Psychological targeting as an effective approach to
digital mass persuasion
S. C. Matza,1, M. Kosinskib,2, G. Navec, and D. J. Stillwelld,2
a
Columbia Business School, Columbia University, New York City, NY 10027; bGraduate School of Business, Stanford University, Stanford, CA 94305;
c
Wharton School of Business, University of Pennsylvania, Philadelphia, PA 19104; and dCambridge Judge Business School, University of Cambridge,
Cambridge, CB2 3EB, United Kingdom
Edited by Susan T. Fiske, Princeton University, Princeton, NJ, and approved October 17, 2017 (received for review June 17, 2017)
People are exposed to persuasive communication across many
different contexts: Governments, companies, and political parties
use persuasive appeals to encourage people to eat healthier,
purchase a particular product, or vote for a specific candidate.
Laboratory studies show that such persuasive appeals are more
effective in influencing behavior when they are tailored to individ-
uals’ unique psychological characteristics. However, the investiga-
tion of large-scale psychological persuasion in the real world has
been hindered by the questionnaire-based nature of psychological
assessment. Recent research, however, shows that people’s psycho-
logical characteristics can be accurately predicted from their digital
footprints, such as their Facebook Likes or Tweets. Capitalizing on
this form of psychological assessment from digital footprints, we
test the effects of psychological persuasion on people’s actual be-
havior in an ecologically valid setting. In three field experiments
that reached over 3.5 million individuals with psychologically tai-
lored advertising, we find that matching the content of persuasive
appeals to individuals’ psychological characteristics significantly al-
tered their behavior as measured by clicks and purchases. Persuasive
appeals that were matched to people’s extraversion or openness-to-
experience level resulted in up to 40% more clicks and up to 50%
more purchases than their mismatching or unpersonalized counter-
parts. Our findings suggest that the application of psychological
targeting makes it possible to influence the behavior of large
groups of people by tailoring persuasive appeals to the psycholog-
ical needs of the target audiences. We discuss both the potential
benefits of this method for helping individuals make better deci-
sions and the potential pitfalls related to manipulation and privacy.
|
persuasion digital mass communication | psychological targeting |
|
personality targeted marketing
P ersuasive mass communication is aimed at encouraging large
groups of people to believe and act on the communicator’s
viewpoint. It is used by governments to encourage healthy be-
haviors, by marketers to acquire and retain consumers, and by
political parties to mobilize the voting population. Research
suggests that persuasive communication is particularly effective
when tailored to people’s unique psychological characteristics
and motivations (1–5), an approach that we refer to as psycho-
logical persuasion. The proposition of this research is simple yet
powerful: What convinces one person to behave in a desired way
might not do so for another. For example, matching computer-
generated advice to participants’ dominance level elicited higher
ratings of source credibility and increased the likelihood of
participants changing their initial opinions in response to the
advice (2). Similarly, participants’ positive attitudes and purchase
intentions were stronger when the marketing message for a
mobile phone was tailored to their personality profile (4). While
these studies provide promising evidence for the effectiveness of
psychological persuasion, their validity is limited by the fact that
they were mainly conducted in small-scale, controlled laboratory
settings using self-report questionnaires. Self-reports are known
to be affected by a whole range of response biases (6), and there
are numerous reasons why people’s natural behavior might differ
12714–12719 | PNAS | November 28, 2017 | vol. 114 | no. 48
from that displayed in the laboratory (7). Consequently, it is
questionable whether—and to what extent—these findings can be
generalized to the application of psychological persuasion in real-
world mass persuasion (see ref. 8 for initial evidence).
A likely explanation for the lack of ecologically valid research
in the context of psychological persuasion is the questionnaire-
based nature of psychological assessment. Whereas researchers
can ask participants to complete a psychological questionnaire in
the laboratory, it is unrealistic to expect millions of people to do
so before sending them persuasive messages online. Recent re-
search in the field of computational social sciences (9), however,
suggests that people’s psychological profiles can be accurately pre-
dicted from the digital footprints they leave with every step they
take online (10). For example, people’s personality profiles have
been predicted from personal websites (11), blogs (12), Twitter
messages (13), Facebook profiles (10, 14–16), and Instagram pic-
tures (17). This form of psychological assessment from digital foot-
prints makes it paramount to establish the extent to which behaviors
of large groups of people can be influenced through the application
of psychological mass persuasion—both in their own interest (e.g.,
by persuading them to eat healthier) and against their best interest
(e.g., by persuading them to gamble). We begin this endeavor in a
domain that is relatively uncontroversial from an ethical point of
view: consumer products.
Significance
Building on recent advancements in the assessment of psycho-
logical traits from digital footprints, this paper demonstrates the
effectiveness of psychological mass persuasion—that is, the ad-
aptation of persuasive appeals to the psychological characteris-
tics of large groups of individuals with the goal of influencing
their behavior. On the one hand, this form of psychological mass
persuasion could be used to help people make better decisions
and lead healthier and happier lives. On the other hand, it could
be used to covertly exploit weaknesses in their character and
persuade them to take action against their own best interest,
highlighting the potential need for policy interventions.
Author contributions: S.C.M. and M.K. designed research; S.C.M., M.K., and D.J.S. per-
formed research; S.C.M. analyzed data; and S.C.M., M.K., G.N., and D.J.S. wrote the paper.
Conflict of interest statement: D.J.S. received revenue as the owner of the myPersonality
Facebook application until it was discontinued in 2012. Revenue was received from dis-
playing ads within the application and charging for a premium personality test. The
revenue received by D.J.S. was unrelated to the studies presented in this paper, and there
will be no future revenue generated from the application. None of the authors received
any compensation for working on the marketing campaigns used to collect data for the
studies presented in this manuscript.
This article is a PNAS Direct Submission.
This open access article is distributed under Creative Commons Attribution-NonCommercial-
NoDerivatives License 4.0 (CC BY-NC-ND).
Data deposition: The data reported in this paper have been deposited in the Open Science
Framework (OSF), https://osf.io/srjv7/?view_only=90316b0c2e06420bbc3a1cc857a9e3c7.
1
To whom correspondence should be addressed. Email: sm4409@gsb.columbia.edu.
2
M.K. and D.J.S. contributed equally to this work.
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
1073/pnas.1710966114/-/DCSupplemental.
www.pnas.org/cgi/doi/10.1073/pnas.1710966114
 Capitalizing on the assessment of psychological traits from dig-
ital footprints, we conducted three real-world experiments that
reached more than 3.7 million people. Our experiments demon-
strate that targeting people with persuasive appeals tailored to
their psychological profiles can be used to influence their behavior
as measured by clicks and conversions. [Click-through rates
(CTRs) are a commonly used digital marketing metric that quan-
tifies the number of clicks relative to number of times the ad was
shown. Conversion rate is a marketing metric that reflects number
of conversions, such as app downloads or online store purchases,
relative to the number of times the ad was shown.] The experi-
ments were run using Facebook advertising, a typical behavioral
targeting platform. As of now, Facebook advertising does not allow
marketers to directly target users based on their psychological
traits. However, it does so indirectly by offering the possibility to
target users based on their Facebook Likes. (Facebook users can
like content such as Facebook pages, posts, or photos to express
their interest in a wide range of subjects, such as celebrities, poli-
ticians, books, products, brands, etc. Likes are therefore similar to
a wide range of other digital footprints—such as web-browsing
logs, purchase records, playlists, and many others. Hence, the
findings based on Facebook Likes are likely to generalize to digital
footprints employed by other advertising platforms.) For example,
if liking “socializing” on Facebook correlates with the personality
trait of extraversion and liking “stargate” goes hand in hand with
introversion, then targeting users associated with each of these
Likes allows one to target extraverted and introverted user seg-
ments (see SI Appendix for a validation of this method).
Studies 1 and 2 target individuals based on their psychological
traits of extraversion and openness-to-experience (18, 19). We
chose these two because they show strong associations with
Facebook Likes (14) and have been frequently investigated in the
context of consumer preferences and persuasive communication
(e.g., ref. 2). We extracted lists of Likes indicative of high and low
levels of each of these traits from the myPersonality.org database
(20). MyPersonality contains the Facebook Likes of millions of
users alongside their scores on the 100-item International Per-
sonality Item Pool (IPIP) questionnaire, a widely validated and
used measure of personality (19). We computed the average
personality trait levels for each Like and selected 10 Likes char-
acterized by the highest and lowest aggregate extraversion and
openness scores (i.e., target Likes). For example, the list of intro-
verted target Likes included Stargate and “Computers,” while the
list of extraverted target Likes contained “Making People Laugh”
or “Slightly Stoopid.” The list of target Likes for low openness
included “Farm Town” and “Uncle Kracker,” while the list for
high openness contained “Walking Life” and “Philosophy” (for
the full lists of target Likes, see SI Appendix, Tables S1 and S7).
Study 3 builds on the findings of studies 1–2 and shows how
psychological persuasion can be used in the context of predefined
behavioral audiences.
Results
Study 1 demonstrates the effects of psychological persuasion on
people’s purchasing behavior. We tailored the persuasive ad-
vertising messages for a UK-based beauty retailer to recipients’
extraversion, a personality trait reflecting the extent to which
people seek and enjoy company, excitement, and stimulation
(18). People scoring high on extraversion are described as en-
ergetic, active, talkative, sociable, outgoing, and enthusiastic;
people scoring low on extraversion are characterized as shy, re-
served, quiet, or withdrawn. Given the specific nature of the
product, we only targeted women. Fig. 1A displays 2 (out of 10)
ads aimed to appeal to women characterized by high versus low
extraversion (we refer to the personality of the audience an ad is
aimed at as ad personality). Using a 2 (Ad Personality: Introverted
vs. Extraverted) × 2 (Audience Personality: Extraverted vs. Intro-
verted) between-subjects, full-factorial design, we ran the Facebook
Matz et al.
A High Extraversion Low Extraversion
Dance like no one's watching Beauty doesn’t have to shout
(but they totally are)
B High Openness Low Openness
Aristoteles? The Seychelles? Unleash your Settle in with an all-time favorite! The
creativity and challenge your imagination with crossword puzzle that has challenged players
an ulimited number of crossword puzzles! for generations.
SOCIAL SCIENCES
Fig. 1. Examples of ads aimed at audiences characterized by high and low
extraversion (A) as well as high and low openness (B). Fig. 1A, Left courtesy
of Caiaimage/Paul Bradbury/OJO+/Getty Images; Fig. 1A, Right courtesy of
Hybrid Images/Cultura/Getty Images.
advertising campaign over the course of 7 d—that is, we placed
the ads on viewers’ Facebook pages as they browsed freely.
Together, the campaign reached 3,129,993 users, attracted
10,346 clicks, and resulted in 390 purchases on the beauty re-
tailer’s website. Table 1 (study 1) provides a detailed overview of
the descriptive campaign statistics across ad sets (see SI Appendix
for more detailed breakdowns). We conducted hierarchical lo-
gistic regression analyses for clicks (click = 1, no click = 0) and
conversions (conversion = 1, no conversion = 0), using the au-
dience personality, the ad personality, and their two-way in-
teraction as predictors. [All of the results reported in this paper
hold when using linear probability models or when testing for
main treatment effects for congruent vs. incongruent conditions
using Chi-square tests, demonstrating the effects’ robustness to
model specification (19, 21).] Users were more likely to purchase
after viewing an ad that matched their personality [Fig. 2;
Audience Personality × Ad Personality interaction; B = 0.90,
SE(B) = 0.21, z = 4.30, P < 0.001]. These effects were robust
after controlling for age and its interactions with ad person-
ality. Averaged across the campaigns, users in the congruent
conditions were 1.54 times more likely to purchase from the online
store than users in the incongruent conditions, χ2(1) = 17.72, odds
ratio (OR) = 1.54 [1.25–1.90], P < 0.001. There was no significant
interaction effect on clicks, χ2(1) < 0.001, OR = 1.0 [0.96–1.04],
P = 0.98.
Study 2 replicates and extends the findings of study 1 by tai-
loring the persuasive advertising messages for a crossword app to
recipients’ level of openness, a personality trait reflecting the
extent to which people prefer novelty over convention (18).
People scoring high on openness are described as intellectually
curious, sensitive to beauty, individualistic, imaginative, and
unconventional. People scoring low on openness are traditional
and conservative and are likely to prefer the familiar over the
unusual. Using the same targeting approach and experimental
design as study 1, we created tailored advertising messages for
both high and low openness (Fig. 1B). The campaign was run on
Facebook, Instagram, and Audience Networks for 12 d.
The campaign reached 84,176 users, attracted 1,130 clicks, and
resulted in 500 app installs. Table 1 (study 2) provides a detailed
PNAS | November 28, 2017 | vol. 114 | no. 48 | 12715
 Table 1. Descriptive statistics of studies 1–3 across ad sets
Condition Reach Clicks CTR Conv CR CPConv ROI

Study 1
Introverted ads 762,197 2,637 0.35% 121 0.016% £7.80 409%
congruent
Introverted ads 791,270 2,426 0.31% 90 0.011% £10.41 300%
incongruent
Extroverted ads 814,308 2,573 0.32% 117 0.014% £8.32 410%
congruent
Extroverted ads 762,218 2,710 0.36% 62 0.008% £15.93 219%
incongruent
Total 3,129,993 10,346 0.33% 390 0.012% £9.85 334%
Study 2
High-openness ad 29,277 427 1.45% 140 0.48% $2.29 —
congruent
High-openness ad 8,926 112 1.25% 37 0.41% $2.71 —
incongruent
Low-openness ad 18,210 296 1.62% 174 0.96% $1.38 —
congruent
Low-openness ad 27,763 295 1.06% 149 0.53% $1.76 —
incongruent
Total 84,176 1,130 1.34% 500 0.59% $1.85 —
Study 3
Standard copy 324,770 1,830 0.56% 1,053 0.32% $3.21 —
Personality-tailored 209,480 1,537 0.73% 784 0.37% $2.91 —
copy
Total 534,250 3,367 0.63% 1,837 0.34% $3.10 —

CPConv = cost per conversion, CR = conversion rate (installs/reach × 100), CTR = click-through rate (clicks/reach ×
100), ROI = return on Investment (profits/spending × 100).

overview of the descriptive campaign statistics across ad sets (see
SI Appendix for more detailed breakdowns). Using the same hi-
erarchical logistic regression analyses as in study 1, we found
significant interaction effects of audience personality and ad
personality on both clicks [B = 0.58, SE(B) = 0.14, z = 4.31, P <
0.001] and conversions [Fig. 2; B = 0.72, SE(B) = 0.22, z = 3.35,
P < 0.001]. These effects were robust after controlling for age,
gender, and their interactions with ad personality. Averaged across
the campaigns, users in the congruent conditions were 1.38 times
more likely to click, χ2(1) = 26.68, OR = 1.38 [1.22–1.56], P < 0.001,
and 1.31 times more likely to install the app, χ2(1) = 8.34, OR =
1.31 [1.09–1.58], P = 0.004, than users in the incongruent condi-
tions. As Fig. 2 illustrates, the significant interaction effect on
installs was mainly driven by the target audiences characterized as
low openness. While people scoring low on openness installed the
app significantly more often when presented with the matching
marketing message, χ2(1) = 22.72, OR = 0.43 [0.29–0.62], P <
0.001, there was no significant difference in install rates to
matching versus mismatching messages among people scoring
high on openness, χ2(1) = 0.97, OR = 0.89 [0.70–1.13],
P = 0.325.
Study 3 builds on the findings of studies 1 and 2 and shows how
psychological persuasion can be used in the context of predefined
audiences (e.g., when marketers have already established a behav-
ioral target group or when health promotions are targeted at a
specific subpopulation at risk). Promoting a bubble shooter game, we

Fig. 2. Interaction effects of audience and ad personality on conversion rates in study 1 (Left) and study 2 (Right).

12716 | www.pnas.org/cgi/doi/10.1073/pnas.1710966114 Matz et al.

followed the company’s preexisting behavioral targeting approach
and targeted the game at Facebook users who were connected with a
selected list of similar games (e.g., Farmville or Bubble Popp).
Mapping these behavioral target Likes onto those available within
the myPersonality database allowed us to identify the psychological
profile of this audience. Building on the finding that the target
audience was highly introverted (Ez = −0.25), we promoted the app
comparing the company’s standard persuasive advertising message
(“Ready? FIRE! Grab the latest puzzle shooter now! Intense ac-
tion and brain-bending puzzles!”) to a psychologically tailored one
(“Phew! Hard day? How about a puzzle to wind down with?”).
Both campaigns were run over the course of 7 d on Facebook.
Together, the campaign reached 534,250 users, attracted
3,173 clicks, and resulted in 1,837 app installs. Table 1 (study 3)
provides a detailed overview of the descriptive campaign statistics
across ad sets. Corresponding to our hypothesis, two χ2 tests
showed that the psychologically tailored ad set attracted signifi-
cantly more clicks, χ2(1) = 58.66, OR = 1.30 [1.22–1.40], P < 0.001,
and installs, χ2(1) = 9.16, OR = 1.15 [1.05–1.27], P = 0.002, than
the standard ad set. CTRs and conversion rates were 1.3 and
1.2 times higher when the persuasive advertising message was
tailored to the psychological profile of the preexisting behavioral
audience (see SI Appendix for evidence that this effect is not ex-
clusively due to the fact that the tailored advertising message was
generally more appealing).
Discussion
The results of the three studies provide converging evidence for
the effectiveness of psychological targeting in the context of real-
life digital mass persuasion; tailoring persuasive appeals to the
psychological profiles of large groups of people allowed us to
influence their actual behaviors and choices. Given that we ap-
proximated people’s psychological profiles using a single Like
per person—instead of predicting individual profiles using peo-
ple’s full history of digital footprints (e.g., refs. 10 and 14)—our
findings represent a conservative estimate of the potential ef-
fectiveness of psychological mass persuasion in the field.
The effectiveness of large-scale psychological persuasion in
the digital environment heavily depends on the accuracy of
predicting psychological profiles from people’s digital footprints
(whether in the form of machine learning predictions from a
user’s behavioral history or single target Likes), and therefore,
this approach is not without limitations. First, the psychological
meaning of certain digital footprints might change over time,
making it necessary to continuously calibrate and update the
algorithm to sustain high accuracy. For example, liking the fan-
tasy TV show “Game of Thrones” might have been highly pre-
dictive of introversion when the series was first launched in 2011,
but its growing popularity might have made it less predictive over
time as its audience became more mainstream. As a rule of
thumb, one can say that the higher the face validity of the rela-
tionships between individual digital footprints and specific psy-
chological traits, the less likely it is that they will change (e.g., it is
unlikely that “socializing” will become any less predictive of ex-
traversion over time). Second, while the psychological assess-
ment from digital footprints makes it possible to profile large
groups of people without requiring them to complete a ques-
tionnaire, most algorithms are developed with questionnaires as the
gold standard and therefore retain some of the problems associated
with self-report measures (e.g., social desirability bias; ref. 22).
Additionally, our study has limitations that provide promising
avenues for future research. First, we focused on the two per-
sonality traits of extraversion and openness-to-experience.
Building on existing laboratory studies, future research should
empirically investigate whether and in which contexts other
psychological traits might prove to be more effective [e.g., need
for cognition (2) or regulatory focus (23)]. Second, we conducted
Matz et al.
extreme group comparisons where we targeted people scoring
high or low on a given personality trait using a relatively narrow
and extreme set of Likes. While the additional analyses reported
in SI Appendix suggest that less extreme Likes still enable ac-
curate personality targeting, future research should establish
whether matching effects are linear throughout the scale and, if
not, where the boundaries of effective targeting lie.
The capacity to implement psychological mass persuasion in
the real world carries both opportunities and ethical challenges.
On the one hand, psychological persuasion could be used to help
individuals make better decisions and alleviate many of today’s
societal ills. For example, psychologically tailored health com-
munication is effective in changing behaviors among patients and
groups that are at risk (24, 25). Hence, targeting highly neurotic
individuals who display early signs of depression with materials
that offer them professional advice or guide them to self-help
literature might have a positive preventive impact on the well-
being of vulnerable members of society. On the other hand,
psychological persuasion might be used to exploit “weaknesses”
in a person’s character. It could, for instance, be applied to target
online casino advertisements at individuals who have psycho-
logical traits associated with pathological gambling (26). In fact,
recent media reports suggest that one of the 2016 US presi-
SOCIAL SCIENCES
dential campaigns used psychological profiles of millions of US
citizens to suppress their votes and keep them away from the
ballots on election day (27). The veracity of this news story is
uncertain (28). However, it illustrates clearly how psychological
mass persuasion could be abused to manipulate people to behave
in ways that are neither in their best interest nor in the best in-
terest of society.
Similarly, the psychological targeting procedure described in
this manuscript challenges the extent to which existing and
proposed legislation can protect individual privacy in the digital
age. While previous research shows that having direct access to
an individual’s digital footprint makes it possible to accurately
predict intimate traits (10), the current study demonstrates that
such inferences can be made even without having direct access to
individuals’ data. Although we used indirect group-level target-
ing in a way that was anonymous at the individual level and thus
preserved—rather than invaded—participants’ privacy, the same
approach could also be used to reveal individuals’ intimate traits
without their awareness. For example, a company could advertise
a link to a product or a questionnaire on Facebook, targeting
people who follow a Facebook Like that is highly predictive of
introversion. Simply following such a link reveals the trait to the
advertiser, without the individuals being aware that they have
exposed this information. To date, legislative approaches in the
US and Europe have focused on increasing the transparency of
how information is gathered and ensuring that consumers have a
mechanism to “opt out” of tracking (29). Crucially, none of the
measures currently in place or in discussion address the tech-
niques described in this paper: Our empirical experiments were
performed without collecting any individual-level information
whatsoever on our subjects yet revealed personal information
that many would consider deeply private. Consequently, current
approaches are ill equipped to address the potential abuse of
online information in the context of psychological targeting.
As more behavioral data are collected in real time, it will
become possible to put people’s stable psychological traits in a
situational context. For example, people’s mood and emotions
have been successfully assessed from spoken and written lan-
guage (30), video (31), or wearable devices and smartphone
sensor data (32). Given that people who are in a positive mood
use more heuristic—rather than systematic—information pro-
cessing and report more positive evaluations of people and
products (33), mood could indicate a critical time period for
psychological persuasion. Hence, extrapolating from what one
PNAS | November 28, 2017 | vol. 114 | no. 48 | 12717
does to who one is is likely just the first step in a continuous
development of psychological mass persuasion.
Methods
Ethical approval was granted by the Department of Psychology Ethics
Committee at the University of Cambridge. Given that the group-level tar-
geting approach applied in studies 1–3 is 100% anonymous on the individual
level (all insights provided by Facebook are summary statistics at the level of
target groups), it is impossible to identify, interact with, and obtain consent
from individual participants.
Study 1.
Selection of target likes. The myPersonality dataset was collected via the
myPersonality Facebook app between 2007 and 2012 (20). Mostly free of charge,
the app allowed its users to take real psychometric tests. Among other validated
tests, users could choose between several versions of the IPIP questionnaire, an
established open-source measure of the five factor model of personality (19).
The five factor model has been shown to have excellent psychometric proper-
ties, including high reliability, convergent and discriminant validity, as well as
robust criterion validity when predicting real-life outcomes (18, 19). Users re-
ceived immediate feedback on their responses and were encouraged to grant
the application access to their personal profile and social network data.
Study 1 used a myPersonality subsample that contained 65,536 unique
Facebook Likes alongside the average personality profile of US-based users
connected to those Likes. The extraversion score of the Facebook Like “Lady
Gaga”, for example, was determined by averaging the z-standardized extra-
version scores of all users in the sample who had liked “Lady Gaga”. To max-
imize the reliability of personality profiles and limit the biases introduced by
differences in traits other than extraversion, we further restricted the dataset
described above to Likes followed by at least 400 users and only considered
Likes that were neutral with respect to the remaining four traits (jzj < 0.20σ).
We finally selected those Likes with the highest (ðzE = 0.54σ, n = 8Þ and lowest
aggregate extraversion scores (ðzE = −0.20σ, n = 8Þ that were available in the
Facebook Interest section at the time. SI Appendix, Table S1 displays the Likes
used to target extraverted and introverted audiences alongside their person-
ality scores and sample sizes.
Advert design. Professional graphic designers created five ads aimed at in-
troverts and five ads aimed at extraverts by manipulating the language and
images used in the advert design. The five extraverted adverts were based on
trait descriptions such as “active, assertive, energetic, enthusiastic, outgoing
and talkative,” whereas the introverted adverts reflected trait descriptions
such as “quiet, reserved, shy, silent, and withdrawn” (18). All ads are dis-
played in SI Appendix, Fig. S2. We validated the manipulation of advert
designs by surveying 38 female judges (16 postgraduate students at the
University of Cambridge Psychology Department and 22 students with no
formal training in psychology). Independent t tests confirmed that both
psychologists and laymen perceived extraverted ads to be more extraverted
than the introverted ads—psychologists: t(196) = 24.77, P < 0.001, d = 3.51
[3.07,3.96]; laymen: t(220) = 15.30, P < 0.001, d = 2.05 [1.73,2.38].
Targeting procedure. We created “introverted” and “extraverted” market
segments by entering the selected target Likes displayed in SI Appendix
Table S1 into the “Interest” section of the Facebook advertising platform.
1. Cesario J, Higgins ET, Scholer AA (2008) Regulatory fit and persuasion: Basic principles
and remaining questions. Soc Personal Psychol Compass 2:444–463.
2. Wheeler SC, Petty RE, Bizer GY (2005) Self-schema matching and attitude change: Situ-
ational and dispositional determinants of message elaboration. J Consum Res 31:787–797.
3. Moon Y (2002) Personalization and personality: Some effects of customizing message
style based on consumer personality. J Consum Psychol 12:313–326.
4. Hirsh JB, Kang SK, Bodenhausen GV (2012) Personalized persuasion: Tailoring per-
suasive appeals to recipients’ personality traits. Psychol Sci 23:578–581.
5. Dubois D, Rucker DD, Galinsky AD (2016) Dynamics of communicator and audience
power: The persuasiveness of competence versus warmth. J Consum Res 43:68–85.
6. Podsakoff PM, Organ DW (1986) Self-reports in organizational research: Problems
and prospects. J Manage 12:531–544.
7. Levitt S, List J (2007) What do laboratory experiments tell us about the real world. J Econ
Perspect 21:153–174.
8. Liberali G, Hauser JR, Urban GL, Braun M (2009) Website morphing. Mark Sci 28:202–223.
9. Lazer D, et al. (2009) Life in the network: The coming age of computational social
science. Science 323:721–723.
10. Kosinski M, Stillwell D, Graepel T (2013) Private traits and attributes are predictable
from digital records of human behavior. Proc Natl Acad Sci USA 110:5802–5805.
11. Marcus B, Machilek F, Schütz A (2006) Personality in cyberspace: Personal web sites as
media for personality expressions and impressions. J Pers Soc Psychol 90:1014–1031.
12. Yarkoni T (2010) Personality in 100,000 words: A large-scale analysis of personality
and word use among bloggers. J Res Pers 44:363–373.
12718 | www.pnas.org/cgi/doi/10.1073/pnas.1710966114
This procedure allowed us to limit the advert recipients to users who were
associated with at least one of our target Likes. At the time the study was
conducted, the Facebook advertising platform only allowed marketers to
enter multiple Likes with OR rather than AND statements. For example, an
extraverted target audience could be created with users who like Making
People Laugh OR “Meeting New People” but not with users who like both.
Therefore, the targeting approach pursued in this paper was based on the
minimum amount of information possible: one single Facebook Like per
person. In addition to the target Likes outlined above, the ad sets were
restricted to female UK residents ages 18–40.
Study 2.
Targeting procedure. Similar to study 1, we selected those Likes with the highest
ðzO = 0.59σ, n = 10Þ and lowest aggregate openness scores ðzO = −0.50σ, n = 10Þ
that were available in the Facebook Interest section at the time. In addition to
the targeting specifications outlined in the main manuscript, we restricted our
ad sets to US residents who were connected to a wireless network at the time
of seeing the ads to facilitate app installs. SI Appendix, Table S6 displays the
Likes used to target audiences low and high in openness alongside their per-
sonality scores and sample sizes.
Ad design. Professional graphic designers and copy editors created two adverts
tailored to high and low openness characteristics by manipulating the lan-
guage and images used in the advert design. While the low-openness advert
was based on trait descriptions such as “down to earth, traditional and
conservative,” the high-openness advert reflected trait descriptions such as
“intellectually curious, creative, imaginative, and unconventional” (17). We
validated the manipulation of advert designs by surveying 22 students at the
University of Cambridge (average age = 23.5 y, 50% female). An in-
dependent t test confirmed that participants perceived the high-openness
ad to be more open-minded than the low-openness ad, t(42) = –4.28, P <
0.001, d = 1.29 [0.62, 1.96].
Study 3.
Targeting procedure. Following the company’s standard behavioral targeting
approach, ad sets were aimed at women aged 35 and above, living in the US,
who were connected to at least one of the mobile games on the company’s
behavioral target list. We further restricted our ad sets to US residents who
were connected to a wireless network at the time of seeing the ads, to fa-
cilitate app installs.
Ad design. Professional copy writers produced an introverted (personality-
tailored) ad text that we subsequently compared with the standard ad text
(the image that was used to advertise the app was kept constant). The two ad
versions are displayed in SI Appendix, Table S9. We validated the manipula-
tion of ad designs by surveying 22 students at the University of Cambridge
(average age = 23.5 y, 50% female). An independent t test confirmed that
participants perceived the personality-tailored ad to be more introverted than
the standard ad, t(41) = –2.77, P = 0.008, d = 0.84 [0.20, 1.48].
ACKNOWLEDGMENTS. We thank Vess Popov, Jochen Menges, Jon Jachimowicz,
Gabriella Harari, Sandrine Mueller, Youyou Wu, Maarten Bos, Michael Norton,
Nader Tavassoli, Joseph Sirgy, Moran Cerf, Pinar Yildirim, and Winter Mason for
their critical reading of earlier versions of the manuscript.
13. Golbeck J, Robles C, Edmondson M, Turner K (2011) Predicting personality from
twitter. 2011 IEEE International Conference on Privacy, Security, Risk, and Trust and
IEEE International Conference on Social Computing (IEEE Computer Society, Wash-
ington, DC), pp 149–156.
14. Youyou W, Kosinski M, Stillwell D (2015) Computer-based personality judgments are
more accurate than those made by humans. Proc Natl Acad Sci USA 112:1036–1040.
15. Park G, et al. (2015) Automatic personality assessment through social media lan-
guage. J Pers Soc Psychol 108:934–952.
16. Schwartz HA, et al. (2013) Personality, gender, and age in the language of social
media: The open-vocabulary approach. PLoS One 8:e73791.
17. Segalin C, Perina A, Cristani M, Vinciarelli A (2017) The pictures we like are our image:
Continuous mapping of favorite pictures into self-assessed and attributed personality
traits. IEEE Trans Affect Comput 8:268–285.
18. McCrae RR, John OP (1992) An introduction to the five-factor model and its appli-
cations. J Pers 60:175–215.
19. Goldberg LR, et al. (2006) The international personality item pool and the future of
public-domain personality measures. J Res Pers 40:84–96.
20. Kosinski M, Matz SC, Gosling SD, Popov V, Stillwell D (2015) Facebook as a research
tool for the social sciences: Opportunities, challenges, ethical considerations, and
practical guidelines. Am Psychol 70:543–556.
21. Ai C, Norton EC (2003) Interaction terms in logit and probit models. Econ Lett 80:123–129.
22. Paulhus DL, Reid DB (1991) Enhancement and denial in socially desirable responding.
J Pers Soc Psychol 60:307–317.
Matz et al.
23. Cesario J, Grant H, Higgins ET (2004) Regulatory fit and persuasion: Transfer from
“Feeling Right.” J Pers Soc Psychol 86:388–404.
24. Noar SM, Benac CN, Harris MS (2007) Does tailoring matter? Meta-analytic review of
tailored print health behavior change interventions. Psychol Bull 133:673–693.
25. Mann T, Sherman D, Updegraff J (2004) Dispositional motivations and message framing:
A test of the congruency hypothesis in college students. Health Psychol 23:330–334.
26. Bagby RM, et al. (2007) Pathological gambling and the five-factor model of person-
ality. Pers Individ Differ 43:873–880.
27. Grassegger H, Krogerus M (2017) The data that turned the world upside down.
Available at https://motherboard.vice.com/en_us/article/how-our-likes-helped-trump-win.
Accessed March 4, 2017.
28. Doward J, Gibbs A (2017) Did Cambridge Analytica influence the Brexit vote and the
US election? Guard. Available at https://www.theguardian.com/politics/2017/mar/04/
nigel-oakes-cambridge-analytica-what-role-brexit-trump. Accessed May 10, 2017.
29. EDPS (2015) Meeting the Challenges of Big Data: A Call for Transparency, User
Control, Data Protection by Design and Accountability (European Data Protection
Matz et al.
Supervisor). Available at https://edps.europa.eu/sites/edp/files/publication/15-11-
19_big_data_en.pdf. Accessed May 14, 2017.
30. Cowie R, et al. (2000) “FEELTRACE”: An instrument for recording perceived emotion in real
time. ISCA Tutorial and Research Workshop (ITRW) on Speech and Emotion (International
Speech Communication Association Archive, Newcastle, Northern Ireland). Available at
https://pdfs.semanticscholar.org/7ef4/c7d2bcb658ba279ef784357fc5d4ee970f88.pdf.
Accessed May 14, 2017.
31. Teixeira T, Wedel M, Pieters R (2012) Emotion-induced engagement in internet video
advertisements. J Mark Res 49:144–159.
32. LiKamWa R, Liu Y, Lane ND, Zhong L (2013) Moodscope: Building a mood sensor
from smartphone usage patterns. Proceeding of the 11th Annual International
Conference on Mobile Systems, Applications, and Services (ACM, New York), pp
389–402.
33. Bagozzi RP, Gopinath M, Nyer PU (1999) The role of emotions in marketing. J Acad
Mark Sci 27:184–206.
SOCIAL SCIENCES
PNAS | November 28, 2017 | vol. 114 | no. 48 | 12719
 Supplementary Information Appendix

Overview

A) Methods and Results of Study 1

B) Methods and Results of Study 2

C) Methods and Results of Study 3

D) Evidence for the scalability of our psychological targeting approach

 A) Study 1

Methods

Table S1 displays the full list of target Likes used in Study 1.

Table S1. Extraverted and introverted target Likes alongside the number and average personality
scores of their followers in the myPersonality database.

Extraverted Introverted
Target Like 𝑧̅𝐸 n Target Like 𝑧̅𝐸 n
Parties 0.62 904 Sanctuary -0.25 424
Making People Laugh 0.56 483 Stargate SG1 -0.23 1,278
Meeting New People 0.56 1,106 Dark Angel -0.20 409
Dancing 0.55 8,594 Stargate -0.19 1,099
Shwayze 0.54 844 Serenity -0.18 2,477
Entourage 0.54 1,934 Stargate Atlantis -0.18 1,291
Rebelution 0.52 752 Battlestar Galactica -0.16 2,399
Slightly Stoopid 0.50 2,322 Computers -0.16 1,540
Mean 0.54 10,917 Mean -0.20 18,666

Validation of Targeting Procedure. Using the same targeting procedure as described for Study 1,
we ran a Facebook advertising campaign that invited users to take the 100-item IPIP personality
questionnaire (1) with instant feedback on their scores. It was displayed in the News Feed and the
right-hand advertising column. The ad is shown in Figure S1.

Figure S1. Ad used to recruit participants via Facebook advertising.

From the 263,000 users in the extraverted target group and the 256,000 users in the introverted
target group to which the ad was displayed, 479 and 563 users clicked on it respectively. The full
100-item IPIP questionnaire was completed by 59 introverted and 60 extraverted users (completion
rate of 12% and 11%). For each user, we calculated the extraversion score by averaging the
responses to the 20 extraversion items and z-standardizing the resulting raw scores across the whole
sample of 119 participants. An independent t-test confirmed that participants who were targeted
with extraverted Likes were significantly more extraverted than those targeted with introverted
Likes (t(117) = 4.29, p < .001, d = 0.79 [0.41,1.16]). The results confirm that as little as one
behavioral record is sufficient to target groups of different personality profiles with the help of
existing behavioral targeting tools.

Table S2. Introverted and extraverted ads used in Study 1.

Introversion
Image accessible here:
http://www.gettyimages.co.uk/license/137548001
Text:
Beauty doesn’t have to shout
Image accessible here:
http://www.gettyimages.co.uk/license/488603139
Text:
Uncover your natural beauty
Image accessible here:
http://www.gettyimages.ca/license/78767137
Text:
Indulge your natural beauty
Image accessible here:
http://www.gettyimages.co.uk/license/496839201
Text:
Find some time to feel yourself.
Image accessible here:
http://www.gettyimages.co.uk/license/184934575
Text:
Beauty isn’t always about being on show
Extraversion
Image accessible here:
http://www.gettyimages.ca/license/476806075
Text:
Dance like no one’s watching (but they totally are)
Image accessible here:
http://www.gettyimages.co.uk/license/476996977
Text:
Bring out your best features and bring on the night
Image accessible here:
http://www.gettyimages.co.uk/license/130899617
Text:
Love the spotlight and feel the moment
Image accessible here:
http://www.gettyimages.co.uk/license/152415201
Text:
Bold characters feel unique
Image accessible here:
http://www.gettyimages.co.uk/license/130406597
Text:
Love the spotlight
Table S3. Raw means for each of the ads and Big Five traits as rated by our samples of
psychologists and laymen.
Psychologist Sample (N = 16)
Ad Openness Conscientiousness Extraversion Agreeableness Neuroticism
Int1 -0.20 1.15 -1.05 0.70 -0.35
Int2 0.39 0.00 -1.56 0.61 -0.44
Int3 -0.25 1.05 -1.00 0.55 -0.75
Int4 0.00 1.50 -1.65 0.15 -0.30
Int5 0.15 1.00 -1.50 0.85 -1.35
Int Total 0.02 0.94 -1.37 0.57 -0.64
Ext1 1.05 -0.33 1.86 -0.24 0.62
Ext2 1.71 -1.05 2.62 0.24 0.25
Ext3 1.68 -1.21 2.58 0.84 0.21
Ext4 1.94 -0.78 2.39 0.78 0.22
Ext5 1.38 -0.05 2.38 0.29 -0.43
Ext Total 1.55 -0.68 2.36 0.38 0.17
Laymen Sample (N = 22)
Ad Openness Conscientiousness Extraversion Agreeableness Neuroticism
Int1 0.18 1.86 -0.59 1.05 -1.41
Int2 1.05 -0.05 -0.55 -0.18 -0.5
Int3 -0.36 1.45 -0.77 0.91 -1.41
Int4 0.26 2.26 -1.13 1.00 -1.09
Int5 -0.14 1.64 -0.64 1.00 -1.32
Int Total 0.20 1.43 -0.74 0.76 -1.15
Ext1 0.55 -0.23 1.32 -1.05 0.95
Ext2 1.95 -1.09 2.18 -0.18 0.45
Ext3 0.36 -0.68 2.05 0.14 0.09
Ext4 1.78 -0.74 2.35 0.64 -0.17
Ext5 0.36 0.41 1.59 0.41 0.36
Ext Total 1.00 -0.47 1.90 -0.01 0.34
Results

Table S4. Descriptive statistics across conditions, broken down by age.

Condition Reach Clicks CTR Conv CR CPConv ROI

Introverted Ads
762,197 2,637 0.35% 121 0.016% £7.80 409%
congruent

Age

18-24 426,453 1,396 0.33% 63 0.015% £7.48 378%

25-34 219,651 799 0.36% 40 0.018% £7.37 444%

35-44 116,093 442 0.38% 18 0.016% £9.92 436%

Introverted Ads
791,270 2,426 0.31% 90 0.011% £10.41 300%
incongruent

Age

18-24 425,140 1,293 0.30% 40 0.009% £11.21 327%

25-34 252,657 784 0.31% 38 0.015% £8.65 307%

35-44 113,473 349 0.31% 12 0.011% £13.33 212%

Extraverted Ads
814,308 2,573 0.32% 117 0.014% £8.32 410%
congruent

Age

18-24 440,748 1,372 0.31% 72 0.016% £6.65 472%

25-34 254,684 798 0.31% 31 0.012% £10.27 360%

35-44 118,876 403 0.34% 14 0.012% £12.60 332%

Extraverted Ads
762,218 2,710 0.36% 62 0.008% £15.93 219%
incongruent

Age

18-24 411,739 1,351 0.33% 34 0.008% £13.66 229%

25-34 233,349 890 0.38% 22 0.009% £15.74 219%

35-44 117,130 469 0.40% 6 0.005% £29.45 191%

Note. CTR = Click-through rate, Conv = Conversions, CR = Conversion rate, CPConv = Cost per
conversion, ROI = Return-on-Investment
Table S5. Descriptive statistics across conditions, broken down by target group and individual ad
creatives.

Target CTR CR CPConv ROI

Group Ad Condition Reach Clicks (in %) Conv (in %) (in £) (in %)

Int Int 1 congruent 157,980 605 0.38% 42 0.027% 4.63 652

Int Int 2 congruent 144,545 467 0.32% 18 0.012% 9.94 302

Int Int 3 congruent 143,689 458 0.32% 10 0.007% 18.06 279

Int Int 4 congruent 164,779 543 0.33% 14 0.008% 13.59 230

Int Int 5 congruent 151,204 564 0.37% 37 0.024% 5.40 558

Int Ext1 incongruent 159,951 567 0.35% 11 0.007% 17.94 200

Int Ext 2 incongruent 152,647 616 0.40% 13 0.009% 15.36 236

Int Ext 3 incongruent 144,147 506 0.35% 16 0.011% 12.31 284

Int Ext 4 incongruent 152,563 522 0.34% 10 0.007% 20.00 143

Int Ext 5 incongruent 152,910 499 0.33% 12 0.008% 16.12 232

Ext Int 1 incongruent 157,229 482 0.31% 11 0.007% 17.31 160

Ext Int 2 incongruent 155,993 481 0.31% 31 0.020% 6.16 582

Ext Int 3 incongruent 157,349 460 0.29% 12 0.008% 16.01 154

Ext Int 4 incongruent 160,466 559 0.35% 15 0.009% 13.14 256

Ext Int 5 incongruent 160,233 444 0.28% 21 0.013% 7.93 360

Ext Ext 1 congruent 163,862 546 0.33% 25 0.015% 7.88 438

Ext Ext 2 congruent 160,146 615 0.38% 22 0.014% 9.05 485

Ext Ext 3 congruent 164,003 436 0.27% 24 0.015% 7.62 422

Ext Ext 4 congruent 160,963 479 0.30% 19 0.012% 10.50 276

Ext Ext 5 congruent 165,334 497 0.30% 27 0.016% 7.21 432

Note. Ext = Extraversion, Int = Introversion. CTR = Click-through rate, Conv = Conversions, CR =
Conversion rate, CPConv = Cost per conversion, ROI = Return-on-Investment
Fig. S2. Interaction effects of audience and ad personality on Click-Through Rates (total N = 3,129,993)

Table S6. Results of two hierarchical logistic regression analyses on CTR and CR.

Click-Through Rate (CTR)

Model 1 Model 2
B SE(B) z B SE(B) z
Target Group -0.12*** 0.03 -4.30 -0.12*** 0.03 -4.34
Ads 0.03 0.03 1.00 0.11 0.03 0.35
Target Group × Ads 0.003 0.04 0.07 -0.04 0.04 -1.12
Age
25-34 - - - 0.06* 0.03 -2.02
35-44 - - - 0.09* 0.04 -2.15
Ads × 25-34 - - - 0.02 0.04 0.37
Ads × 35-44 - - - 0.06 0.05 1.02
Conversion Rate (CR)
Model 1 Model 2
B SE(B) t B SE(B) t
Target Group -0.33* 0.14 -2.40 -0.34* 0.14 -2.47
Ads -0.67*** 0.16 -4.28 -0.46* 0.18 -2.54
Target Group × Ads 0.90*** 0.21 4.30 0.91*** 0.21 4.34
Age
25-34 - - - 0.32* 0.15 2.12
35-44 - - - 0.08 0.21 0.36
Ads × 25-34 - - - -0.46* 0.23 -2.05
Ads × 35-44 - - - -0.45 0.32 -1.41
Note. The reference categories for Target Group and Ads are ‘Introverted’. The reference category for Age is 18-24.
 B) Study 2

Methods

Table S7. Target Likes for low and high openness alongside the number and average personality
scores of their followers in the myPersonality database.
High Openness Low Openness
Target Like 𝑧̅𝑂 n Target Like 𝑧̅𝑂 n
Waking Life 0.78 883 Farm Town -0.59 1,094
Film 0.63 628 Bubble Popp -0.57 589
Philosophy 0.62 3,669 Uncle Kracker -0.55 427
Thelonious Monk 0.61 533 NCIS: LA -0.49 1,343
The Fountain 0.59 1,172 My Town -0.49 1,097
Siddhartha 0.57 701 Island Paradise -0.48 1,238
Astrology 0.55 525 MindJolt Games -0.47 2,391
Meditation 0.54 923 Watching TV -0.47 1,719
Poetry 0.53 3,734 My City Life -0.47 597
The Fall 0.51 443 Scary Movie -0.47 438
Mean 0.59 13,211 Mean -0.51 10,933

Targeting Procedure.

We used the same targeting procedure as outlined in Study 1. It is noteworthy that the two matching
ad sets were shown to users much more frequently than their two mismatching counterparts
initially. Assuming that the Facebook advertising algorithm prioritizes high-performing ads, this
finding can be taken as additional support for our matching hypothesis. However, in order to
achieve a more balanced experimental design (with a comparable number of impressions across ad
sets), we made minor adjustments to the bidding procedure throughout the campaign. For each
adjustment, we tested whether the change had a noticeable impact on the key outcomes of click-
through and conversion rates to avoid biases. The fact that none of the adjustments did have a
substantial impact further underlines the robustness of effects.

Table S8. Raw means for each of the ads and Big Five traits, as rated by our online panel

Ad Openness Conscientiousness Extraversion Agreeableness Neuroticism

LO 3.09 5.82 2.45 3.68 2.68
HO 5.00 2.91 5.41 4.14 4.50
Results

Table S9. Descriptive stastics across conditions, broken down by age and gender.

Condition Reach Clicks CTR Conv CR CPConv

LO Ads congruent

Age

18-24 3,779 42 1.11% 27 0.71% $1.59

25-34 5,320 75 1.41% 50 0.94% $1.26

35-44 4,172 64 1.53% 39 0.93% $1.36

45-54 3,027 66 2.18% 30 0.99% $1.57

55-64 1,343 39 2.90% 21 1.56% $1.12

65+ 569 10 1.76% 7 1.23% $1.58

Gender

Female 14,170 224 1.58% 137 0.97% $1.37

Male 4,040 72 1.78% 37 0.92% $1.41

LO Ads incongruent
Age

18-24 9,393 66 0.70% 34 0.36% $2.08

25-34 6,500 69 1.06% 35 0.54% $1.59

35-44 4,190 37 0.88% 22 0.53% $1.72

45-54 3,726 51 1.37% 28 0.75% $1.52

55-64 2,495 39 1.56% 17 0.68% $2.07

65+ 1,459 33 2.26% 13 0.89% $1.58

Gender

Female 21,508 223 1.04% 116 0.54% $1.78

Male 6,255 72 1.15% 33 0.53% $1.69

HO Ads congruent

Age

18-24 7,217 69% 0.96 24 0.33% $2.72

25-34 7,809 105% 1.34 42 0.54% $1.93

35-44 6,595 105% 1.59 34 0.52% $1.93

 45-54 4,377 75% 1.71 25 0.57% $2.30

55-64 2,131 50% 2.35 10 0.47% $3.27

65+ 1,148 23% 2.00 5 0.44% $3.74

Gender

Female 24,440 357% 1.46 117 0.48% $2.29

Male 4,837 70% 1.45 23 0.48% $2.28

HO Ads incongruent
Age

18-24 1,565 14% 0.89 5 0.32% $2.91

25-34 2,505 24% 0.96 12 0.48% $2.15

35-44 2,024 25% 1.24 4 0.20% $5.28

45-54 1,569 26% 1.66 6 0.38% $3.27

55-64 884 16% 1.81 6 0.68% $2.00

65+ 379 7% 1.85 4 1.06% $1.78

Gender

Female 7,178 100% 1.39 35 0.49% $2.31

Male 1,748 12% 0.69 2 0.11% $9.68

Figure S3. Interaction effects of audience and ad personality on Click-Through Rates (total N = 84,176).
Table S10. Results of two hierarchical logistic regression analyses on CTR and CR.

Click-Through Rate (CTR)

Model 1 Model 2
B SE(B) z B SE(B) z
Target Group -0.43*** 0.08 -5.20 -0.42 0.08 -4.96
Ads -0.26* 0.11 -2.35 -0.26 0.18 -1.45
Target Group × Ads 0.58*** 0.14 4.32 0.61*** 0.14 4.47
Age
25-34 - - - 0.32* 0.13 2.49
35-44 - - - 0.29* 0.14 2.06
45-54 - - - 0.68*** 0.14 5.02
55-64 - - - 0.89*** 0.15 5.92
65+ - - - 0.96*** 0.18 5.26
Gender - - - 0.08 0.10 0.79
Ads × 25-34 - - - -0.02 0.19 -0.10
Ads × 35-44 - - - 0.21 0.20 1.03
Ads × 45-54 - - - -0.06 0.20 -0.30
Ads × 55-64 - - - -0.01 0.22 -0.28
Ads × 65+ - - - -0.19 0.28 -0.67
Ads × Gender - - - -0.26 0.16 -1.66
Conversion Rate (CR)
Model 1 Model 2
B SE(B) z B SE(B) z
Target Group -0.58*** 0.11 -5.19 -0.55*** 0.11 -4.87
Ads -0.84*** 0.18 -4.63 -0.82** 0.29 -2.89
Target Group × Ads 0.72*** 0.22 3.35 0.72*** 0.22 3.32
Age
25-34 - - - 0.35* 0.17 2.10
35-44 - - - 0.34 0.18 1.87
45-54 - - - 0.53** 0.19 2.88
55-64 - - - 0.73*** 0.21 3.52
65+ - - - 0.77** 0.26 2.98
Gender - - - -0.08 0.14 -0.56
Ads × 25-34 - - - 0.13 0.29 0.45
Ads × 35-44 - - - -0.03 0.32 -0.09
Ads × 45-54 - - - -0.05 0.32 -0.15
Ads × 55-64 - - - -0.23 0.38 -0.60
Ads × 65+ - - - -0.16 0.46 -0.35
Ads × Gender - - - -0.18 0.25 -0.70
Note. The reference categories for Target Group and Ads are ‘Introverted’. The reference category for Age is 18-24
years.
 C) Study 3

Methods

Table S11. Raw means for each of the ads and Big Five traits, as rated by our online panel.

Ad Openness Conscientiousness Extraversion Agreeableness Neuroticism

Standard 4.54 2.90 5.59 4.05 5.04
Tailored 3.77 3.32 4.38 3.91 4.09

Results

Online Experiment. To rule out the possibility that the personality-tailored ad was simply more
appealing and effective overall, we conducted an experiment on Amazon Mechanical Turk. The
survey was completed by 98 workers. We excluded one worker who had not passed the attention
check, which left us with 97 valid responses (average age = 35.34, 50% female). Participants
indicated their opinions about the standard and the personality-tailored ad copy using a five-point
scale from 1 = Strongly Disagree to 5 = Strongly Agree. We used the following four questions: “I
find this ad to be persuasive”, “Overall, I like this ad”, “I’m interested in learning more about the
app after seeing this ad”, and “This is an effective ad” (adapted from , 2). With an average
Cronbach’s alpha of α = 0.94 across the two ads, the scale reliability of our preference measure
was found to be excellent. Participants subsequently completed the BFI, an established measure of
the five factor Model of personality (3).

We used linear regression to predict participants’ attitudes about the two ads from participants’
dichotomized extraversion level (0 = introverted 1 = extraverted), the ad copy (0 = introverted, 1 =
standard), as well as their two-way interaction. The average ratings across the four conditions are
shown in Figure S5.
Figure S4. Average ratings across the four conditions (total N = 97). The error bars indicate the
standard error of measurement (SEM).

Table S11 displays the results of the linear regression analysis. The analysis revealed a significant
main effect of the ad design, such that the personality-tailored ad was perceived as significantly
more appealing and effective than the standard ad. However, as the significant interaction effect as
well as Figure S6 illustrate, this effect was almost exclusively driven by introverted participants.
While introverted participants showed a significant preference for the introverted ad (t(94) = 3.55,
p < 0.001), there was no significant preference among extraverted participants (t(96) = 0.35, p =
0.727). The findings indicate that the results obtained from the Facebook campaigns and presented
in the main manuscript are unlikely to be explained by the fact that the introverted ad was generally
more appealing than the extraverted ad.

Table S12. Results of a linear regression analysis of participant extraversion, ad copy as well as
their interaction on participants’ preference ratings.
B SE(B) z
Participant Extraversion 0.09 0.20 -3.52
Ad -0.72*** 0.20 0.42
Participant Extraversion × Ad 0.65* 0.29 2.26
The reference ad is ‘Introverted’.
 D) Evidence for the scalability of our psychological targeting approach

We conducted additional analyses on all five personality traits in the five factor Model (4, 5) to
explore the potential scope and scalability of psychological targeting by testing the predictive
accuracy of individual target Likes in a hold-out validation procedure that resembles a real-life
targeting setting.

Method

Dataset and Analysis. We used the latest datasets made available on myPersonality.org in
November 2015 and selected Facebook Likes that were followed by at least 50 U.S. Facebook
users who had completed the full 100-item IPIP questionnaire (1). This left us with 32,712 unique
Likes from 73,085 users (12,507,096 unique user-Like pairs). In order to obtain two independent
samples, we randomly split the dataset into two subsamples (subsample A = 36,570 consumers,
subsample B = 36,515 consumers). Each of the datasets was subsequently used as both a training
and testing dataset (Fold 1: A = training, B = testing; Fold 2: B = training, A = testing). For each
of the two folds, the validation procedure developed in three steps (see Figure S6 for a visual
illustration of Fold 1). First, we used the training sample to calculate the average personality
profiles of all Likes by averaging the z-standardized scores on each of the five personality traits
across all participants in our sample who had endorsed the particular Like. Similar to Study 1, the
personality profile of the Facebook Like “Lady Gaga” therefore describes the average personality
profile of consumers in our sample who followed her Page. Second, we looked at the individual
user-Like pairs in our testing dataset and compared the predicted personality of the user (based on
the Like personality) with their actual self-reported personality. All user-Like pairs for which the
predicted personality matched the actual personality were classified as “correctly targeted.” Third,
for each Like and personality trait, we calculated the percentage of participants classified as
“correctly targeted” as the fold-specific classification accuracy. The following example describes
the procedure for the trait of openness and a single Facebook Like. Let us assume, for example,
that we calculated an average openness score of 𝑧̅𝑂 = .50𝜎 for the Facebook Like “Lady Gaga” in
the training sample. Given that the openness score was above average (𝑧̅𝑂 = .50𝜎 > 0), the
Facebook Like “Lady Gaga” can be considered as a target Like for high openness (Step 1). Next,
we compared the actual personality scores of all the users who have liked “Lady Gaga” in our
testing sample to the prediction made by the target Like. In the case of the Facebook Like “Lady
Gaga” (high openness), all participants whose z-scores were above the average openness score (𝑧𝑂
> 0) were classified as “correctly targeted,” while all participants whose z-scores were below the
average openness score (𝑧𝑂 < 0) were classified as “incorrectly targeted” (Step 2). Finally, we
calculated the classification accuracy of the Like “Lady Gaga” as the proportion of participants
classified as “correctly targeted” relative to the total number of participants who were associated
with this Like (Step 3).

Figure S5. Hold-out validation procedure (Fold 1) used to determine the classification accuracy of
all 32,712 Facebook Likes.

One of the defining features of Facebook Likes in the context of classification accuracy is their
level of discriminative power, which is equal to their absolute average z-score. For example, one
would expect a Facebook Like with an average extroversion z-score of 𝑧̅𝐸 = .50𝜎 to be more
effective in defining an extroverted target group than a Facebook Like with an average extroversion
z-score of 𝑧̅𝐸 = 0.20𝜎. Similarly, one would expect a Facebook Like with an average extroversion
z-score of 𝑧̅𝐸 = −.50𝜎 to be more effective in defining an introverted target group than a Facebook
Like with an average extroversion z-score of 𝑧̅𝐸 = −0.20𝜎. Therefore, taking the absolute value
of the personality z-scores associated with a Facebook Like provides an indication of its level of
discriminative power: The higher the absolute z-score, the better the Like should be at
discriminating between consumers that are high and low on the trait. In order to investigate the
classification accuracy of Facebook Likes at different levels of discriminative power, we binned
all of the 32,712 Facebook Likes in our sample according to their average absolute trait score,
rounded to the closest decimal point. The bin |𝑧̅𝑜 | = 0.8𝜎, for example, contained all Likes with
an average openness score of 0.75 ≥ |𝑧̅𝑜 | > 0.85𝜎, as well as all Likes with an average openness
score of −0.85 ≥ |𝑧̅𝑜 | > −0.75𝜎.

Results

Figure S7 displays the classification accuracy of Facebook Likes (Y-axis) dependent on their level
of discriminative power (X-axis; see web appendix G for exact values). A classification accuracy
of 50% is the level of chance. As expected, the classification accuracy of Facebook Likes increases
significantly with their level of discriminative power. While the average classification accuracy
across all traits is only 58% for Likes with a discriminative power of .02, it goes up to 69% for
Likes with a discriminative power of 0.8. Importantly, however, there are considerable differences
in the classification accuracy between personality traits. While openness was found to be the most
successful in targeting participants outside of the training sample (maximum classification
accuracy of 82%), agreeableness was found to be the least successful (maximum classification
accuracy of 61%). In line with research on the predictability of personality traits from whole user
profiles (7, 10), this finding suggests that some traits can be predicted more successfully from
digital footprints than others.
Figure S6. Relationship between the cross-validated classification accuracy (Y-axis) and the
levels of target Likes’ discriminative power (total N = 73,085). Facebook Likes were binned
by their average absolute trait score, rounded to the closest decimal point.

Taken together, these findings suggest that it is feasible to implement personality targeting on a
scale that is much larger than what has been empirically demonstrated in Studies 1–3 (e.g., using
different personality traits and Likes of varying discriminant power). However, it is important to
take a closer look at the potential limits in scalability. Indeed, while an increase in the level of
discriminative power results in a higher targeting accuracy of Likes across all traits, it limits the
number of Likes that can be used for targeting. In order to examine this inverse relationship, we
calculated the number of Facebook Likes at every level of discriminative power. Figure S8
illustrates the trade-off between accuracy and reach by plotting the average level of accuracy (Y-
axis) against the relative number of Likes available for targeting (X-axis). For example, while on
average 26% of Facebook Likes have a discriminative power of |𝑧̅| = .30, only 2.5% of Likes
have a discriminative power of |𝑧̅| = .60.

Although a high level of discriminative power thus allows marketers to increase their level of
accuracy when targeting consumers of different personalities, it simultaneously reduces the pool
of consumers that can be targeted. The optimum balance between accuracy and reach likely
depends on a marketing campaign’s specific requirements. If it is essential to reach a large number
of consumers, marketers might need to lower their level of accuracy. However, if the priority is to
maximize targeting accuracy, the potential reach of campaigns might be restricted to a smaller
number of consumers. Given the enormous size of Facebook, however, a relatively small number
of Likes can often be sufficient to target a fairly large number of consumers. Even though Study 1
only used eight target Likes, for example, the estimated audience size was N = 6,000,000 in each
of the groups.

Figure S7. Trade-off between the accuracy and availability of target Likes at different levels of
discriminant power.

References
1. Goldberg LR, et al. (2006) The international personality item pool and the future of public-
domain personality measures. J Res Pers 40(1):84–96.
2. Hirsh JB, Kang SK, Bodenhausen G V (2012) Personalized persuasion: tailoring
persuasive appeals to recipients’ personality traits. Psychol Sci 23(6):578–81.
3. John OP, Srivastava S (1999) The Big Five trait taxonomy: History, measurement, and
theoretical perspectives. Handb Personal Theory Res 2(1999):102–138.
4. McCrae RR, John OP (1992) An introduction to the five-factor model and its applications.
J Pers 60(2):175–215.
5. Goldberg LR (1999) A broad-bandwidth, public domain, personality inventory measuring
the lower-level facets of several five-factor models. Personal Psychol Eur 7:7–28.