and nonprofits like those described above with an effective
and low-cost way to deflect policing by overzealous trademark owners.

III. OBSERVATIONS AND IMPLICATIONS

The final Part addresses three issues. First, it observes that courts
combine factors and analyze them together. Occasionally they do so
overtly. Factor folding occurs across all the factors, including the troika
of mark similarity, actual confusion, and competitive proximity. A
strong showing on one factor may prevent the need to show another.
Second, decision-makers tend to start limiting the factors that they
choose to consider when confronted with complex decision processes.
At some point, decision-makers will stop analyzing new information
and instead commit to a decision and then work backward to vindicate
it. This adaptation has allowed 63% of litigants to receive an early
resolution on the merits. Third, it explains how the empirical analysis
provides a blueprint for algorithmic adjudication using AI, taking the
reader from conception to execution to identifying and addressing its
limitations.

A. Factor Folding
While likelihood of confusion factors may present themselves as
discrete categories, the dataset reveals that courts do not regard them
as such. Courts instead combine factors and analyze them together.
This is called “folding” and in likelihood of confusion analyses the
courts notoriously fold the factors together, using the presence of one
factor as a proxy analysis for another.
1346 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

For instance, in J-B Weld Co., LLC v. Gorilla Glue Co.,355 the Eleventh
Circuit used similarity as a proxy for intent.356 Instead of making an
adverse finding outright, a court may sometimes shift the plaintiff’s
burden of proof to require the defendant instead to disprove bad
faith.357 The direction of a court’s substitution bias is not a one-way
street. On occasion, courts leaned on lack of evidence of actual
confusion to vindicate imitation of successful product features.358 In
this way, the likelihood of confusion factors operate not as
independent elements along orthogonal lines but as a sliding scale: the
more closely the products compete, the more likely it is that a new
product whose design arrogates the atypical qualities of the old
product will confuse consumers.
Sometimes the combination is obvious. For instance, courts treat
actual confusion as an indicator of mark strength.359 One court
explained that “[i]f buyers are confused between two sources, then this
also means that they must have recognized plaintiff’s designation as a
trademark and associated it only with the plaintiff.”360 Another court
observed that “where the parties’ marks are identical and their goods
are in very close competitive proximity, a highly sophisticated
consumer may be the most vulnerable to confusion.”361
This blending was not confined to the likelihood of confusion
factors but extended to fair use. One court explained actual confusion
gets to “the heart of the nominative fair use situation.”362 At other
times, the logical connection is more tenuous, suggesting a negative

355. 978 F.3d 778, 790 (11th Cir. 2020).

356. Id. (“[W]here a defendant attempts to copy a plaintiff’s product ‘as closely as
possible’ and uses the plaintiff’s product design as a model, it may be ‘inferred that
[defendant] purposely chose a mark which was very similar to [plaintiff’s] in order to
benefit from the reputation [plaintiff]’s mark had already achieved.”).
357. Kiki Undies Corp. v. Promenade Hosiery Mills, Inc., 411 F.2d 1097, 1101 (2d
Cir. 1969) (“[W]here the allegedly infringing mark is identical to the registered mark,
and its use began subsequent to the plaintiff’s trade-mark registration, the defendant
must carry the burden of explanation and persuasion.”).
358. George Basch Co. v. Blue Coral, Inc., 968 F.2d 1532, 1541 (2d Cir. 1992)
(“Absent confusion, imitation of certain successful features in another’s product is not
unlawful . . . .”).
359. Vital Pharms., Inc. v. Monster Energy Co., No. 19-60809-CIV, 2021 WL
3371942, at *43 (S.D. Fla. Aug. 3, 2021) (stating that several courts have held that
“actual confusion is an indicium of secondary meaning”); Am. Sci. Chem., Inc. v. Am.
Hosp. Supply, 690 F.2d 791, 793 (9th Cir. 1982).
360. 2 MCCARTHY ON TRADEMARKS § 15:11; see also Popular Bank of Fla. v. Banco
Popular de P.R., 9 F. Supp. 2d 1347, 1358 (S.D. Fla. 1998) (making the same point).
361. Kohler Co. v. Bold Int’l FZCO, 422 F. Supp. 3d 681, 730 n.20 (E.D.N.Y. 2018).
362. Juul Labs, Inc. v. 4X PODS, 509 F. Supp. 3d 52, 64 (D.N.J. 2020).
2022] TRADEMARK CONFUSION REVEALED 1347

form of coherence-based reasoning. For instance, in ascertaining mark

similarity, one court considered the nature of the purchasing process,
reasoning that where marks are similar but used in different contexts
or on different visual displays, the risk of confusion is minimized.363
This practice of factor folding happens across all factors, including
with the trio of key factors of mark similarity, actual confusion, and
competitive proximity, where a strong showing on one factor may be
sufficient. A strong showing on one factor may prevent the need to
show another.364 Factors thus trump each other, with competitive
proximity often trumping mark similarity without explaining why one
factor should take precedence over another.365
Notably, only 9% of cases in the dataset expressly acknowledge
“folding” factors. Most do not, regardless of the procedural posture in
the case. This makes it more difficult for appellate courts and
commentators to hold lower courts accountable for their analysis when
this “folding” occurs. This phenomenon underscores the importance
of minimizing coherence-based reasoning by having courts focus on a
few factors when making likelihood of confusion determinations.

B. Early Off-Ramps
Courts generally agree that “application of the factors is a highly fact-
intensive inquiry both as to the assessment of the evidence concerning
each factor and as to the overall synthesis of factors and the
evidence.”366 A context-specific inquiry guides courts towards the
material aspects of product source or affiliation germane to the
consuming public’s understanding.367 Given their marching orders,
one might expect judges to weigh the likelihood of confusion factors

363. Sensient Techs. Corp. v. SensoryEffects Flavor Co., 613 F.3d 754, 765 (8th Cir.
2010); E.A. Sween Co. v. A & M Deli Express, Inc., No. 17 CV 2514, 2018 WL 1283682,
at *4 (E.D.N.Y. Mar. 9, 2018).
364. See e.g., ConAgra, Inc. v. George A. Hormel & Co., 990 F.2d 368, 371 (8th Cir.
1993) (“[W]hen ‘products are closely related, less similarity in trademarks is necessary
to support a finding of infringement.’” (quoting SquirtCo v. Seven-Up Co., 628 F.2d
1086, 1091 (8th Cir. 1980))); Select Comfort Corp. v. Baxter, 996 F.3d 925, 934 (8th
Cir. 2021) (“[T]he relative importance of any given factor is influenced greatly by how
the other factors might apply.”).
365. Reply All Corp. v. Gimlet Media, LLC, 843 F. App’x 392, 396 (2d Cir. 2021)
(“[W]hile the two marks undoubtedly share aural and typographic similarities, they
are unlikely to appear in the marketplace in a similar manner.”).
366. Select Comfort Corp., 996 F.3d at 933–34.
367. Id. at 934 (“Common sense is inherent in the factors, and the factors, properly
applied, should try to capture a holistic view of the normal experiences for any given
industry, product, or service.”).
1348 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

carefully.368 However, as seen in this study, that is not what happens in

practice, nor indeed more generally.369
When confronting complex decision processes, decision-makers
tend to limit the factors they consider.370 After a certain point, judges
will stop analyzing new information, instead committing to their
decision first and then working backwards to rationalize it. Some
courts opt for a holistic weighing of the factors rather than attempting
piecemeal arithmetic.371 Others emphasize case-by-case determination,
and in so doing, underscore flexibility in applying a multitude of
factors.372
To resist a movant’s summary judgment motion, the non-moving
party must establish, through pleadings, depositions, answers to
interrogatories, admissions, and affidavits in the record that there is no
genuine issue of material fact.373 To put it differently, summary
judgment needs to be based on undisputed material facts that show
there is “only one conclusion a trier of fact could reasonably draw.”374
The “factors require a fact-intensive inquiry not suitable for a motion
to dismiss.”375 Some appellate courts caution district courts to only
grant summary judgment on the issue of likelihood of confusion
“sparingly.”376
However, expedient determinations serve the ends of justice for
both sides in litigation. Summary judgments provide a quick and

368. See Zubulake v. UBS Warburg LLC, 217 F.R.D. 309, 322 (S.D.N.Y. 2003) (noting
the tendency towards this type of application).
369. Anthony E. Chavez, Using Legal Principles to Guide Geoengineering Deployment, 24
N.Y.U. ENVTL. L.J. 59, 93 (2016) (“Decision makers, however, often do not apply multi-
factor—or multi-principle—tests as they are intended.”).
370. Beebe, supra note 6, at 1601.
371. Noble v. United States, 231 F.3d 352, 359 (7th Cir. 2000).
372. John S. Applegate, Worst Things First: Risk, Information, and Regulatory Structure
in Toxic Substances Control, 9 YALE J. REG. 277, 302 (1992).
373. ServPro Intell. Prop., Inc. v. Blanton, 451 F. Supp. 3d 710, 722 (W.D. Ky. 2020).
374. Health Net v. U.S.A. Healthnet, Inc., No. CV 92-3925 KN, 1993 WL 209558, at
*1 (C.D. Cal. May 12, 1993).
375. GeigTech E. Bay LLC v. Lutron Elecs. Co., 352 F. Supp. 3d 265, 285 (S.D.N.Y.
2018); see also Eliya, Inc. v. Kohl’s Dep’t Stores, No. 06 Civ 195, 2006 WL 2645196, at
*13 n.2 (S.D.N.Y. Sept. 13, 2006) (“[A]n application of the so-called Polaroid factors on
this motion to dismiss would be inappropriate because it would involve premature
factfinding.”).
376. Fortune Dynamic, Inc. v. Victoria’s Secret Stores Brand Mgmt., Inc., 618 F.3d
1025, 1039 (9th Cir. 2010); see also Rearden LLC v. Rearden Com., Inc., 683 F.3d 1190,
1210 (9th Cir. 2012) (“Given the open-ended nature of this multi-prong inquiry, it is
not surprising that summary judgment on ‘likelihood of confusion’ grounds is
generally disfavored.”).
2022] TRADEMARK CONFUSION REVEALED 1349

inexpensive off-ramp for parties to dispose of a case when no real issues

call for a trial. The ability of courts to wield this important judicial tool
protects defendants against frivolous lawsuits and plaintiffs from
incurring unnecessary costs.377 Streamlining the test by consolidating
and trimming down the factors will enable courts to get to the heart of
the inquiry expeditiously. Simplifying the likelihood of confusion
lowers the temperature and makes it easier for owners to determine
when to protect their interests.
Many courts are more willing to move ahead with summary
judgments even where factors are in dispute and the evidence is not
obvious on the basis that “as with any other issue of fact, summary
judgment remains appropriate when no jury reasonably could have
ruled in the non-moving party’s favor.”378 Others are willing to do so
when most of the relevant factors weigh in the movant’s favor,
including at least one “key factor.”379 Yet others maintain “a finding of
a likelihood of confusion ‘need not be supported by a majority’ of the
digits.”380 Non-movants resisting summary judgment must show “how
additional discovery on these issues would create a genuine issue of
fact” material to movants’ claim for trademark infringement.381
In a Ninth Circuit case from the dataset, the court held as a matter
of law that the trademark owner was entitled to summary judgment
where the marks were identical, the goods were related, and the
marketing channels overlapped.382 A small set of key factors helps
structure the likelihood of confusion inquiry and gives notice of
pertinent issues and relevant evidence; this creates a more solid basis

377. Elaine Kussurelis, Canada’s Summary Trial Procedure: A Viable Alternative to

Summary Judgment on Trademark Likelihood of Confusion Actions in the United States, 50 U.
MIAMI INTER-AM. L. REV. 165, 168 (2019) (observing summary judgments “can be a
powerful trademark litigation weapon for either plaintiffs or defendants”).
378. RXD Media, LLC v. IP Application Dev. LLC, 986 F.3d 361, 375 (4th Cir. 2021);
see also EST Inc. v. Royal-Grow Prods., LLC, 526 F. Supp. 3d 943, 956–57 (D. Kan. 2021)
(“The evidence is far from one-sided and leads to no obvious answer.”); Collins v. U.S.
Dep’t of Veterans Affs., 497 F. Supp. 3d 885, 894 (S.D. Cal. 2020) (“[S]ummary
judgment is still proper in trademark infringement cases where, as here, no genuine
issue of material fact exists.”).
379. See, e.g., RXD Media, LLC, 986 F.3d at 375 (“Based on the record before us, we
hold that a jury could not have reasonably concluded that RXD’s use of the ‘ipad’ mark
was unlikely to cause consumer confusion.”).
380. Future Proof Brands, LLC v. Molson Coors Beverage Co., 982 F.3d 280, 298
(5th Cir. 2020).
381. Better Angels Soc’y, Inc. v. Inst. for Am. Values, Inc., 419 F. Supp. 3d 765, 780
(S.D.N.Y. 2019).
382. Nissan Motor Co. v. Nissan Comput. Corp., 378 F.3d 1002, 1019 (9th Cir.
2004).
1350 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

for predicting case outcomes and may even be sufficient to justify an

inference that there is a likelihood of confusion.383
Similarly, courts can also rule on a motion to dismiss under Rule
12(b)(6) if the complaint contains facts that state a claim to relief that
is plausible on its face that the plaintiff’s claims are barred as a matter
of law.384 Where plaintiffs can substantiate the plausibility of their
claims, courts will deny the motion to dismiss.385 Plaintiffs have
successfully done so on competitive proximity,386 actual confusion,387
mark strength,388 mark similarity, and “bridging the gap.”389
This Article reveals that courts rely on a small number of factors to
economize their decisions to give parties an early off-ramp. With either
party as the movant or cross-motion, summary judgments comprised
48% of cases in the dataset. Motions to dismiss by either party made up
10% of cases, and other postures, mostly preliminary injunction
motions, made up 21%. These collectively indicate that 79% of
litigants seek an early resolution. Of these, 63% received a resolution
with either plaintiff or defendant winning on the merits.
Some district courts treat the likelihood of confusion as a question
of fact, requiring proof of each element of each factor and
categorically precluding summary judgment.390 Others treat the

383. Future Proof Brands, LLC, 982 F.3d at 298.

384. Uber Inc. v. Uber Techs., Inc., 521 F. Supp. 3d 455, 462 (S.D.N.Y. 2021); see
also id. at 464 (“At the Rule 12(b)(6) stage, the Polaroid analysis is limited to the facts
alleged in the Complaint and any documents integral thereto.”).
385. See Ashcroft v. Iqbal, 556 U.S. 662, 678 (2009) (holding that “[t]o survive a
motion to dismiss,” a complaint need only “contain sufficient matter, accepted as true,
to ‘state a claim to relief that is plausible on its face.’” (quoting Bell Atlantic Corp. v.
Twombly, 550 U.S. 544, 570 (2007))).
386. Uber Inc., 521 F. Supp. 3d at 464 (“It suffices to note that the Complaint
plausibly alleges that defendants’ services are in competitive proximity with the
plaintiff’s graphic design and marketing services.”).
387. Id. (“[T]he Complaint’s descriptions of confusion among businesses, official
bodies and members of the public provide some factual support for the plausibility of
plaintiff’s claims.”).
388. Id. (“Among the other Polaroid factors, the Complaint plausibly alleges that
plaintiff’s Uber mark is arbitrary and distinctive, that the parties’ marks are similar,
and that defendants’ putative entry into the display-advertising market may result in
plaintiff bridging the ‘gap’ between its services and those of defendants.”).
389. Id.
390. Flower Mfg., LLC v. CareCo, LLC, 466 F. Supp. 3d 797, 820 (N.D. Ohio 2020)
(“Because each of Flower’s claims requires proof of that element, they cannot survive
summary judgment.”).
2022] TRADEMARK CONFUSION REVEALED 1351

likelihood of confusion as a matter of law, whether outright391 or

paying lip service to its factual dimensions.392 Appellate courts are split
along the same lines, either reviewing the lower courts’ “ultimate
conclusion about likelihood of success for clear error,”393 or “de novo,
using the same legal standards [the lower court] employed.”394
Even among those who profess fidelity to the fact/law distinction in
theory, applying that distinction in practice is not easy. As the Sixth
Circuit put it, “[a]ny dispute about the evidence that pertains to the
eight factors presents a factual issue . . . [and] whether a given set of
foundational facts establishes a likelihood of confusion is a legal
conclusion.”395 In contrast, the Second Circuit expressed that
determining whether one of the Polaroid factors favors one party or
another is a legal judgment reviewed de novo.396
The data also revealed a steady increase in the affirmance rate of
lower courts’ decisions between 2016 and 2021. This is because
appellate courts generally defer to lower court finding of facts but give
less deference to questions of law.397 But, more significantly, appellate
courts seem either unaware or complicit in the practice of lower courts
folding some factors and ignoring others. It would be interesting and
worth further study to see whether this is a practice of “wink-and-nod”
between the lower and appellate courts or if this state of affairs was
purely coincidental.

391. FCOA, LLC v. Foremost Title & Escrow Servs., LLC, 416 F. Supp. 3d 1381, 1387
(S.D. Fla. 2019) (“[I]n trademark infringement cases, courts in this Circuit have
decided the issue of likelihood of confusion as a matter of law.”).
392. Yellowfin Yachts, Inc. v. Barker Boatworks, LLC, 898 F.3d 1279, 1289 (11th Cir.
2018) (“Although likelihood of confusion is a question of fact, it may be decided as a
matter of law.”).
393. Future Proof Brands, LLC v Molson Coors Beverage Co., 982 F.3d 280, 298
(5th Cir. 2020); AWGI, LLC v. Atlas Trucking, Co., LLC, 998 F.3d 258, 264 (6th Cir.
2021) (“We review the district court’s finding of fact for clear error . . . .” (quoting
Premium Freight Mgmt., LLC v PM Eng’g Sols., Inc., 906 F.3d 403, 406 (6th Cir.
2018))).
394. Yellowfin Yachts, Inc., 898 F.3d at 1289.
395. Progressive Distrib. Servs., Inc. v. United Parcel Serv., Inc., 856 F.3d 416, 427
(6th Cir. 2017); see also Champions Golf Club, Inc. v. The Champions Golf Club, Inc.,
78 F.3d 1111, 1116 (6th Cir. 1996) (“Whether there is a likelihood of confusion is a
mixed question of fact and law.”).
396. Tiffany & Co. v. Costco Wholesale Corp., 971 F.3d 74, 86 (2d Cir. 2020)
(“[I]nsofar as the determination of whether one of the Polaroid factors favors one party
or another involves a legal judgment—which it often does—we must review that
determination de novo.”).
397. Select Comfort Corp. v. Baxter, 996 F.3d 925, 934 (8th Cir. 2021) (“[W]e
review the likelihood of confusion determination as a finding of fact.”).
1352 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

Figure 15: Affirmance on Appeal

2016 2017 2018 2019 2020 2021

100.0% 87.5% 80.0% 84.6% 81.3% 81.8%

Scholars have long debated whether AI can replicate human legal

reasoning.398 Edward Levi described how common law rules evolve in
his classic text An Introduction to Legal Reasoning.399 Judges begin by
identifying factors that have legal salience to explain case outcomes.
Once those rules fail to yield sensible results, judges alter them
accordingly.400 Legal reasoning rests on analogies, but this fact-specific
method also makes developing coherence in the case of precedent an
elusive task. For this reason, AI’s ability to detect patterns in judicial
opinions is of great interest to scholars.401
At its heart, the likelihood of confusion inquiry seeks to ascertain the
probability that a defendant’s use of its trademarks will confuse
consumers.402 Making the likelihood of confusion more rule-like, both
through the doctrinal reformation of the standard and through the
application of AI, makes it easier for appeals courts to scrutinize and
overturn deviant lower court decisions and allows lower courts to

398. See e.g., Stephen M. McJohn, Review of Artificial Legal Intelligence, 12 HARV. J.L. &
TECH. 241, 243–44 (1998) (noting that the way neural networks learn through
adjustment makes the technology capable of performing legal reasoning; however, the
technology would not understand the reasons behind its conclusion, making it less
valuable to the legal field).
399. Edward H. Levi, An Introduction to Legal Reasoning, 15 U. CHI. L. REV. 501, 501–
03 (1948).
400. Id.; see also RONALD DWORKIN, LAW’S EMPIRE 400–03 (1986) (explaining the
factors that a judge might weigh when considering whether to change the law
prospectively, noting that they are bound by precedent and the integrity of law).
401. See, e.g., McJohn, supra note 398, at 241 (offering commentary on another
scholar’s contribution to discussions around the topic).
402. Leah Chan Grinvald, Shaming Trademark Bullies, 2011 WIS. L. REV. 625, 636
(2011) (“This liability standard refers to the probability (not the actuality or
possibility) that consumers will be confused by the same or similar trademarks.”).
2022] TRADEMARK CONFUSION REVEALED 1353

distinguish dubious precedent based on facts.403 Trademark’s troika of

actual confusion, mark similarity, and competitive proximity paves the
road for AI to fill the final piece of the equation to simplify the
likelihood of confusion.

C. Deploying Artificial Intelligence

Legal scholarship on AI and trademarks is scarce.404 It is surprising
given AI’s centrality in both the consumer marketing literature and
trademark’s centrality in IP protection.405 Sonia Katyal and Aniket
Kesari argued “as a general matter, that AI should be of interest to
anyone studying trademarks and the role that they play in economic
decision making.”406 They point to AI deployment by the government
in trademark image recognition, classifying goods and services, and
identifying descriptive terms.407 This Article explains how a dataset
such as the one used here might be a rudimentary prototype for a
grander form of AI-enabled likelihood of confusion analysis that courts
and litigants might deploy in the future.

403. Frederick Schauer, Formalism, 97 YALE L.J. 509, 541–42 (1988) (noting errors
are more easily detectable under rules).
404. Sonia K. Katyal & Aniket Kesari, Trademark Search, Artificial Intelligence, and the
Role of the Private Sector, 35 BERKELEY TECH. L.J. 501, 504 (2020) (“[S]urprisingly, very
little legal scholarship has addressed the potential role for AI in the context of
trademarks.”). The few examples available include Dev Gangjee, Eye, Robot: Artificial
Intelligence and Trade Mark Registers, in TRANSITION AND COHERENCE IN INTELLECTUAL
PROPERTY LAW (Niklas Bruun, Graeme B. Dinwoodie, Marianne Levin & Ansgar Ohly
eds., 2021), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3467627; Anke
Moerland & Conrado Freitas, Artificial Intelligence and Trade Mark Assessment, in
ARTIFICIAL INTELLIGENCE & INTELLECTUAL PROPERTY (Jyh-An Lee, Reto M. Hilty, &
Kung-Chung Liu eds., 2021), https://papers.ssrn.com/sol3/papers.cfm?
abstract_id=3683807.
405. See Trademarks, Copyright and Patents: Should Business Owners Really Care About
IP?, VARNUM (May 1, 2019), https://www.varnumlaw.com/newsroom-publications-
trademarks-copyrights-and-patents-why-business-owners-should-care-about-ip
[https://perma.cc/VST8-XF56] (“A trademark is one of the most important business
assets that a company will ever own because it identifies and distinguishes the company
and its products/services in the marketplace from its competitors.”).
406. Katyal & Kesari, supra note 404, at 505 (“AI will fundamentally transform the
trademark ecosystem, and the law will need to evolve as a result.”).
407. See, e.g., DAVID FREEMAN ENGSTROM ET AL., GOVERNMENT BY ALGORITHM:
ARTIFICIAL INTELLIGENCE IN FEDERAL ADMINISTRATIVE AGENCIES 49–50 (2020), https://
www-cdn.law.stanford.edu/wp-content/uploads/2020/02/ACUS-AI-Report.pdf
[https://perma.cc/SR9J-UEUN] (describing a USPTO prototyping using deep
learning model using an unsupervised approach to generate visually similar images
from a database).
1354 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

1. Conception
AI gives courts the capability to scour case reports to assess how past
courts weighed effects and stress-test theories of confusion against real-
world data.408 AI can match the results against depositions and other
preprocessed evidence to provide quicker and more consistent
analyses, unlike the binarily coded factors in this study.409 Principal
component analysis can identify factors carrying the greatest weight in
functions and zero in on the most important dimensions of datasets to
show the stampeding likelihood of confusion factors.410
AI expands the scope of cases so that courts can dispense cases
summarily. It can significantly reduce the time and effort needed to
analyze a case, and courts can apply consistently evolving legal
principles, even when the facts are idiosyncratic.411 It can also avoid the
risk of judges engaging in side-by-side mark comparison and ensure
they apply the real-world purchasing context. The results from AI
recommendations challenge judges’ prior assumptions, providing a
check against coherence-based reasoning. Simon’s research shows that
confronting people with merits of the opposite side reduced the effect
of coherence shifts by about 50%.412 In particular, his study moderated
jury instruction by expressly requesting jury members to “take some
time to seriously consider the possibility that the opposite side has a
better case.”413 Other legal studies similarly showed that asking lawyers
to consider the weaknesses in their side or reasons that the judge might
rule against them mitigated bias.414
The beauty of AI-enabled likelihood of confusion analysis is that it
can reach outcomes we cannot define in advance of the AI being run
as “good” or “better” than the untrained neural network interrogates
itself via the process of trial and error. In addition, convolutional
neural networks can abstract local features from examples, for
instance, by recognizing specific facts in opinions. They would also
account for interactions among indicators that escape even expert

408. See Daryl Lim, Confusion, Simplified, BERKELEY TECH. L.J. (Forthcoming, 2022)
[hereinafter Lim, Confusion, Simplified].
409. Id.
410. Id.
411. Pamela Samuelson, Unbundling Fair Uses, 77 FORDHAM L. REV. 2537, 2541–42
(2009).
412. Simon, Third View of the Black Box, supra note 225, at 544 (noting that “[m]ore
studies are required to gain a better sense of the effects of the debiasing intervention”).
413. Id. at 571.
414. See Linda Babcock et al., Creating Convergence: Debiasing Biased Litigants, 22 L. &
SOC. INQUIRY 913, 920–21 (1997).
2022] TRADEMARK CONFUSION REVEALED 1355

witnesses and contextualize and associate information with known

factors to provide predictions based on untrained parameters.415
Finally, unsupervised data mining algorithms can zero in on data
clusters and probe those clusters to find other abstractions.416
Moreover, programming the AI to maximize reward in a
predetermined environment allows it to directly optimize policy
performance rather than learning from old data417 by updating the
agent’s policy using good estimates of a particular policy’s advantage
relative to another policy.418 Conceivably, variations of the algorithm
will predict litigation risk and the business implications of marketing
and sales decisions.419 As Dev Gangjee put it, “it is extremely tempting
to be guided by clearly defined percentages of similarity.”420 A. S. Li,
A.J.C. Trappey, and C.V. Trappey sketched out how that model might
work. The data set combines trademark litigation ontology and text
mining to extract features from cases to build a machine-readable
database like case content analysis.421

2. Execution
Like many AI datasets, case content analysis treats the content of
opinions as generic data.422 Coding and counting cases imply that

415. Similarly, AI-based support vector machines (SVMs) can find relationships
between sets of trademark infringement cases while handling outlier or mislabeled
cases, allowing SVM to crunch abrogated case law. See e.g., AURÉLIEN GÉRON, HANDS-ON
MACHINE LEARNING WITH SCIKIT-LEARN, KERAS, AND TENSORFLOW 155–67 (Nicole Tache
ed., 2d ed. 2019) (explaining how SVMs work and how they can be helpful).
416. PEDRO DOMINGOS, THE MASTER ALGORITHM 210 (2015).
417. Proximal Policy Optimization, OpenAI Spinning Up, OPENAI (2018) https://
spinningup.openai.com/en/latest/algorithms/ppo.html [https://perma.cc/GB72-
ZWGX].
418. See Brian S. Haney, AI Patents: A Data Driven Approach 19 CHI.-KENT J. INTELL.
PROP. 407, 439 (2020) (explaining the advantage function).
419. Katyal & Kesari, supra note 404, at 533 (“Indeed, predictive analytics can prove
to be transformative in helping businesses both create and sustain a strong presence
in the marketplace, predicting the outcome of filing suit, sending a cease-and-desist,
articulating various claims, or deciding whether and for how much to settle. And this
is just the tip of the iceberg. Imagine every aspect of a trademark claim--its probable
outcome automated, calculated, predicted and ready for real-time decision-making.”).
420. See Gangjee, supra note 404, at 13.
421. A. S. Li, A. J. C. Trappey, & C. V. Trappey, Intelligent Identification of Trademark
Case Precedents Using Semantic Ontology, in TRANSDISCIPLINARY ENGINEERING FOR COMPLEX
SOCIO-TECHNICAL SYSTEMS—REAL-LIFE APPLICATIONS (Jerzy Pokojski et al. eds. 2020).
422. Hall & Wright, supra note 29, at 83.
1356 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

information in one opinion is potentially relevant to another.423 In a

pre-AI world, an army of legal scholars might attempt to map all
likelihood of confusion cases comprehensively. Indeed, they would
need to endure many hours of time-consuming and repetitious
reading and extracting of the necessary information to code each case,
draw interferences, and report on trends, as was done in this Article’s
writing.
Case content analysis is suitable for automating because the same set
of information must be keyed into many cells in the same case. This
requires coders to eyeball each cell for accuracy given the tedious,
repetitive data entry, resulting in avoidable human errors and copy-paste
tasks.424 Nevertheless, as seen in this Article, the result is useful,
capable of determining the weight courts have placed on various legal
and non-legal factors, identifying which factors judges use to
“stampede” others, revealing trends across time, and other relevant
parametric factors that may typically escape conventional wisdom.425
The algorithm would pick out keywords and assign appropriate
weights to each variable with AI. For example, factors “in favor of”426 or
“favors”427 would signify a positive correlation to one side. Similarly,
phrases like “marks are strong,”428 “high degree of care,”429 “marks
are . . . identical,”430 and “weighs heavily in favor of,”431 would be
assigned greater, or in the case of “neutral,”432 “weighs neither for

423. See Peter J. Hammer & William M. Sage, Antitrust, Health Care Quality, and the
Courts, 102 COLUM. L. REV. 545, 561 (2002) (explaining that the coding exercise does
not determine the law, but instead treats opinions as data).
424. Serena Lim & Nandini Nayar Sharma, Document Drafting—Less Is More, SING. L.
GAZETTE (Nov. 21, 2021), https://lawgazette.com.sg/practice/tech-talk/document-
drafting-less-is-more.
425. See e.g., Barton Beebe, An Empirical Study of U.S. Copyright Fair Use Opinions,
1978–2005, 156 U. PA. L. REV. 549, 591 (2008) (describing a study of cases regarding
the fair use reasoning used by different courts).
426. Fletcher’s Original State Fair Corny Dogs, LLC v. Fletcher-Warner Holdings
LLC, 434 F. Supp. 3d 473, 485 (E.D. Tex. 2020).
427. E.g., Fleet Feet, Inc. v. Nike Inc., 419 F. Supp. 3d 919, 943 (M.D.N.C. 2019)
(“This factor favors Fleet Feet.”), appeal dismissed, 986 F.3d 458 (4th Cir. 2021).
428. New Balance Athletics, Inc. v. USA New Bunren Int’l Co., 424 F. Supp. 3d 334,
347 (D. Del. 2019).
429. Id. at 348.
430. Id. at 347.
431. Better Angels Soc’y, Inc. v. Inst. For Am. Values, Inc., 419 F. Supp. 3d 765, 777
(S.D.N.Y. 2019); Life After Hate, Inc. v. Free Radicals Project, Inc., 410 F. Supp. 3d
891, 908 (N.D. Ill. 2019).
432. Illinois Tool Works Inc. v. J-B Weld Co., 419 F. Supp. 3d 382, 399 (D. Conn.
2019).
2022] TRADEMARK CONFUSION REVEALED 1357

[n]or against,”433 “slightly in favor of,”434 would be afforded less weight.

The algorithm would also recognize and capture variables like rivalry
(“direct competitors”435).
Automation saves dataset preparers a substantial amount of time.
Studies on automating conveyancing work show a time savings of
90%.436 Less skilled and lower‑cost staff can quickly and accurately
generate datasets, lowering costs, time, and effort to produce complex
datasets, freeing up scholars to focus on higher-value work.437 The user
selects a smart template and answers a questionnaire presented by the
template to generate an opinion.438 The AI then uses the training data
to assemble a custom opinion.439
Likelihood of confusion opinions contain logic-dependent
conditional clause variations which incorporate the factors. The
algorithm could compare the qualitative and quantitative factors
presented in each case to its markers as a first step. Cases presenting
the same set of facts would reach the same outcome as precedential
cases presenting the same set of markers. AI will need to specify the
weight of factors not expressly entailed by rules or precedents. Once
algorithms produce their recommendation, judges could accept or
reject the AI’s recommendation, like how Amazon consumers choose
to make another purchase based on Amazon’s recommendations of
their earlier purchases and browsing history.440
The algorithm randomly plays out certain results, learns—with input
from data scientists in each iteration—adjusts its weights and
parameters, and chooses advantageous moves with increasing
finesse.441 The feedback loop causes the algorithm’s nodes to change
their weights, so case precedents refined by new case law and market

433. New Balance Athletics, Inc, 424 F. Supp. 3d, at 349.

434. Delta Forensic Eng’r, Inc. v. Delta V Biomechanics, Inc., 402 F. Supp. 3d 902,
910 (C.D. Cal. 2019).
435. Smartling, Inc. v. Skawa Innovation Ltd., 358 F. Supp. 3d 124, 139 (D. Mass.
2019).
436. Lim & Sharma, supra note 424.
437. Id.
438. Id.
439. Id.
440. See SPANDANA SINGH, NEW AMERICA, WHY AM I SEEING THIS? HOW VIDEO AND E-
COMMERCE PLATFORMS USE RECOMMENDATION SYSTEMS TO SHAPE USER EXPERIENCE 22
(Mar. 2020) (discussing how Amazon’s recommendation system drives user purchases
and contributes to revenue generation on the platform).
441. See TERRENCE J. SEJNOWSKI, THE DEEP LEARNING REVOLUTION 20 (2018)
(describing how the AlphaZero used machine learning algorithm to play chess).
1358 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

data may eventually yield a different, better outcome over time.442 This
allows adjudicating to become less a question of “ideology plus facts
plus law equal the outcome” and more a question of whether the data
supports the parties’ legal outcome or if on appeal, one that the lower
court advanced.

3. Limitations
As with any AI system, there are limitations, some generic, some
specific, that its implementers need to keep in mind. First, this Section
identifies the main limitations pertinent to the discussion: “garbage-in,
garbage-out,” biases, contextualizing purchasing conditions, and
coding challenges. Then, this Section discusses each one in turn.

a. “Garbage-in, garbage-out”
First, the saying “garbage-in, garbage-out” applies to the training
dataset. The algorithm applies the judge’s expertise through the
opinions coded in the training data while minimizing unreliability.443
The case law may be doctrinally flawed but remain good law.
Nonetheless, the algorithm can implement the likelihood of confusion
factors more consistently than both the human judges who decide the
precedential cases in the dataset and the judges applying those
precedents. Moreover, judges adjudicating live cases can compare the
model’s prediction with the ground truth and adjust the model’s
parameters, minimizing the error between these two values over time.
As algorithms gain additional knowledge about the probabilities of
occurrence, ambiguity disappears, and the choices become clearer.444
Scholars and AI service providers agree that AI augments human
decision-making and does not displace it.445 As LawPanel’s founder put

442. See Stephen McJohn & Ian McJohn, Fair Use and Machine Learning, 12 NE. U. L.
REV. 99, 135 (2020).
443. Dawes, supra note 216, at 575 (“[A] linear model distills underlying policy . . .
from otherwise variable behavior (e.g., judgments affected by context effects or
extraneous variables).”).
444. See Gary Charness & Dan Levin, When Optimal Choices Feel Wrong: A Laboratory
Study of Bayesian Updating, Complexity, and Affect, 95 AM. ECON. REV. 1300, 1300 (2005)
(describing a comparable heuristic form of processing new information).
445. See Gangjee, supra note 404, at 11 (“Experience till date therefore suggests that
AI algorithms are intended to augment human judgment—to effectively sift through
ever increasing volumes of registration data—and not to replace it.”); see also
COMPUMARK, ARTIFICIAL INTELLIGENCE, HUMAN EXPERTISE: HOW TECHNOLOGY AND
TRADEMARK EXPERTS WORK TOGETHER TO MEET TODAY’S IP CHALLENGES 5 (2018)
(“While AI and neural networks will play an expanding role in CompuMark
solutions . . . they are intended to complement, not replace, human analysts.”).
2022] TRADEMARK CONFUSION REVEALED 1359

it, “AI will speed up legal research, but it will not replace advice
formulation . . . [since it] only works on repetitive tasks in a very
tightly-defined domain.”446 Nevertheless, Katyal and Kesari are
optimistic that the gap can be closed as data scientists enrich the
dataset with more data points and human-AI teams.447 They report how
experts are continuing to highlight the need for human oversight and
participation, particularly when it comes to complex cognitive tasks in
trademark doctrines.448
On appeal, the variability of decisions can reveal some idea of the
extent of noise. A three-judge circuit appeals court or nine-justice
Supreme Court bench would provide an additional check. Salib
observes that “there will be an adjustment period as courts develop
doctrine about what constitutes credible scientific practice in
algorithmic design. Such bumps on the road, however, are the cost of
admission if generalist judges are to continue playing any major role
governing our increasingly complex world.”449

b. Biases
Second, system architects need to address data biases in adopting
the technology and in deploying AI. For example, with supervised
machine learning, humans classify the data. This introduces bias, such
as training an AI on the similarity of signs. One trainer might
determine a similarity between two given signs, while another might
not.450 As a result, AI may replicate and perpetuate data biases.451
Coding is not value-neutral, and biases may seep into the algorithmic
code, filtering into training data and the weights judges assign to the
algorithm.452 Bias could also come from the algorithms being trained

446. Tim Lince, “No Imminent AI Apocalypse”—Tech Expert Rejects Predictions of Mass
Job Losses in Trademark Industry, WORLD TRADEMARK REV. (Feb. 1, 2018), https://www.
worldtrademarkreview.com/brand-management/no-imminent-ai-apocalypse-tech-
expert-rejects-predictions-mass-job-losses [https://perma.cc/RAF4-FHPV].
447. See Katyal & Kesari, supra note 404, at 526.
448. Id. at 533.
449. Peter N. Salib, Artificially Intelligent Class Actions, 100 TEX. L. REV. 519, 552
(2022).
450. See Lim, Confusion, Simplified, supra note 408.
451. See id.
452. See Dan L. Burk, Algorithmic Fair Use, 86 U. CHI. L. REV. 283, 283 (2019)
(describing how design values of algorithms can reflect biases); see, e.g., David Lehr &
Paul Ohm, Playing with the Data: What Legal Scholars Should Learn About Machine
Learning, 51 U.C. DAVIS L. REV. 653, 669–701 (2017) (explaining that because humans
make algorithms and humans have bias, the biases of humans are reflected in both
algorithms themselves and how humans use them).
1360 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

using biased data, such as prior decisions from judges who are biased
themselves, and from the way humans interpret the data produced by
AI systems.453 In addition, reinforcement learning techniques may
embed bias, raising the risk of what Thomas Nachbar labeled
“snowballing unfairness.”454 Codes are based on earlier program
decisions and the constant integration of new information, prompts a
continual search for purpose.455
Moerland and Freitas provide an example of bias in action:
[W]hen teaching an AI to establish a pattern of similarity of signs,
one could easily ascertain a similarity between two given signs, while
someone else would not. Even if case law regarding similarity of signs
is used as training data, courts sometimes come to differing
outcomes for the same cases. Bias in data will be replicated when
used by the AI technology, as it lacks the ability to filter out slightly
incorrect interpretations.456
The lack of a standardized method to weigh factors systematically
exacerbates the risk of bias. The likelihood of confusion factors have
no weights assigned, eroding the ability to apply the tests objectively or
in a manner that can be replicated.457 AI helps integrate data and
provides a statistical prediction based on input variables. Humans are
superior at selecting and coding information but poor at integrating
it.458
Daniel Kahneman, Cass Sunstein, and Olivier Sibony recommend
assigning probabilities rather than absolute values or binary “yes” or
“no” judgments.459 Numerical thresholds could help adjudicate
infringement cases. For example, computer scientists could build a
model that requires judges to rate the three core likelihood of
confusion factors on a scale of 0–10. If the marks were completely
different, the judge would rate it ‘0’ (the lowest rating possible), but if

453. See Gangjee, supra note 404, at 11 (“[W]here the data for a machine learning
approach is derived from judicial content analysis—past decisions by human tribunals
where factors are coded and correlations derived—the algorithm will behave like the
human decision maker it is modelled after, warts and all.”).
454. Thomas B. Nachbar, Algorithmic Fairness, Algorithmic Discrimination, 48 FLA.
STATE L. REV. 509, 522 (2021).
455. Id. at 548.
456. Moerland & Freitas, supra note 404, at 282.
457. See Menard, Inc. v. Comm’r of Internal Revenue, 560 F.3d 620, 622–23 (7th
Cir. 2009) (“Multifactor tests with no weight assigned to any factor are bad enough
from the standpoint of providing an objective basis for a judicial decision; multifactor
tests when none of the factors is concrete are worse.”).
458. See, e.g., Dawes, supra note 216, at 573.
459. See KAHNEMAN, SIBONY & SUNSTEIN, supra note 176, at 218.
2022] TRADEMARK CONFUSION REVEALED 1361

the marks were simple counterfeits, the judge would rate it ‘10’ (the
highest rating). Thus, the algorithm would set a numerical threshold
for finding confusion that maps to case law and the balance of
probabilities. Over time, the algorithm would provide more granular
information about the characteristics driving outcomes in likelihood
of confusion cases. In this way, the algorithm would imitate judges,
granting a low score to a particular factor and a consequently lower
success rate to plaintiffs.
Daniel Kahneman, Cass Sunstein, and Olivier Sibony also
recommend relying more heavily on rules like judicial sentencing
guidelines.460 The trio of factors again provides that framework.
Importantly, the results from AI recommendations challenge judges’
prior assumptions, providing a check against coherence-based
reasoning.461 For instance, confronting people with merits of the
opposite side reduced the effect of coherence shifts by about 50%.462
Legal studies similarly showed that asking lawyers to consider the
weaknesses in their side or reasons that the judge might rule against
them mitigated bias.463
Finally, to address the issue of “snowballing unfairness,” flooding the
system with voluminous data may help. As Moerland and Freitas note,
“with large amounts of data, incidental bias may not influence the rule
that the AI learns from the data.”464 They reassuringly report that “[AI]
training is continuous and subject to high standards of reliability. Error
measures are used as well as pilot studies on unseen data to determine
how the AI tool performs its tasks.”465
Done well, trademark algocracy will minimize biases from human
decision-making without compounding those biases with its own.466 In
the years ahead, ethics teams will likely become an essential
department in antitrust agencies and economic consultancies such as
finance, legal, marketing, and human resource departments. These

460. Id. at 258.

461. See generally Lim, Saving Substantial Similarity, supra note 281, at 634 (discussing
the safety valve role of juries in challenging the assumptions of judges).
462. Simon, Third View of the Black Box, supra note 225, at 544 (noting that “[m]ore
studies are required to gain a bettr sense of the effects of the debiasing intervention”).
463. Babcock et al., supra note 414, at 920–21 (describing results from an
experiment which showed that when subjects consciously considered weaknesses in
their cases, their biases were mitigated).
464. See Moerland & Freitas, supra note 404, at 282.
465. Id. at 281.
466. See Meghan J. Ryan, Secret Conviction Programs, 77 WASH. & LEE L. REV. 269, 286
(2020) (describing how computer programs can effectively help judges avoid injecting
their own biases when making judicial decisions).
1362 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

teams can help decision-makers weigh the benefits and harms of AI

procedures and recommendations, flag their implications, develop
guidelines, and help clarify ethical conflicts.467

c. Contextualizing the purchasing conditions

Third, the algorithm needs to replicate how a human perceives a
mark in the marketplace.468 As we saw in Part I, a human fact-finder
trying to contextualize the marketplace faces a difficult task. AI has the
additional burden of delivering on that promise of an objective
assessment.469 The algorithm will need to account for the relevant
consumer type, competitors, circumstances of purchase, and the end-
use of the product. The AI tool is unlikely to contextualize and
juxtapose them against case law to compare situations and determine
a likely outcome. The data scientist will need to acquire this
information and structure it in a manner that the algorithm can
automatically process.
One way to contextualize purchasing conditions is for AI to
maximize a preset reward without the need for continual human
supervision. In this case, the reward is whether, on balance, consumers
would be confused.470 Instead, the algorithm chooses an action in the
environment’s initial state—representing a moment in time—randomly
explores the environment, gathers information about the environment,
develops an optimal policy, and optimizes performance by “expressing
the relationship between the value of a state and the values of future
states.”471

467. Lim Sun & Jeffrey Chan Kok Hui, Moving AI Ethics Beyond Guidelines, STRAITS
TIMES (Dec. 16, 2020), https://www.straitstimes.com/opinion/moving-ai-ethics-
beyond-guidelines-0 [https://perma.cc/R4WD-76TD].
468. Moerland & Freita, supra note 404, at 284 (“This leads us to the finding that
the assessment is one of degree and requires reasoning from the perspective of the
relevant public. It is questionable as to how far AI technology can reflect this human-
centric approach.”).
469. Katyal & Kesari, supra note 404, at 532 (“Others have expressed similar
concerns, noting that determining trademark distinctiveness, the relevant public, the
proper classification of goods and services, among other elements, are so subjective
that they pose challenges to the development of AI in trademark law.”).
470. See Haney, supra note 418, at 430 (describing how reward can act as a feedback
mechanism).
471. Id. at 437.
2022] TRADEMARK CONFUSION REVEALED 1363

Figure 16: Reinforcement Learning472

As it continues to the next state, the agent receives a reward and a

set of choices, the algorithm selects an action, and the environment
returns a reward and the next state.473 The reward teaches the
algorithm what to do and formalizes the goal’s idea.474 Through this
iteration, it learns to take actions optimizing a reward, which would be,
say, mark similarity.475 In essence, the total reward mirrors the legal
“algorithm” we call trademark law’s likelihood of confusion.
This feature allows the algorithm to navigate dynamic market
environments without stopping the environment before computing.476
To the extent variables in its dataset need modification, AI training
techniques use autoencoders to update word embeddings, machine
translation, document clustering, sentiment analysis, and paraphrase
detection.477 Stacking autoencoders on top of each other allows the
first autoencoder to focus on encoding features at one level of
abstraction.478 The next autoencoder uses the earlier output to
recognize fact patterns and encode more abstract features.479 Defining
features broadly helps avoid overfitting, which happens when the

472. RICHARD S. SUTTON & ANDREW G. BARTO, REINFORCEMENT LEARNING: AN

INTRODUCTION 48 (2018).
473. EUGENE CHARNIAK, INTRODUCTION TO DEEP LEARNING 113 (2018).
474. Id.
475. Jennifer Barry, Daniel T. Barry, & Scott Aaronson, Quantum Partially Observable
Markov Decision Processes, PHYS. REV., No. 032311, 2014, at 1, 2, https://
journals.aps.org/pra/abstract/10.1103/PhysRevA.90.032311.
476. Id.
477. See Venkata Krishna Jonnalagadda, Sparse, Stacked and Variational Autoencoder,
MEDIUM (Dec. 6, 2018), https://medium.com/@venkatakrishna.jonnalagadda/
sparse-stacked-and-variational-autoencoder-efe5bfe73b64 [https://perma.cc/ZH2D-
JTNW] (describing the various types of and uses for autoencoders).
478. Id.
479. Gideon Lewis-Kraus, The Great A.I. Awakening, N.Y. TIMES, (Dec. 14, 2016),
https://www.nytimes.com/2016/12/14/magazine/the-great-ai-awakening.html
[https://perma.cc/46B3-8DKV].
1364 AMERICAN UNIVERSITY LAW REVIEW [Vol. 71:1285

learner fits the function to the data.480 Overfitting also happens in legal
reasoning when one ties a rule to the facts. The solution is to include
more training examples and test the function against other test
examples.481

d. Coding challenges
Fourth, the likelihood of confusion factors do not currently lend
themselves to easy coding by a machine, given the coherence-based
reasoning and non-uniformity in how courts operationalize those
factors, as discussed in this Article. Finally, the Eighth Circuit reminds
us that “factors do not operate in a mathematically precise formula.”482
Again, the issue is real but not insurmountable. Courts can do their
part by employing more rule-like formulations when applying the
likelihood of confusion standard, such as the “rules of thumb”
advanced in this Article. Courts can also standardize their lexicon,
enabling them to present their judicial opinions in a way more
amenable to machine learning. Finally, courts could and should also
standardize their treatment of absent factors in the likelihood of
confusion inquiry—do these factors favor either party and if so, in what
way?483 This templating exercise helps rationalize and consolidate
disparate variations into a reusable asset that captures and preserves
the substantial knowledge of experienced judges.
Finally, the algorithm will need to distinguish between cases from
courts at different levels of the judicial hierarchy. Stare decisis tells us
that Supreme Court cases take precedence over court of appeal cases,
which in turn take precedence over district court cases. However,
empirical legal studies routinely ignore the weight stare decisis endows
in coding datasets.484 It matters little if the Supreme Court or a district
court looked at likelihood of confusion if the variable of interest is
competitive proximity. The algorithm will need to consider judicial
hierarchy, the appellate jurisdiction of regional circuit courts, and
similar factors as appropriate.

480. GÉRON, supra note 415, at 26–28.

481. Id. at 29.
482. Frosty Treats, Inc. v. Sony Comput. Ent. Am., Inc., 426 F.3d 1001, 1008 (8th
Cir. 2005).
483. Easy Spirit, LLC v. Skechers U.S.A., Inc., 515 F. Supp. 3d 47, 74 (S.D.N.Y. 2021)
(explaining that absence of actual confusion favors junior user).
484. See e.g., Lim, Saving Substantial Similarity, supra note 281.
2022] TRADEMARK CONFUSION REVEALED 1365

CONCLUSION
Congress built a degree of indeterminacy into the likelihood of
confusion standard as a feature and not a bug. Over the years, however,
the jurisprudential roots of trademark law became unruly and tangled.
Unwanted variability and bias in judgments cause serious problems by
including complex and irrelevant factors, including financial loss and
rampant unfairness. Meanwhile, simple rules and algorithms have
developed with technological strides presenting big advantages over
human judges. Three core factors, combined with two safe harbors and
AI, would enable courts to reach consistent and accurate results. A
simplified framework promotes fair play, safeguards expressive uses,
and enhances access to justice.
This Article presented a contemporary empirical analysis of the
likelihood of confusion factors and how they interact. Conventional
wisdom teaches us that courts should comprehensively traverse each
factor and that likelihood of confusion cases generally require jury
determination. The data reveals that neither is true. Instead, courts
provide early off-ramps to litigants by “economizing” using a handful
of factors or by “folding” factors within each other. The findings also
indicate which forums are pro-defendant and which are pro-plaintiff,
the impact of rivalry and fair use on outcomes, and the Ninth Circuit’s
dominance. This Article also showed how AI systems could use
empirical studies as training data to help stakeholders make likelihood
of confusion analyses. A familiar yet concise, precise, and efficient
framework helps preempt, counsel, and adjudicate disputes. In this
way, the likelihood of confusion standard can attain the amphibious
benefits of becoming more rule-like while retaining its suppleness.