What Twitter Knows: Characterizing Ad Targeting Practices ... · 978-1-939133-17-5 Open access to the Proceedings of the 29th USENIX Security Symposium is sponsored by USENIX. What

This paper is included in the Proceedings of the 29th USENIX Security Symposium

August 12ndash14 2020978-1-939133-17-5

Open access to the Proceedings of the 29th USENIX Security Symposium

is sponsored by USENIX

What Twitter Knows Characterizing Ad Targeting Practices User Perceptions and Ad Explanations

Through Usersrsquo Own Twitter DataMiranda Wei University of Washington University of Chicago Madison Stamos and Sophie Veys University of Chicago Nathan Reitinger and Justin Goodman University of Maryland Margot Herman University of Chicago Dorota Filipczuk

University of Southampton Ben Weinshel University of Chicago Michelle L Mazurek University of Maryland Blase Ur University of Chicago

httpswwwusenixorgconferenceusenixsecurity20presentationwei

What Twitter Knows Characterizing Ad Targeting Practices User Perceptionsand Ad Explanations Through Usersrsquo Own Twitter Data

Miranda Wei4 Madison Stamos Sophie Veys Nathan Reitinger Justin GoodmanMargot Herman Dorota Filipczukdagger Ben Weinshel Michelle L Mazurek Blase Ur

University of Chicago University of Maryland daggerUniversity of Southampton4University of Chicago and University of Washington

AbstractAlthough targeted advertising has drawn significant attentionfrom privacy researchers many critical empirical questionsremain In particular only a few of the dozens of targetingmechanisms used by major advertising platforms are wellunderstood and studies examining usersrsquo perceptions of adtargeting often rely on hypothetical situations Further it isunclear how well existing transparency mechanisms fromdata-access rights to ad explanations actually serve the usersthey are intended for To develop a deeper understanding ofthe current targeting advertising ecosystem this paper en-gages 231 participantsrsquo own Twitter data containing ads theywere shown and the associated targeting criteria for measure-ment and user study We find many targeting mechanismsignored by prior work mdash including advertiser-uploaded listsof specific users lookalike audiences and retargeting cam-paigns mdash are widely used on Twitter Crucially participantsfound these understudied practices among the most privacyinvasive Participants also found ad explanations designedfor this study more useful more comprehensible and overallmore preferable than Twitterrsquos current ad explanations Ourfindings underscore the benefits of data access characterizeunstudied facets of targeted advertising and identify potentialdirections for improving transparency in targeted advertising

1 Introduction

Social media companies derive a significant fraction of theirrevenue from advertising This advertising is typically highlytargeted drawing on data the company has collected aboutthe user either directly or indirectly Prior work suggests thatwhile users may find well-targeted ads useful they also findthem ldquocreepyrdquo [40 42 58 61 70] Further users sometimesfind targeted ads potentially embarrassing [3] and they may(justifiably) fear discrimination [4 15 21 47 57 59 60 73]In addition there are questions about the accuracy of cate-gorizations assigned to users [7 12 21 71 72] Above allusers currently have a limited understanding of the scope andmechanics of targeted advertising [17 22 48 50 54 70 77]

Many researchers have studied targeted advertising largelyfocusing on coarse demographic or interest-based target-ing However advertising platforms like Twitter [63] andGoogle [27] offer dozens of targeting mechanisms that arefar more precise and leverage data provided by users (egTwitter accounts followed) data inferred by the platform (egpotential future purchases) and data provided by advertisers(eg PII-indexed lists of current customers) Further becausethe detailed contents and provenance of information in usersrsquoadvertising profiles are rarely available prior work focusesheavily on abstract opinions about hypothetical scenarios

We leverage data subjectsrsquo right of access to data collectedabout them (recently strengthened by laws like GDPR andCCPA) to take a more comprehensive and ecologically validlook at targeted advertising Upon request Twitter will pro-vide a user with highly granular data about their accountincluding all ads displayed to the user in the last 90 daysalongside the criteria advertisers used to target those ads allinterests associated with that account and all advertisers whotargeted ads to that account

In this work we ask What are the discrete targeting mech-anisms offered to advertisers on Twitter and how are theyused to target Twitter users What do Twitter users think ofthese practices and existing transparency mechanisms A to-tal of 231 Twitter users downloaded their advertising-relateddata from Twitter shared it with us and completed an on-line user study incorporating this data Through this methodwe analyzed Twitterrsquos targeting ecosystem measured partici-pantsrsquo reactions to different types of ad targeting and ran asurvey-based experiment on potential ad explanations

We make three main contributions First we used our231 participantsrsquo files to characterize the current Twitter ad-targeting ecosystem Participants received ads targeted basedon 30 different targeting types or classes of attributes throughwhich advertisers can select an adrsquos recipients These typesranged from those commonly discussed in the literature (eginterests age gender) to others that have received far lessattention (eg audience lookalikes advertiser-uploaded listsof specific users and retargeting campaigns) Some partic-

USENIX Association 29th USENIX Security Symposium 145

ipantsrsquo files contained over 4000 distinct keywords 1000follower lookalikes and 200 behaviors Participantsrsquo files alsorevealed they had been targeted ads in ways that might beseen as violating Twitterrsquos policies restricting use of sensi-tive attributes Participants were targeted using advertiser-provided lists of users with advertiser-provided names con-taining ldquoDLX_Nissan_AfricanAmericansrdquo ldquoChristian Audi-ence to Excluderdquo ldquoRising Hispanics | Email Openersrdquo andmore They were targeted using keywords like ldquotransgenderrdquoand ldquomexican americanrdquo as well as conversation topics likethe names of UK political parties These findings underscorehow data access rights facilitate transparency about targetingas well as the value of such transparency

Second we investigated participantsrsquo perceptions of thefairness accuracy and desirability of 16 commonly observedtargeting types Different from past work using hypotheticalsituations we asked participants about specific examples thathad actually been used to target ads to them in the past 90days Whereas much of the literature highlights usersrsquo nega-tive perceptions of interest-based targeting [42 61] we foundthat over two-thirds of participants agreed targeting based oninterest was fair the third most of the 16 types In contrastfewer than half of participants agreed that it was fair to targetusing understudied types like follower lookalike targetingtailored audience lists events and behaviors Many target-ing types ignored by prior work were the ones viewed leastfavorably by participants emphasizing the importance of ex-panding the literaturersquos treatment of ad-targeting mechanisms

Third we probe a fuller design space of specificity read-ability and comprehensiveness for ad explanations Althoughad explanations are often touted as a key part of privacy trans-parency [24] we find that existing ad explanations are incom-plete and participants desire greater detail about how ads weretargeted to them Compared to Twitterrsquos current explanationparticipants rated explanations we created to be significantlymore useful helpful in understanding targeting and similarto what they wanted in future explanations

Our approach provides a far richer understanding of theTwitter ad ecosystem usersrsquo perceptions of ad targeting andad explanation design than was previously available Our re-sults emphasize the benefits of advertising transparency insurfacing potential harms associated with increasingly accu-rate and complex inferences Our findings also underscorethe need for a more ethical approach to ad targeting that canmaintain the trust of users whose data is collected and used

2 Related Work

We review prior work on techniques for targeted advertisingassociated transparency mechanisms and user perceptions

21 Targeted Advertising Techniques

Web tracking dates back to 1996 [38] The online ad ecosys-tem has only become more sophisticated and complex sinceCompanies like Google Facebook Bluekai and many otherstrack usersrsquo browsing activity across the Internet creatingprofiles for the purpose of sending users targeted advertis-ing Commercial web pages contain an increasing number oftrackers [52] and much more data is being aggregated aboutusers [13] Many studies have examined tools to block track-ing and targeted ads finding that tracking companies can stillobserve some of a userrsquos online activities [2 10 11 19 30]

Social media platforms have rich data for developing exten-sive user profiles [7 12 57] augmenting website visits withuser-provided personal information and interactions with plat-form content [7] This data has included sensitive categorieslike lsquoethnic affinityrsquo [8] and wealth Even seemingly neutralattributes can be used to target marginalized groups [57]

To date studies about user perceptions of ad-targetingmechanisms have primarily focused on profiles of usersrsquodemographics and inferred interests (eg yoga travel) re-gardless of whether the studies were conducted using usersrsquoown ad-interest profiles [12 20 50] or hypothetical scenar-ios [17 36] Furthermore most studies about advertising onsocial media have focused on Facebook [7 25 57 71] Whilesome recent papers have begun to examine a few of the dozensof other targeting mechanisms available [7 72] our studyleverages data access requests to characterize the broad set oftargeting types in the Twitter ecosystem much more compre-hensively than prior work in terms of both the mechanismsconsidered and the depth of a given userrsquos data examined

Newer techniques for targeting ads go beyond collectinguser data in several ways that may be less familiar to bothusers and researchers For example since 2013 Facebook [23]and Twitter [9] have offered ldquocustomrdquo or ldquotailoredrdquo audiencetargeting which combine online user data with offline dataAdvertisers upload usersrsquo personally identifiable information(PII) such as their phone numbers and email addresses gath-ered from previous transactions or interactions in order to linkto usersrsquo Facebook profiles This offline data can also includedata supplied by data brokers [72] often pitched to advertis-ers as ldquopartner audiencesrdquo [32] or even PII from voter andcriminal records [7] These features can be exploited by adver-tisers to target ads to a single person [25] or evade restrictionsabout showing ads to people in sensitive groups [57]

Another newer form of targeting is lookalike-audience tar-geting which relies on inferences about users relative to otherusers For example on Facebook advertisers can reach newusers with similar profiles as their existing audience [39] Thisfeature can be exploited as a biased input group will lead toan output group that contains similar biases [57] Services areincreasingly implementing lookalike targeting [56] To ourknowledge we are the first to study user perceptions of theselesser-known forms of targeting with real-world data

146 29th USENIX Security Symposium USENIX Association

22 Transparency Mechanisms

Ad and analytics companies increasingly offer transparencytools [16 29] These include ad preference managers [12]which allow users to see the interest profiles that platformshave created for them and ad explanations or descriptionsof why a particular advertiser displayed a particular ad toa user [7] Nevertheless a disparity remains between infor-mation available to advertisers and information visible tousers [50 58] Although researchers have documented adver-tisersrsquo use of a multitude of attributes including sensitive onesthey rarely appear in user-facing content [7 15 50 74 75]Facebookrsquos ad preferences are vague and incomplete [7] no-tably leaving out information from data brokers [72]

To shed light on the black box of advertising researchershave developed ldquoreverse engineeringrdquo tools that can extractsome information about targeting mechanisms associated ex-planations and inferences that have been made Techniquesinclude measuring the ads users see [67101115343575]purchasing ads in controlled experiments [47172] and scrap-ing companiesrsquo ad-creation interface [25 57 71 72 74] ad-interest profiles [71215166075] and ad explanations [67]Unfortunately these excellent tools are limited by the diffi-culty of scaling them (as they require making many requestsper user) and by companies continually making changes totheir interfaces perhaps in part to thwart such tools [43]

23 Perceptions of Targeting amp Transparency

Users do not understand advertising data collection and target-ing processes [717204554] They instead rely on imprecisemental models [58] or folk models [2277] While some userslike more relevant content [40] and understand that ads sup-port free content on the web [42] many others believe track-ing browser activity is invasive [42 53] Users are concernedabout discrimination [47] or bias [21] inaccurate inferencesand companies inferring sensitive attributes such as health orfinancial status [50 70] Studies have shown that when userslearn about mechanisms of targeted advertising their feelingstowards personalization become more negative [53 58 61]

To an increasing extent studies have looked into the designand wording of transparency tools [5 37 74] Unfortunatelythese tools are meant to provide clarity but can be confus-ing due to misleading icons [36] or overly complicated lan-guage [37 54] Improving the design of transparency toolsis important because vague ad explanations decrease usersrsquotrust in personalized advertising while transparency increasesparticipantsrsquo likelihood to use that service [20] and to appreci-ate personalization [54 70] Users want to know the specificreasons for why they saw an ad [17] and want more controlover their information by being able to edit their interest pro-files [31 41] Users continually express concern about theirprivacy [18 28] but cannot make informed decisions if infor-mation about how their data is used is not transparent [58]

Ad explanations are a particularly widespread form of trans-parency [7 17] Sadly prior work has found current explana-tions incomplete [77172] and companion ad-interest profilesto be both incomplete [15] and inaccurate [1216] While stud-ies have examined existing ad explanations [7 20 71 72] orengaged in speculative design of new explanations [20] sur-prisingly little work has sought to quantitatively test improvedexplanations We build on this work by quantitatively compar-ing social media platformsrsquo current ad explanations with newexplanations we designed based on prior user research [1720]Emphasizing ecological validity we test these explanationsusing ads that had actually been shown to participants whileexplaining the true reasons those ads had been targeted tothem leveraging the participantrsquos own Twitter data

3 Method

To examine Twitter ad targeting data we designed an on-line survey-based study with two parts First participantsfollowed our instructions to request their data from TwitterUpon receipt of this data a few days later they uploaded theadvertising-relevant subset of this data and completed a surveythat instantly incorporated this data across two sections

Section 1 of the survey elicited participantsrsquo reactions to dif-ferent targeting types such as follower lookalike targeting andinterest targeting We selected 16 commonly observed target-ing types many of which have not previously been exploredin the literature In Section 2 we conducted a within-subjectsexperiment measuring participantsrsquo reactions to six poten-tial ad explanations including three novel explanations wecreated by building on prior work [17 20] as well as approx-imations of Twitter and Facebookrsquos current ad explanationsWe also asked participants about their general Twitter usageWe concluded with demographic questions Our survey wasiteratively developed through cognitive interviews with peo-ple familiar with privacy research as well as pilot testing withpeople who were not Below we detail our method

31 Study Recruitment

We recruited 447 participants from Prolific to request theirTwitter data paying $086 for this step The median comple-tion time was 73 minutes We required participants be at least18 years old live in the US or UK and have a 95+ approvalrating on Prolific Additionally participants had to use Twitterat least monthly and be willing to upload their Twitter ad datato our servers During this step we requested they paste intoour interface the ad interest categories Twitter reported forthem in their settings page If a participant reported 10 orfewer interests (another indicator of infrequent usage) we didnot invite them to the survey

To give participants time to receive their data from Twitterwe waited several days before inviting them back A total of


254 participants completed the survey The median comple-tion time for the 231 valid participants (see Section 41) was315 minutes and compensation was $700

To protect participantsrsquo privacy we automatically ex-tracted and uploaded only the three Twitter files related toadvertising ad-impressionsjs personalizationjsand twitter_advertiser_listpdf The JavaScript filead-impressionsjs contained data associated with adsseen on Twitter in the preceding 90 days includingthe advertiserrsquos name and Twitter handle targeting typesand values and a timestamp An example of this JSONdata is presented in our online Appendix A [1] Thefile twitter_advertiser_listpdf contained advertiserswho included the participant in a tailored audience list as wellas lookalike audiences in which Twitter placed the participant

32 Survey Section 1 Targeting Types

Our goal for the first section of the survey was to compara-tively evaluate user awareness perceptions and reactions tothe targeting types advertisers frequently use to target ads onTwitter We wanted to include as many targeting types as pos-sible while ensuring that a given participant would be likelyto have seen at least one ad targeted using that type If we hadincluded all 30 types we would have only been able to show afew participants an ad relying on the more obscure types andwould likely not have had a sufficient number of participantsto meaningfully carry out our statistical analyses In our pilotdata only 16 targeting types appeared in the data of morethan half of our pilot participants therefore we opted to usethese 16 in the survey The 16 targeting types were as fol-lows follower lookalike location tailored audience (list)keyword age conversation topic interest tailored audi-ence (web) platform language behavior gender moviesand TV shows event retargeting campaign engager andmobile audience We refer to a specific attribute of a type asan instance of that type For example language targeting hasinstances like English and French and event targeting hasinstances including ldquo2019 Womenrsquos World Cuprdquo and ldquoBackto School 2019rdquo These targeting types are described in detailin Section 43 Twitterrsquos definitions are given in our onlineAppendix B [1] Using a mixed between- and within-subjectsdesign we showed each participant four randomly selectedtargeting types chosen from however many of the 16 typeswere in that userrsquos ad impressions file Prior work has coveredonly a fraction of these 16 targeting types Furthermore ask-ing questions about instances from participantsrsquo own Twitterdata increased the ecological validity of our study comparedto the hypothetical scenarios used in prior work

For each targeting type we repeated a battery of questionsFirst we asked participants to define the targeting type in theirown words Next we gave a definition of the term adaptedfrom Twitter for Business help pages [63] We then showedparticipants one specific instance of the targeting type drawn

Figure 1 Example ad shown in Section 2 of the survey Partici-pants always saw the ad before the corresponding explanation

from their Twitter data (eg for keyword ldquoAccording to yourTwitter data you have searched for or Tweeted about catsrdquo)Finally we showed participants the five most and five leastfrequent instances of that targeting type in their Twitter data (ifthere were fewer than 10 instances we showed all available)as well as an estimate of how many ads they had seen in thelast three months that used that targeting type

At this point the participant had seen a definition of thetargeting type as well as several examples to aid their under-standing We then asked questions about participantsrsquo com-fort with perception of the fairness of perceptions of theaccuracy of and desire to be targeted by the type For thesequestions we asked participants to rate their agreement on a5-point Likert scale from strongly agree to strongly disagreeHereafter we say participants ldquoagreedrdquo with a statement asshorthand indicating participants who chose either ldquoagreerdquo orldquostrongly agreerdquo Similarly we use ldquodisagreed as shorthandfor choosing ldquodisagree or ldquostrongly disagree We also askedparticipants to explain their choices in free-text responses toconfirm that participants were understanding our constructs asintended The text of all questions is shown in Appendix E [1]

33 Survey Section 2 Ad Explanations

Our goal for the second section of the survey was to charac-terize user reactions to the ad explanations companies likeTwitter and Facebook currently provide on social media plat-forms as well as to explore whether ideas proposed in pastwork (but not quantitatively tested on a large scale) could leadto improved explanations To that end we used participantsrsquoTwitter data to craft personalized ad explanations for ads thatwere actually displayed to them on Twitter within the last 90days We tested six different ad explanations

Rather than trying to pinpoint the best design or contentthrough extensive AB testing we instead constructed ourexplanations as initial design probes of prospective ad expla-nations that are more detailed than those currently used bymajor social media platforms The explanations differed inseveral ways allowing us to explore the design space Our


within-subjects design invited comparison among the explana-tions which helped participants to evaluate the them as wellas answer the final question of this section ldquoPlease describeyour ideal explanation for ads on Twitter

To study reactions to widely deployed ad explanations ourfirst two explanations were modeled on those Twitter andFacebook currently use They retained the same informationand wording but were recreated in our visual theme for con-sistency and to avoid bias from participants knowing theirorigin The first was based on Twitterrsquos current ad explana-tion (Fig 2a) which features most commonly but not alwaystwo of the many possible ad targeting types interest andlocation (most frequently at the level of country) Notablyads themselves can be targeted to more granular locations andusing many more targeting types Twitterrsquos current explana-tion does not present these facets to users We also adaptedone of Facebookrsquos current ad explanations (Fig 2b) whichuses a timeline to explain tailored audience targeting andincorporates age and location These explanations representtwo major platformsrsquo current practices

Because current ad explanations are vague and incom-plete [7 72] we wanted to explore user reactions to potentialad explanations that are more comprehensive and also inte-grate design suggestions from prior work [17 20] We thuscreated two novel explanations Detailed Visual (Fig 2c) andDetailed Text (Fig 2d) that showed a more comprehensiveview of all the targeting types used including lesser-knownyet commonly used targeting types like follower lookalikemobile audience event and tailored audience The distinctionbetween our two conditions let us explore the communicationmedium While we hypothesized that Detailed Visual wouldperform better than Detailed Text we wanted to probe thetrade-off between the comprehensiveness and comprehensi-bility of text-based explanations

While ad explanations should be informative and intelligi-ble they should also nudge users to think about their choicesregarding personalized advertising We designed our thirdnovel ad explanation ldquoCreepy (Fig 2e) to more stronglynudge participants toward privacy by including informationlikely to elicit privacy concerns This explanation augmentedour broader list of targeting types with information the partic-ipant leaks to advertisers such as their device browser andIP address This explanation also used stronger declarativelanguage such as ldquoyou are instead of ldquoyou may

Finally we designed a generic Control explanation(Fig 2f) that provided no targeting information This expla-nation was designed to be vague and meaningless Followingother work [29 76] Control provides a point of comparison

Our ad explanations are the result of several iterations ofdesign After each iteration we discussed whether the de-signs met our goal of creating a spectrum of possibilitiesfor specificity readability and comprehensiveness We thenredesigned the explanations until we felt that they were satis-factory based on both pilot testing and group discussion

Participants were shown ad explanations in randomized or-der Each explanation was preceded by an ad from their dataand customized with that adrsquos targeting criteria We createda list of all ads a participant had been shown in the last 90days and sorted this list in descending order of the number oftargeting types used To filter for highly targeted ads we se-lected six ads from the beginning of this list Participants whohad fewer than six ads in their Twitter data saw explanationsfor all of them After each explanation we asked questionsabout whether the ad explanation was useful increased theirtrust in advertisers and more

The six ad explanations collectively represent a spectrumof possible ad explanations in terms of specificity Controlrepresents a lower bound Creepy represents an upper boundand the others fall in between

34 Analysis Method and Metrics

We performed quantitative and qualitative analyses of surveydata We provide descriptive statistics about Twitter data files

Because each participant saw only up to four of the 16 tar-geting types in survey Section 1 we compared targeting typesusing mixed-effects logistic regression models These areappropriate for sparse within-subjects ordinal data [49 62]Each model had one Likert question as the outcome and thetargeting type and participant (random effect) as input vari-ables We used interest targeting as our baseline because itis the most widely studied targeting type Interest targetingis also commonly mentioned in companiesrsquo advertising dis-closures and explanations in contrast to most other targetingtypes we investigated (eg tailored audience) Appendix I [1]contains our complete regression results

To investigate how targeting accuracy impacted partici-pant perceptions we also compared the accuracy of targetingtype instances (self-reported by participants) to participantsrsquoresponses to the other questions for that targeting type Toexamine correlation between these pairs of Likert responseswe used Spearmanrsquos ρ which is appropriate for ordinal data

To compare a participantrsquos Likert responses to the six dif-ferent ad explanations they saw we used Friedmanrsquos ranksum test (appropriate for ordinal within-subjects data) as anomnibus test We then used Wilcoxon signed-rank tests tocompare the other five explanations to the Twitter explana-tion which we chose as our baseline because Twitter currentlyuses it to explain ads We used the Holm method to correctp-values within each family of tests for multiple testing

We qualitatively analyzed participantsrsquo free-response an-swers to five questions about targeting types and ad explana-tions through an open coding procedure for thematic analysisOne researcher made a codebook for each free-response ques-tion and coded participant responses A second coder inde-pendently coded those responses using the codebook made bythe first The pair of coders for each question then met to dis-cuss the codebook verifying understandings of the codes and


(a) Twitter ad explanation

(b) Facebook ad explanation

(c) Detailed Visual ad explanation

(d) Detailed Text ad explanation

(e) Creepy ad explanation

(f) Control ad explanationFigure 2 The six ad explanations tested using a hypotheticalad to demonstrate all facets of the explanations

combining codes that were semantically similar Inter-coderreliability measured with Cohenrsquos κ ranged from 053 to 091

for these questions Agreement gt 04 is considered ldquomod-eraterdquo and gt 08 ldquoalmost perfectrdquo [33] To provide contextwe report the fraction of participants that mentioned specificthemes in these responses However a participant failing tomention something is not the same as disagreeing with it sothis prevalence data should not be considered generalizableAccordingly we do not apply hypothesis testing

35 EthicsThis study was approved by our institutionsrsquo IRB As socialmedia data has potential for abuse we implemented manymeasures to protect our participantsrsquo privacy We did not col-lect any personally identifiable information from participantsand only identified them using their Prolific ID numbers Ad-ditionally we only allowed participants to upload the threefiles necessary for the study from participantsrsquo Twitter dataall other data remained on the participantrsquos computer Thesethree files did not contain personally identifiable informationIn this paper we have redacted potential identifiers found intargeting data by replacing numbers with letters with anddates with MM DD or YYYY as appropriate

To avoid surprising participants who might be uncomfort-able uploading social media data we placed a notice in ourstudyrsquos recruitment text explaining that we would requestsuch data As some of our participants were from the UK andProlific is located in the UK we complied with GDPR

36 LimitationsLike all user studies ours should be interpreted in the contextof its limitations We used a convenience sample via Prolificthat is not necessarily representative of the population whichlessens the generalizability of our results However prior worksuggests that crowdsourcing for security and privacy surveyresults can be more representative of the US population thancensus-representative panels [51] and Prolific participantsproduce higher quality data than comparable platforms [46]We may have experienced self-selection bias in that potentialparticipants who are more privacy sensitive may have been un-willing to upload their Twitter data to our server Nonethelesswe believe our participants provided a useful window intouser reactions While we did find that the average charactercount of free response questions decreased over the courseof the survey (ρ = minus0399 p lt 001 between question orderand average character number) we were satisfied with thequalitative quality of our responses Responses included inour analysis and results were on-topic and complete

We were also limited by uncertainty in our interpretation ofthe Twitter data files at the time we ran the user study Twittergives users their data files without documentation definingthe elements in these files For instance each ad in the datafile contains a JSON field labeled ldquomatchedTargetingCriteriardquothat contains a list of targeting types and instances It was


initially ambiguous to us whether all instances listed had beenmatched to the participant or whether this instead was a fulllist of targeting criteria specified by the advertiser regardlessof whether each matched to the participant The name of thisfield suggested the former interpretation However the pres-ence of multiple instances that could be perceived as mutuallyexclusive (eg non-overlapping income brackets) and Twit-ter telling advertisers that some targeting types are ldquoORedrdquowith each other (see online Appendix F Figure 6 [1]) madeus question our assumption Members of the research teamdownloaded their own data and noticed that most ldquomatched-TargetingCriteriardquo were consistent with their own character-istics We made multiple requests for explanations of thisdata from Twitter including via a GDPR request from anauthor who is an EU citizen (see online Appendix C [1])We did not receive a meaningful response from Twitter formore than 45 months by which point we had already run theuser study with softer language in survey questions and adexplanations than we might otherwise have used UltimatelyTwitterrsquos final response reported that the instances shown un-der ldquomatchedTargetingCriteriardquo indeed were all matched tothe user confirming our initial interpretation

Because we wanted to elicit reactions to ad explanations forads participants had actually been shown our comparisons ofad explanations are limited by peculiarities in participantsrsquo adimpressions data If an ad did not have a particular targetingtype associated with it then that targeting type was omit-ted from the explanation The exception was Visual whichtold participants whether or not each targeting type was usedFurther 38 participantsrsquo files contained data for fewer thansix ads In these cases we showed participants explanationsfor all ads in their file The targeting types and specific ex-ample instances randomly chosen for each participant hadinherent variance Some targeting types had more potentialinstances than others Some instances undoubtedly seemedcreepier or more invasive than others even within the sametargeting type To account for these issues we recruited sev-eral hundred participants and focused on comparisons amongtargeting types and explanations interpreting our results ac-cordingly Additionally the more detailed explanations wereless readable and participants may have been more likely toskim them We performed a broad exploration of the designspace in an effort to understand what features participantsliked and disliked There is a trade-off between readabilityand comprehensiveness that future work should address

4 Results

In this section we first characterize current ad-targeting prac-tices by analyzing our 231 participantsrsquo Twitter data We thenreport participantsrsquo reactions to targeting mechanisms as wellas to six potential ad explanations from our online survey

We observed 30 different targeting types in use some withthousands of unique instances Participantsrsquo perceptions of

fairness comfort and desirability differed starkly by typebut comfort and desirability generally increased with the per-ceived accuracy of the targeting Further all three ad expla-nations we designed (based on the literature) outperformedexplanations currently deployed on Twitter and Facebook

41 ParticipantsWe report on data from the 231 participants who uploadedtheir Twitter data completed all parts of the study and wroteon-topic answers to free-response prompts Our participantshad been on Twitter for between 1 month and 123 years withan average of 66 years Two-thirds of participants reportedspending under an hour a day on Twitter Among participants528 identified as female 840 reported at least some col-lege education and 208 percent reported some backgroundin computer science or IT When asked early in the surveyparticipants only recognized an average of 16 companies(min 0 max 8) out of a random sample of 10 companies thathad shown them ads in the past 90 days Interestingly morethan 50 participants reported looking at their files before thesurvey Although this may have biased participants regardingspecific ads shown this is unlikely given both the large num-ber of files found in the original data download and the largesize of the ad-impressionsjs files containing per-ad dataParticipants would have had to parse many blocks like the onein Appendix A [1] and particularly notice the specific oneswe asked about

42 Aggregate Overview of TargetingParticipants had an average of 10466 ads in their files (min1 max 14035) a full histogram of ad impressions is shownin Appendix H Figure 8 [1] Our 231 participantsrsquo data filescollectively contained 240651 ads that had been targeted withat least one targeting type As detailed in Table 1 we observed30 different targeting types with 45209 unique instances ofthose targeting types

Usage of the different targeting types varied greatly asshown in Figure 3 (left) The most commonly used typeswere location (992 of all ads) and age (723) The leastcommonly used was flexible audience lookalikes (02) Asingle ad could be targeted using multiple instances of a giventype but Language age and gender targeting always usedone instance In contrast follower lookalikes and keywordsoften employed multiple instances 60 and 49 instances onaverage per ad respectively The largest set we observed was158 behavior instances Figure 3 (center) shows how oftenmultiple instances were used to target a given ad

For nine targeting types we observed fewer than ten uniqueinstances (eg male and female were the only two genderinstances) In contrast keywords (25745) follower looka-likes (8792) and tailored lists (2338) had the most uniqueinstances across participants For many targeting types the


Figure 3 Summaries of our 231 participantsrsquo Twitter ad data Left The fraction of ads seen by each participant that included eachtargeting type Center Instances of each targeting type per ad Right Unique instances of each targeting type per participant

median participant encountered dozens or hundreds of uniqueinstances of that type as shown in Figure 3 (right)

43 Detailed Usage of Targeting TypesNext we detail how each targeting type was used to target adsto our participants Based on the source of the data underlyingeach type we grouped the targeting types into three clustersThe first two clusters mdash targeting types related to user demo-graphics and targeting types related to user psychographics(behaviors and interests) mdash use information collected directlyby Twitter In contrast the third cluster consists of targetingtypes using data provided by prospective advertisers

431 Twitter Demographic Targeting Types

The first of our three clusters consists of demographic-basedtargeting We include in this category characteristics aboutboth a person and their device(s) Sometimes users directlyprovide this information to Twitter (eg providing a birthdate upon registration) In other cases Twitter infers this data

Advertisers commonly used demographics to targetbroad audiences Language was used frequently with En-glish being the most popularly targeted (208 participants)Age targeting was also extremely common yet also usedcoarsely (only 23 unique instances) ldquo18 and uprdquo was themost frequently targeted value 8311 of participants weretargeted on this attribute Many age instances overlapped (egldquo18 and uprdquo ldquo18 to 24rdquo ldquo18 to 34rdquo ldquo18 to 49rdquo) The five mostfrequently observed locations were the US UK Los AngelesLondon and Chicago We also observed locations as granularas ZIP codes (eg 44805 for Ashland OH) Different adsfor a single participant were sometimes targeted to multiplenon-overlapping locations demonstrating that their Twitterlocation changed over time Gender targeting was much lessfrequently used than language age or location Almost 70of gender instances targeted women The READMEtxt fileaccompanying data downloads says that Twitter infers a userrsquosgender if they did not provide it our analysis (and others [44])support this assertion We also found that this inference may

change over time 199 were targeted as male in some adsand female in others

Twitter also collects data about usersrsquo devices for tar-geting [67] Platform was used to target ads to users of iOS(115 participants) desktop (115) and Android (98) In total14605 ads were targeted to iOS users while 8863 were tar-geted to Android users The most frequently targeted devicemodels were the iPhone 8 Galaxy Note 9 iPhone 8 Plusand iPhone 7 Participants were often associated with multi-ple instances (eg both Android Lollipop and Jelly Bean) oreven targeted cross-OS (eg both Android Marshmallow andiOS 124) Twitter also offers targeting of Twitter users on anew device 626 of the 243 instances we observed were todevices Twitter designated as 1 month old (as opposed to 23 or 6 months) Advertisers also targeted by carrier mostcommonly to T-Mobile (21 participants) and O2 (19)

432 Twitter Psychographic Targeting Types

We next discuss targeting types related to participantsrsquo psy-chographic attributes which users provide via Twitter activityor which are inferred by Twitterrsquos algorithms Psychographicattributes relate to a userrsquos lifestyle behavioral or attitudinalpropensities [26] Although ldquobehavioral targetingrdquo is com-monly used in industry and research as an umbrella term forall forms of psychographic targeting we describe the rangeof targeting based on user behaviors and attitudes as psycho-graphic in contrast to the specific behavior targeting typeoffered by Twitter While some participants may be awareof the inferences that could be made about them from theirTwitter activity many likely are not [73]

Some of the most frequently used psychographic tar-geting types are based directly on usersrsquo Twitter activityFollowers of a user id which targets all followers of the sameTwitter account was used 590502 times in our data Out ofthe five of the most commonly targeted values four were re-lated to news agencies WSJ nytimes TheEconomistwashingtonpost and BillGates Keywords which are se-lected by advertisers and approved by Twitter [65] was themost unique targeting type with a total of 25745 distinct


Unique Most FrequentlyTargeting Type Total Uses Instances Observed Instance

Source Twitter (Demographic)Language 350121 4 EnglishAge 173917 23 18 and upPlatform 32351 4 iOSLocation 31984 566 United StatesOS version 7382 29 iOS 100 and aboveDevice model 2747 36 iPhone 8Carriers 1442 11 T-Mobile UKGender 1327 2 FemaleNew device 236 4 1 monthWiFi-Only 108 1 WiFi-Only

Source Twitter (Psychographic)Followers of a user ID 590502 138 nytimesFollower lookalikes 242709 8792 netflixConversation topics 128005 2113 FoodKeyword 91841 25745 parentingBehavior 35088 854 US - Household income

$30000-$39000Interest 25284 206 ComedyMovies and TV shows 22590 548 Love IslandEvent 17778 198 2019 Womenrsquos World CupRetargeting campaign 15529 1842 Retargeting campaign

engager Retargeting engagement type 11185 5 Retargeting engagement

type Retargeting user engager 2184 218 Retargeting user engager

Retargeting lookalikes 229 66 Nielson Online - Website

Visitors - FinanceIn

Source AdvertiserTailored audience (list) 113952 2338 Lifetime Suppression

[Installs] (Device Id)Mobile audience 21631 478 Purchase Postmates - Local

Restaurant DeliveryTailored audience (web) 18016 550 Quote FinishTailored audience CRM lookalikes 1179 22 Samba TV gt Mediacom -

Allergan - Botox ChronicFlexible audience 382 12 iOS gt Recently Installed

(14days) No CheckoutMobile lookalikes 141 23 Install New York Times

Crossword IOS AllFlexible audience lookalike 7 2 All WBGs Android

Purchase Events

Source Unknown (as labeled by Twitter)Unknown 927 179 Unknown

Table 1 Targeting types observed in our 231 participantsrsquoTwitter data We report how many of the 240651 ads weretargeted by that type as well as the number of unique instancesof that type and the most frequently observed instance Wegroup targeting types by their source (advertisers or Twitter) indicates targeting types also studied in the user survey

instances Keywords varied greatly in content and specificityranging from ldquotechnologyrdquo and ldquofoodrdquo to ldquofirst homerdquo (usedby realtorcom) and ldquoidiopathic thrombocytopenic purpurardquo(used by WEGO Health) We identified several keywords aspotentially violating Twitter policies prohibiting targeting tosensitive categories ldquosuch as race religion politics sex lifeor healthrdquo [65 69] Examples include ldquoostomyrdquo ldquoGayrdquo andldquolatinasrdquo (see Table 2 for more) Twitter infers conversationtopic instances based on usersrsquo Twitter activity (Tweets clicksetc) allowing advertisers to target narrow populations abouta third of our unique conversation instances were in onlyone userrsquos ad data The top five topics however were broadldquotechnologyrdquo ldquofoodrdquo ldquotravelrdquo ldquosoccerrdquo and ldquofashionrdquo

Inferences made for interests targeting are one stepmore abstract they are inferred from the accounts a user

follows (and the content from those accounts) as well as theirdirect activities The top five interests were similar to thetop five conversation topics ldquocomedyrdquo ldquotech newsrdquo ldquotech-nologyrdquo ldquomusic festivals and concertsrdquo and ldquosoccerrdquo Othertargeted interests were more specific such as ldquovintage carsrdquoand ldquoscreenwritingrdquo

Similarly to interests the event and movies and TV showstargeting types appear to rely on both a userrsquos direct activitiesand on inferences to label users as interested in offline eventsand entertainment These targeting types most commonlyreflected sports (ldquo2019 Womenrsquos World Cuprdquo 2713 instancesldquoMLB Season 2019rdquo 1344 instances) and popular shows suchas ldquoLove Islandrdquo ldquoStranger Thingsrdquo and ldquoGame of Thronesrdquo

Highly targeted psychographic targeting types arebased on Twitter algorithms Follower Lookalikes target-ing is even more indirect the targeted users are labeled assharing interests or demographics with followers of a par-ticular account despite not actually following that accountFollower lookalikes is the second most individualized target-ing type in our dataset (after keywords) with 8792 distincttargeted values A majority of these values (4126) were as-sociated with a single participant (eg one participant wastargeted as a follower look-alike of FDAOncology while26 users were targeted as follower lookalikes of Speaker-Pelosi) However a few well-known handles were frequentlythe focus of lookalikes netflix (used in targeting 5199 ads)espn (3608) and nytimes (3440)

Behavior targeting one specific targeting type offered byTwitter within the full range of psychographic targeting typesis based on inferences drawn from proprietary algorithms Ourmost commonly observed instances were related to income orlifestyles (eg ldquoUS - Household income $30000 - $39999rdquoldquoUS - ExecutiveC-suiterdquo ldquoUS - Presence in household yes rdquoldquoUS - Fit momsrdquo) Some were surprisingly specific ldquoHome in-surance expiration month 10 Octoberrdquo ldquoUS - Likely to switchcell phone providersrdquo ldquoCountry Club Climbers - SuburbanEmpty Nesters K59rdquo and ldquoUS - Animal charity donorsrdquo

Finally Twitter offers four retargeting types based on pre-vious user engagement with ads There were 15814 uses(1812 unique instances) of retargeting campaign targetingwhich targets users who responded to an advertiserrsquos priorcampaign The ambiguous naming of these instances (ldquoRetar-geting campaign engager rdquo) makes them hard tointerpret in detail Retargeting user engager used 707 timesis similarly vague Retargeting custom audience lookaliketargeting which combines retargeting with Twitterrsquos looka-like algorithms was very rarely used in our data

433 Advertiser Targeting Types

The final category of targeting types use advertiser-providedinformation Instead of providing any targeting data Twitteronly facilitates matching to Twitter users via Twitter user-names email addresses or other identifiers Notably adver-


tiser targeting types are also the most covert from a userrsquosperspective while Twitter-provided data could potentiallybe deduced from the standard user interface (eg interestsbased on likes or Retweets) targeting types using advertiser-provided data are completely unrelated to Twitter activity

Tailored audience (lists) match Twitter users to lists up-loaded by advertisers We found 113952 instances of listtargeting across 2338 unique lists companies using listtargeting the most were Anker (22426 instances) Post-mates (11986) Rockstar Games (8494) and Twitter Sur-veys (3131) Tailored lists often used words like lsquoNegativersquolsquoHoldoutrsquo and lsquoBlacklistrsquo which we hypothesize referenceconsumers who previously opted out of receiving targetedads or content via other mediums Advertisers may also uselist targeting for targeting offline purchasers as list namesincluded the words lsquoPurchasersquo and lsquoBuyersrsquo Many lists usenaming schemes that make it difficult or impossible to dis-cern the contents of the lists (eg ldquo__rdquoldquo_MM_YY__rdquo)

We identified several lists with names that sug-gest targeting on attributes prohibited by Twit-terrsquos policies (see Table 2) including financial status(ldquoYYYY account status balance duerdquo) race (ldquo_Nis-san_AfricanAmericans_YYYYMMrdquo) religion (ldquoChristianAudience to Excluderdquo) or sex life (ldquoLGBT SuppressionListrdquo) [66] Tailored audience (web) also consists ofadvertiser-collected lists of website visitors eg ldquoStartedNew Credit Card Applicationrdquo or ldquoRegistered but notActivated User on Cloudrdquo This targeting type thereforeconnects usersrsquo potentially sensitive browsing activity to theirTwitter accounts in ways that may violate Twitterrsquos healthadvertising policies [64]

Tailored audience CRM lookalike targeting combinesadvertiser lists with the lookalike algorithm to find Twit-ter users who may be similar to known current or poten-tial customers We observed this mechanism being usedin incredibly specific ways such as to find users similarto ldquoQSR Ice Cream Frozen Yogurt Frequent Spenderrdquo orldquoFrozen_Snacks_Not_Frozen_Yogurt_Or_Ice_Cream_Used_in_last_6_months_Principal_Shoppers_Primary_Fla_Vor_Ice_rdquo both used by advertiser Dairy Queen

Twitter also offers targeting types that enable cross-platform tracking Mobile audience targets Twitter users whoalso use an advertiser-owned mobile app (ie ldquopeople whohave taken a specific action in your app such as installs orsign-upsrdquo [68]) Instances reflect the userrsquos status with theapp app name and mobile platform eg ldquoInstall GeminiBuy Bitcoin Instantly ANDROID Allrdquo and ldquoInstall Lumen -Over 50 Dating IOS Allrdquo Mobile audience lookalike target-ing which combines the prior mechanism with the lookalikealgorithm was rarely used Flexible audience targeting al-lows advertisers to combine tailored audiences (lists web ormobile) using AND OR and NOT operations We observedseven ads using this type all from one advertiser

Targeting Value Policy Advertiser(s)

Keywordsostomy Health ConvaTec Stoma UKunemployment Financial Giant Eagle JobsGay Sex Life HampM United Kingdommexican american Race Just Mercy Doctor Sleep

The Kitchen MovieAfricanAmerican Race sephoranative Race sephorahispanics Race sephoralatinas Race sephoramexican Race sephora-Racist Religion xbox

Conversation TopicsLiberal Democrats (UK) Politics Channel 5 Irina von

Wiese MEP

Tailored Audience (List)YYYY account status balance due(translated from Mandarin Chinese)

Financial Anker

segment_Control | Rising Hispanics | EmailOpeners_

Race Big Lots

segment_Control | Rising Hispanics |

Non-Opener_Race Big Lots

lowastlowastlowast_Nissan_AfricanAmericans_YYYYMM Race NissanChristian Audience to Exclude Religion nycHealthyLGBT Suppression List Sex Life nycHealthyASL Marketing gt Hispanic Millennials -

Race Verizon

Tailored Audience (Web)Website Retargeting - Tagrissocom (a siteabout lung cancer therapy)

Health Test Lung Cancer

Table 2 Examples of targeted ads that could be seen as vi-olating Twitterrsquos keyword targeting policy (see Appendix FFigure 7 [1]) or Twitterrsquos privacy policy ldquo our ads policiesprohibit advertisers from targeting ads based on categoriesthat we consider sensitive or are prohibited by law such asrace religion politics sex life or healthrdquo [69]

Finally for the curiously-named targeting type unknown25 participants were associated with a single instance (ldquoUn-known ) all related to the advertiser ldquoTwitter Surveys

44 Participant Reactions to Targeting Types

One key benefit of our study design is that we could ask par-ticipants questions about advertising criteria actually usedin ads they saw Participants answered questions about upto four randomly selected targeting types filtered by thosepresent in their uploaded data Advertisers used certain target-ing types more often than others meaning different numbersof participants saw each type (see Appendix G Table 4 [1])

441 Fairness Comfort Desirability and Accuracy

Participants perceived language age and interest target-ing to be the most fair with 863 720 and 690 agree-ing respectively (Figure 4) Overall few participants thoughtany given targeting type was unfair to use no type had morethan 50 of participants disagree that its use would be fair(Figure 4 General Fair) Tailored audience (list) which wasperceived as least fair overall was still roughly evenly split


Tailored lists

Mobile

Tailored web

Behavior

Event

Lookalikes

Retargeting

MovieTV

Location

Gender

Keyword

Conversation

Platform

Interest

Age

Language

0 25 50 75 100

General Fair

0 25 50 75 100

General Comfortable

0 25 50 75 100

General Want

0 25 50 75 100

Specific Want

0 25 50 75 100

Specific Accurate

Stronglyagree

Agree Neither DisagreeStronglydisagree

Figure 4 Participantsrsquo level of agreement to questions about targeting types in general and specific instances

between participants agreeing and disagreeing Compared tothe regression baseline (interest) participants were signifi-cantly more likely to find language targeting fair (OR = 448p lt 0001) Retargeting campaign age and platform target-ing were not statistically different from interest (α = 005)Participants found all other targeting types significantly lessfair than interest (OR = 00607minus0401 all p lt 005)

To dig deeper into perceptions of fairness we asked par-ticipants to elaborate on their Likert-scale answers in a free-response question gathering a total of 898 responses Partic-ipants had varying conceptions of the meaning of fairnessSome equated fairness with utility some equated fairnesswith comfort and some equated fairness with accuracy ofthe information Across all targeting types the most commonrationale used to judge fairness were that targeting is usefulto the user in some way (248) For instance participantsmentioned that they preferred to see relevant rather than ran-dom ads if they had to see ads at all and that advertisingallows them to access Twitter for free 146 said that tar-geting was fair because the advertiser benefited in some waynamely by increased effectiveness of advertising These tworationales centered on deriving benefits either for advertisersor users but failed to consider the privacy or data autonomyof the participant Others considered that Twitter is a publicplatform ldquoTwitter is pretty much a public arena if I wereshouting about various topics in a town square people wouldinfer my interests from that and potentially attempt to profitfrom themrdquo (P191) Participantsrsquo rationales seemed to as-sume that personalized targeting types like these must be usedfor advertising Only a few suggested profiting off of usersrsquoprivate information was fundamentally unfair

Perceptions of comfort largely aligned with percep-tions of fairness with small exceptions For example par-ticipants rated gender and keyword targeting as more fairthan location targeting but were curiously more comfortablewith location than gender and keyword (Figure 4 GeneralComfortable) Some participantsrsquo comments suggested dis-

comfort may relate to whether participants understood howdata about them was obtained P184 commented ldquoIrsquom notsure how they would know my income level Disturbingrdquo

We were also curious about participantsrsquo desire for ad-vertising that used each targeting type and found generalaffirmation with some strong opposition to specific in-stances We told participants to assume the number of adsthey would see would stay the same and asked them to con-sider how much they would want to see ads targeted with agiven type for both a specific instance of that type and fortype generally As an example 538 of participants whosaw an instance of event targeting disagreed that it describedthem accurately and 650 disagreed that they would want tosee advertising based on that specific example However only250 disagreed that they would want to see ads utilizingevent targeting in general

In the general case participants were significantly morelikely to want ads that used language targeting than theregression-baseline interest (OR = 33 p = 0004) All othertargeting types were significantly less wanted than interest(OR = 01minus04 all p lt 005)

Participants found specific instances of some demo-graphic targeting types to be very accurate but other psy-chographic types to be very inaccurate More than half ofparticipants strongly agreed that a specific instances of lan-guage age platform gender location targeting was accuratefor them while more than half strongly disagreed that re-targeting tailored web and mobile targeting was accurate(Figure 4 Specific Accurate) Participants were more likelyto agree that specific instances of platform language genderand age targeting described them accurately compared to aspecific instance of interest (OR = 29minus 97 all p lt 001)Specific instances of movies and TV shows location and be-havior targeting were not significantly different from interestin agreed accuracy (α = 005) while all remaining signifi-cant targeting types were less likely to be rated as accurate(OR = 01minus05 all p lt 005) As we found in their initial free-


Response ρ p

General Fair 0332 lt001General Comfortable 0366 lt001General Want 0341 lt001Specific Comfortable 0614 lt001Specific Want 0732 lt001

Table 3 Spearmanrsquos ρ correlation between participantsrsquo agree-ment with Specific Accurate (ldquoSpecific instance describesme accuratelyrdquo) and their other Likert-scale responses

response reactions to uses of a particular targeting type in theirdata if participants perceived an instance of targeting to beaccurate it was generally well-received Participants seemedto enjoy seeing information being accurately reflected aboutthemselves as P189 described about conversation targetingldquoI am okay with this Itrsquos cool how accurate it isrdquo

As shown in Table 3 the accuracy of a specific instanceof a targeting type was significantly correlated with all ofour other measurements of participantsrsquo perceptions Thatis when participants disagreed that a specific instance ofa targeting type described them accurately they were alsosignificantly less likely to be comfortable with that instancebeing used (ρ = 0614 p lt 0001) and to want to see more adsbased on that instance (ρ = 0732 p lt 0001) We found simi-lar correlations for perceptions of the use of a targeting typegenerally It is possible that inaccuracy leads to perceptionsof discomfort and unwantedness it is also possible that whenpeople see ads they find undesirable they are less likely tobelieve the associated targeting is accurate

Even if a majority of people are comfortable with certaintargeting in the abstract it is important to understand andpotentially design for those who feel less comfortable Toexplore this we looked for participants who consistently dis-agreed with questions about fairness comfort and desirabilityIn particular for each of the questions presented in Figure 4besides Specific Accurate we generated a median responsefor each participant of the up to four targeting types they wereasked questions about From this we found only 23 of our 231participants disagreed or strongly disagreed as their medianresponse for all 4 questions

442 Targeting Types Awareness and Reactions

We were also interested in participantsrsquo familiarity with ormisconceptions of the various targeting types Before partici-pants were given any information about a targeting type weshowed them the term Twitter uses to describe that type [63]and asked them to indicate their current understanding or bestguess of what that term meant in the context of online adver-tising Nearly all participants had accurate mental models oflocation age gender and keyword targeting likely becausethese types are fairly well-known and straightforward Further93 of participants asked about interest correctly defined itsuggesting it is also relatively straightforward In fact some

participants confused other targeting types with interest tar-geting ldquoI have never heard this term before Irsquom guessing thatthey target ads based on your followersrsquo interests as wellrdquo(P161 on follower lookalike targeting)

Tailored audience (list) behavior and mobile audiencetargeting were the least well understood with 964970 and 100 of participants respectively providing anincorrect or only partially correct definition The first tworely on offline data being connected with participantsrsquo onlineaccounts but most participants incorrectly defined the termonly based on online activities Mobile audience targetingwas misunderstood due to different interpretations of ldquomobilerdquo(eg P122 guessed ldquoadvertising based on your phone net-workrdquo) or other mobile details The correct answer relatesto the userrsquos interactions with mobile apps Participants alsofrequently believed a targeting type meant advertising thattype of thing (eg an event) as opposed to leveraging userdata about that thing for targeting ads (eg targeting a productonly to users who attended an event)

While 636 of participants who were asked to define lan-guage targeting correctly referenced the userrsquos primary lan-guage many of the 288 who incorrectly defined it posed amore involved and potentially privacy-invasive definition ldquoIsuppose that language targeting would be communicating in away that is targeted to how that specific person communicatesFor example as a millennial I would want to see language thatis similar to how I speak rather than how someone who is myparents age would speakrdquo (P76) Platform targeting was sim-ilarly misunderstood with some participants believing thatthis was the practice of targeting by social media platformuse or even political platform ldquoIt looks at my list of peopleI follow and sends me ads based on what they believe mypolitical stance isrdquo (P147) We also found evidence acrosstargeting types of the belief that advertising is based on sur-reptitious recordings of phone audio For example P231 saidof conversation targeting ldquoGiven what I know about howphone microphones are always on I would guess that itrsquoswhen ads pop up based on what Irsquove said in a conversationrdquo

45 Participant Responses to Ad Explanations

We examined reactions to our ad explanations among the 193participants who saw all six variants Our approximation ofTwitterrsquos current explanation served as our primary basis ofcomparison We also report qualitative opinions about whatwas memorable perceived to be missing or would be ideal

451 Comparing Our Ad Explanations to Twitterrsquos

Overall participants found explanations containingmore detail to be more useful as shown in Figure 5 Unsur-prisingly Control was the least useful explanation only 313of participants agreed it was useful This is significantly lessthan our Twitter baseline where 488 agreed (V = 63445


Control

Creepy

Detailed Text

Detailed Visual

Facebook

Twitter

0 25 50 75 100

Useful

0 25 50 75 100

Understand How Targeted

0 25 50 75 100

Concerned About Privacy

0 25 50 75 100

Increased Trust

0 25 50 75 100

Want Similar

Stronglyagree


Figure 5 Participantsrsquo level of agreement to questions about ad explanations

p lt 0001) The Facebook explanation was comparable to theTwitter explanation (414 agreed V = 35200 p = 0154) Incontrast the three explanations we designed were rated as sig-nificantly more useful than Twitterrsquos (V = 13525ndash23360 allp lt 0001) Specifically 636 712 and 786 of partici-pants respectively agreed the Detailed Text Detailed Visualand Creepy explanations were useful

The usefulness ratings closely resembled responses towhether the explanation provided ldquoenough information tounderstand how the ad was chosen for me Again Twitterperformed better than only Control (V 59060 p lt 0001)and did not significantly differ from Facebook (V = 42610p = 0091) Participants agreed our explanationsmdashDetailedText Detailed Visual Creepymdashwere most helpful in under-standing how they were targeted all three significantly dif-fered from Twitter (V = 19280ndash28780 all p le 0001)

We saw a different trend for privacy concern 772 of par-ticipants agreed Creepy made them ldquomore concerned aboutmy online privacyrdquo compared to 348 for Twitter and just282 for the Control Privacy concern for Creepy was signif-icantly higher than for Twitter (V = 9895 p lt 0001) BothFacebook and Detailed Text also exhibited significantly moreconcern than Twitter (V = 18210 28350 p = 0002 0015)but to a lesser extent Respondents reported comparable pri-vacy concern for the Twitter explanation as for Detailed Visualand Control (V = 20295 37510 p = 00800064)

Transparency and usefulness generally did not trans-late to increased trust in an advertiser In fact only a mi-nority of participants agreed that they trusted the advertisermore as a result of any provided ad explanation Only the De-tailed Visual explanation increased trust significantly relativeto Twitter (V = 16955 p lt 0001)

A majority of participants agreed they would ldquowant an adexplanation similar to this one for all ads I see on Twitter forour Creepy (688) Detailed Visual (644) and DetailedText (549) versions Agreement for these was significantlylarger (V = 17985ndash21320 all p lt 0001) than the 398who wanted Twitter-like Participants significantly preferredTwitter to the Control (V = 68315 p lt 0001) but not toFacebook (V = 42490 p = 0339)

452 Qualitative Responses to Ad Explanations

Participants want detail and indicators of non-use Weasked participants what they found most memorable abouteach ad explanation For Control Facebook and Twitter mostmemorable was how little detail they gave about how partici-pants were targeted (307 216 and 133 of participantsrespectively) By comparison 163 (Detailed Text) 79(Visual) and 31 (Creepy) of participants noted a lack ofdetail as the most memorable part Conversely 817 foundthe amount of detail in Creepy to be the most memorable partfollowed by 612 for Visual These findings may be becauseCreepy included the most information and Detailed Visualindicated which targeting types were not used

Ambiguity was perceived as missing information Wealso asked participants what information if any they thoughtwas missing from each ad explanation We wanted to helpparticipants identify what information could be missing soour within-subjects design featured randomly-shown variantsthat demonstrated information that could be included In linewith the quantitative results for usefulness our Detailed Vi-sual Detailed Text and Creepy explanations performed bestwith 612 589 and 530 of participants respectivelyanswering nothing was missing Conversely Facebook Con-trol and Twitter performed less well with 692 673 and524 respectively of participants stating that some informa-tion was missing or unclear For Detailed Text and DetailedVisual among the most commonly noted missing informationrelated to our use of ldquomayrdquo and ldquomightrdquo about which criteriaactually were matched the participant This was necessitatedby the ambiguity of the Twitter files (prior to receiving aclarification from Twitter see Section 36 for details) ForFacebook the most commonly missing information was as-sociated with the hashed tailored audience list several wrotethat they did not know what a hashed list was P125 wroteldquoThe nature of the list mentioned should be clarified in somedetail Itrsquos unfair to be put on a list without access to what thelist is and who compiled it and who has access to itrdquo

Describing their ideal Twitter ad explanation 468 ofparticipants wanted to see the specific actions (eg what theyTweeted or clicked on) or demographics that caused them tosee a given ad 342 wanted to know more about how theadvertiser obtained their information They also wanted clearlanguage (190) and settings for controlling ads (134)


5 Discussion

We study Twitterrsquos targeted advertising mechanisms whichcategorize users by demographic and psychographic at-tributes as determined from information provided by the userprovided by advertisers or inferred by the platform Whileprior work has surfaced and studied user reactions to ad target-ing as a whole [20 70] or specific mechanisms like inferredinterests [17] our work details advertisersrsquo use of 30 uniquetargeting types and investigates user perceptions into 16 ofthem These distinct types including follower lookalikesand tailored audiences are rarely studied by the academiccommunity but frequently used by advertisers (see Table 1)Our participants expressed greater discomfort with some ofthese less studied targeting types highlighting a need forfuture work

We complement existing work on Facebook ad trans-parency by investigating ad explanations on a different plat-form Twitter and using participantsrsquo own Twitter data toevaluate them Strengthening prior qualitative work [20] wequantitatively find that our participants preferred ad expla-nations with richer information than currently provided byFacebook and Twitter We also find significant user confusionwith ldquohashedrdquo lists a term introduced to ad explanations byFacebook in 2019 [55] to explain how platforms match userdata to information on advertiser-uploaded lists for tailoredaudience targeting (called custom audiences on Facebook)

Can sensitive targeting be prohibited in practice Wefind several instances of ad targeting that appear to violateTwitterrsquos stated policy prohibiting targeting on sensitive at-tributes Such targeting is often considered distasteful and insome cases may even be illegal We observed these instancesmost commonly in targeting types where advertisers providecritical information keywords (where advertisers can pro-vide any keyword of choice subject to Twitter acceptance)and variations of tailored audiences where the advertiserprovides the list of users to target Potentially discriminatorykeywords are a problem that Twitter could theoretically solvegiven a sufficiently accurate detection algorithm or (morelikely) manual review Tailored audiences however are morepernicious Advertisers can use any criteria to generate a listWe were only able to identify potentially problematic casesbecause the list name which is under advertiser control hap-pened to be meaningfully descriptive It would be trivial for anadvertiser to name a list generically to skirt scrutiny callinginto question whether Twitterrsquos policy on sensitive attributeshas (or can have) any real force in practice It also raises largerconcerns about regulating potentially illegal or discriminatorypractices as long as tailored audiences remain available

More accuracy fewer problems Similarly to prior workwe found that the perceived inaccuracy of targeting instancescorrelates with users having less desire for such targeting

to be used for them [14 17] This has potentially danger-ous implications If accuracy reduces discomfort this mayappear to justify increasing invasions of privacy to obtain ever-more-precise labels for users However participantsrsquo free-textresponses indicate an upper bound where increasing accu-racy is no longer comfortable For example P220 noted thata specific instance of location targeting was ldquovery accurate but I donrsquot really like how they are able to do that withoutmy knowledge and even show me ad content related to mylocation because I choose not to put my specific location onmy Twitter account in any way for a reasonrdquo Future workshould investigate how and when accuracy crosses the linefrom useful to creepy

Transparency A long way to go This work also con-tributes a deeper understanding of ad explanations amid sub-stantial ongoing work on transparency as perhaps the onlyway for the general public to scrutinize the associated costsand benefits Participants found our ad explanations whichprovide more details significantly more useful understand-able and desirable than currently deployed ad explanationsfrom Twitter and Facebook However our results also high-light a significant challenge for transparency platform andadvertiser incentives Some of our proposed explanations de-spite being more useful decreased participant trust in theadvertiser which clearly presents a conflict of interest Thisconflict may explain why currently deployed explanations areless complete or informative than they could be

Finally our results suggest it is insufficient to simply re-quire data processing companies to make information avail-able While the option to download advertising data is a strongfirst step key aspects of the ad ecosystem mdash such as the ori-gins of most targeting information mdash remain opaque In addi-tion even as researchers with significant expertise we strug-gled to understand the data Twitter provided (see Section 36)This creates doubt that casual users can meaningfully under-stand and evaluate the information they receive However ourparticipants indicated in free response answers that they foundthe transparency information provided in our study useful andthat it illuminated aspects of tracking they had not previouslyunderstood making it clear that comprehensible transparencyhas value We therefore argue that transparency regulationsshould mandate that raw data files be accompanied by cleardescriptions of their contents and researchers should developtools and visualizations to make this raw data meaningful tousers who want to explore it

Acknowledgments

We gratefully acknowledge support from the Data Trans-parency Lab and Mozilla as well as from a UMIACS contractunder the partnership between the University of Marylandand DoD The views expressed are our own


References

[1] Online appendix httpswwwblaseurcompapersusenix20twitterappendixpdf

[2] Gunes Acar Christian Eubank Steven Englehardt MarcJuarez Arvind Narayanan and Claudia Diaz The WebNever Forgets Persistent Tracking Mechanisms in theWild In Proc CCS 2014

[3] Lalit Agarwal Nisheeth Shrivastava Sharad Jaiswal andSaurabh Panjwani Do Not Embarrass Re-ExaminingUser Concerns for Online Tracking and Advertising InProc SOUPS 2013

[4] Muhammad Ali Piotr Sapiezynski Miranda BogenAleksandra Korolova Alan Mislove and Aaron RiekeDiscrimination Through Optimization How FacebookrsquosAd Delivery Can Lead to Skewed Outcomes In ProcCSCW 2019

[5] Hazim Almuhimedi Florian Schaub Norman SadehIdris Adjerid Alessandro Acquisti Joshua Gluck LorrieCranor and Yuvraj Agarwal Your Location has beenShared 5398 Times A Field Study on Mobile AppPrivacy Nudging In Proc CHI 2015

[6] Athanasios Andreou Maacutercio Silva Fabriacutecio Ben-evenuto Oana Goga Patrick Loiseau and Alan MisloveMeasuring the Facebook Advertising Ecosystem InProc NDSS 2019

[7] Athanasios Andreou Giridhari Venkatadri Oana GogaKrishna P Gummadi Patrick Loiseau and Alan MisloveInvestigating Ad Transparency Mechanisms in SocialMedia A Case Study of Facebookrsquos Explanations InProc NDSS 2018

[8] Julia Angwin and Terry Parris Facebook Lets Adver-tisers Exclude Users by Race ProPublica October 282016

[9] ashrivas More Relevant Ads with TailoredAudiences Twitter Blog December 2013httpsblogtwittercommarketingen_usa2013more-relevant-ads-with-tailored-audienceshtml

[10] Rebecca Balebako Pedro Leon Richard Shay Blase UrYang Wang and Lorrie Faith Cranor Measuring theEffectiveness of Privacy Tools for Limiting BehavioralAdvertising In Proc W2SP 2012

[11] Muhammad Ahmad Bashir Sajjad Arshad and WilliamRobertson Tracing Information Flows Between AdExchanges Using Retargeted Ads In Proc USENIXSecurity 2016

[12] Muhammad Ahmad Bashir Umar Farooq MaryamShahid Muhammad Fareed Zaffar and Christo WilsonQuantity vs Quality Evaluating User Interest ProfilesUsing Ad Preference Managers In Proc NDSS 2019

[13] Muhammad Ahmad Bashir and Christo Wilson Diffu-sion of User Tracking Data in the Online AdvertisingEcosystem In Proc PETS 2018

[14] Rena Coen Emily Paul Pavel Vanegas AletheaLange and GS Hans A User-CenteredPerspective on Algorithmic PersonalizationMasterrsquos thesis Berkeley School of Informa-tion httpswwwischoolberkeleyeduprojects2016user-centeredperspective-algorithmic-personalization 2016

[15] Amit Datta Michael Carl Tschantz and Anupam DattaAutomated Experiments on Ad Privacy Settings InProc PETS 2015

[16] Martin Degeling and Jan Nierhoff Tracking and Trick-ing a Profiler Automated Measuring and Influencing ofBluekairsquos Interest Profiling In Proc WPES 2018

[17] Claire Dolin Ben Weinshel Shawn Shan Chang MinHahn Euirim Choi Michelle L Mazurek and Blase UrUnpacking Privacy Perceptions of Data-Driven Infer-ences for Online Targeting and Personalization In ProcCHI 2018

[18] Serge Egelman Adrienne Porter Felt and David WagnerChoice Architecture and Smartphone Privacy TherersquosA Price for That In Workshop on the Economics ofInformation Security 2012

[19] Steven Englehardt and Arvind Narayanan Online Track-ing A 1-Million-Site Measurement and Analysis InProc CCS 2016

[20] Motahhare Eslami Sneha R Krishna Kumaran Chris-tian Sandvig and Karrie Karahalios CommunicatingAlgorithmic Process in Online Behavioral AdvertisingIn Proc CHI 2018

[21] Motahhare Eslami Kristen Vaccaro Karrie Karahaliosand Kevin Hamilton ldquoBe Careful Things Can Be Worsethan They Appearrdquo Understanding Biased Algorithmsand Usersrsquo Behavior around Them in Rating PlatformsIn Proc AAAI 2017

[22] Motahhare Eslami Kristen Vaccaro Min Kyung LeeAmit Elazari Bar On Eric Gilbert and Karrie Kara-halios User Attitudes towards Algorithmic Opacity andTransparency in Online Reviewing Platforms In ProcCHI 2019


[23] Facebook Coming Soon New Ways to Reach PeopleWhorsquove Visited Your Website or Mobile App FacebookBusiness October 15 2013 httpswwwfacebookcombusinessnewscustom-audiences

[24] Facebook Introducing New Requirements for CustomAudience Targeting Facebook Business June 2018httpswwwfacebookcombusinessnewsintroducing-new-requirements-for-custom-audience-targeting

[25] Irfan Faizullabhoy and Aleksandra Korolova Face-bookrsquos Advertising Platform New Attack Vectors andthe Need for Interventions In Proc ConPro 2018

[26] Faye W Gilbert and William E Warran PsychographicConstructs and Demographic Segments Psychologyand Marketing 12223ndash237 1995

[27] Google About Audience Targeting GoogleAds Help 2020 httpssupportgooglecomgoogle-adsanswer2497941hl=en

[28] Ruogu Kang Stephanie Brown Laura Dabbish and SaraKiesler Privacy Attitudes of Mechanical Turk Workersand the US Public In Proc USENIX Security 2014

[29] Saranga Komanduri Richard Shay Greg Norcie andBlase Ur Adchoices Compliance with Online Behav-ioral Advertising Notice and Choice Requirements ISA Journal of Law and Policy for the Information Society7603 2011

[30] Georgios Kontaxis and Monica Chew Tracking Protec-tion in Firefox for Privacy and Performance In ProcW2SP 2015

[31] Balachander Krishnamurthy Delfina Malandrino andCraig E Wills Measuring Privacy Loss and the Im-pact of Privacy Protection in Web Browsing In ProcSOUPS 2007

[32] KyleB Introducing Partner Audiences TwitterBlog March 5 2015 httpsblogtwittercommarketingen_usa2015introducing-partner-audienceshtml

[33] J Richard Landis and Gary G Koch The Measurementof Observer Agreement for Categorical Data Biomet-rics 33(1)159ndash174 1977

[34] Mathias Leacutecuyer Guillaume Ducoffe Francis Lan An-drei Papancea Theofilos Petsios Riley Spahn AugustinChaintreau and Roxana Geambasu XRay Enhancingthe Webrsquos Transparency with Differential CorrelationIn Proc USENIX Security 2014

[35] Mathias Lecuyer Riley Spahn Yannis Spiliopolous Au-gustin Chaintreau Roxana Geambasu and Daniel HsuSunlight Fine-grained Targeting Detection at Scale withStatistical Confidence In Proc CCS 2015

[36] Pedro Leon Justin Cranshaw Lorrie Faith Cranor JimGraves Manoj Hastak Blase Ur and Guzi Xu Whatdo Online Behavioral Advertising Privacy DisclosuresCommunicate to Users In Proc WPES 2012

[37] Pedro Leon Blase Ur Richard Shay Yang Wang Re-becca Balebako and Lorrie Cranor Why Johnny CanrsquotOpt out A Usability Evaluation of Tools to Limit OnlineBehavioral Advertising In Proc CHI 2012

[38] Adam Lerner Anna Kornfeld Simpson TadayoshiKohno and Franziska Roesner Internet Jones and theRaiders of the Lost Trackers An Archaeological Studyof Web Tracking from 1996 to 2016 In Proc USENIXSecurity 2016

[39] Qiang Ma Eeshan Wagh Jiayi Wen Zhen Xia RobertOrmandi and Datong Chen Score Look-Alike Audi-ences In Proc ICDMW 2016

[40] Aleksandar Matic Martin Pielot and Nuria OliverldquoOMG How did it know thatrdquo Reactions to Highly-Personalized Ads In Proc UMAP 2017

[41] Jonathan R Mayer and John C Mitchell Third-partyWeb Tracking Policy and Technology In Proc IEEESampP 2012

[42] Aleecia M McDonald and Lorrie Faith Cranor Ameri-cansrsquo Attitudes About Internet Behavioral AdvertisingPractices In Proc WPES 2010

[43] Jeremy B Merrill and Ariana Tobin FacebookMoves to Block Ad Transparency Tools mdash Includ-ing Ours ProPublica January 28 2019 httpswwwpropublicaorgarticlefacebook-blocks-ad-transparency-tools

[44] Katie Notopoulos Twitter Has Been GuessingYour Gender And People Are Pissed BuzzfeedNews May 2017 httpswwwbuzzfeednewscomarticlekatienotopoulostwitter-has-been-guessing-your-gender-and-people-are-pissed

[45] Jason R C Nurse and Oliver Buckley Behind theScenes a Cross-Country Study into Third-Party Web-site Referencing and the Online Advertising EcosystemHuman-centric Computing and Information Sciences7(1)40 2017

[46] Eyal Peer Laura Brandimarte Sonam Somat andAlessandro Acquisti Beyond the turk Alternative plat-forms for crowdsourcing behavioral research In Journalof Experimental Social Psychology 2017


[47] Angelisa C Plane Elissa M Redmiles Michelle LMazurek and Michael Carl Tschantz Exploring UserPerceptions of Discrimination in Online Targeted Ad-vertising In Proc USENIX Security 2017

[48] Emilee Rader and Rebecca Gray Understanding UserBeliefs About Algorithmic Curation in the FacebookNews Feed In Proc CHI 2015

[49] J H Randall The Analysis of Sensory Data by Gener-alized Linear Model Biometrical Journal 1989

[50] Ashwini Rao Florian Schaub and Norman Sadeh WhatDo They Know About Me Contents and Concerns ofOnline Behavioral Profiles In Proc ASE BigData 2014

[51] Elissa M Redmiles Sean Kross and Michelle LMazurek How Well Do My Results Generalize Com-paring Security and Privacy Survey Results from MTurkWeb and Telephone Samples In Proc IEEE SampP 2019

[52] Franziska Roesner Tadayoshi Kohno and David Wether-all Detecting and Defending Against Third-Party Track-ing on the Web In Proc USENIX Security 2012

[53] Sonam Samat Alessandro Acquisti and Linda BabcockRaise the Curtains The Effect of Awareness About Tar-geting on Consumer Attitudes and Purchase IntentionsIn Proc SOUPS 2017

[54] Florian Schaub Aditya Marella Pranshu Kalvani BlaseUr Chao Pan Emily Forney and Lorrie Faith CranorWatching Them Watching Me Browser Extensionsrsquo Im-pact on User Privacy Awareness and Concern In ProcUSEC 2016

[55] Ramya Sethuraman Why Am I Seeing This WeHave an Answer for You Facebook Blog March2019 httpsaboutfbcomnews201903why-am-i-seeing-this

[56] Matt Southern LinkedIn Now Lets MarketersTarget Ads to lsquoLookalike Audiencesrsquo SearchEngine Journal March 20 2019 httpswwwsearchenginejournalcomlinkedin-now-lets-marketers-target-ads-to-lookalike-audiences299547

[57] Till Speicher Muhammad Ali Giridhari Venkatadri Fil-ipe Ribeiro George Arvanitakis Fabriacutecio BenevenutoKrishna Gummadi Patrick Loiseau and Alan MislovePotential for Discrimination in Online Targeted Adver-tising In Proc FAT 2018

[58] Darren Stevenson Data Trust and Transparency inPersonalized Advertising PhD thesis University ofMichigan 2016

[59] Latanya Sweeney Discrimination in Online Ad Deliv-ery CACM 56(5)44ndash54 2013

[60] Michael Carl Tschantz Serge Egelman Jaeyoung ChoiNicholas Weaver and Gerald Friedland The Accuracyof the Demographic Inferences Shown on Googlersquos AdSettings In Proc WPES 2018

[61] Joseph Turow Jennifer King Chris Jay Hoofnagle AmyBleakley and Michael Hennessy Americans RejectTailored Advertising and Three Activities that Enable ItTechnical report Annenberg School for Communication2009

[62] Gerhard Tutz and Wolfgang Hennevogl Random effectsin ordinal regression models Computational Statisticsand Data Analysis 1996

[63] Twitter Ad Targeting Best Practices for Twitter TwitterBusiness 2019 httpsbusinesstwittercomentargetinghtml

[64] Twitter Healthcare Twitter Business 2019httpsbusinesstwittercomenhelpads-policiesrestricted-content-policieshealth-and-pharmaceutical-products-and-serviceshtml

[65] Twitter Keyword targeting Twitter Busi-ness 2019 httpsbusinesstwittercomenhelpcampaign-setupcampaign-targetingkeyword-targetinghtml

[66] Twitter Policies for conversion tracking andtailored audiences Twitter Business 2019httpsbusinesstwittercomenhelpads-policiesother-policy-requirementspolicies-for-conversion-tracking-and-tailored-audienceshtml

[67] Twitter Target based on how people accessTwitter Twitter Business 2019 httpsbusinesstwittercomentargetingdevice-targetinghtml

[68] Twitter Intro to Tailored Audiences TwitterBusiness 2020 httpsbusinesstwittercomenhelpcampaign-setupcampaign-targetingtailored-audienceshtml

[69] Twitter Twitter Privacy Policy httpstwittercomenprivacy 2020 Accessed February 13 2020

[70] Blase Ur Pedro Giovanni Leon Lorrie Faith CranorRichard Shay and Yang Wang Smart Useful ScaryCreepy Perceptions of Online Behavioral AdvertisingIn Proc SOUPS 2012


[71] Giridhari Venkatadri Athanasios Andreou Yabing LiuAlan Mislove Krishna Gummadi Patrick Loiseau andOana Goga Privacy Risks with Facebookrsquos PII-basedTargeting Auditing a Data Brokerrsquos Advertising Inter-face In Proc IEEE SampP 2018

[72] Giridhari Venkatadri Piotr Sapiezynski Elissa Red-miles Alan Mislove Oana Goga Michelle Mazurekand Krishna Gummadi Auditing Offline Data Brokersvia Facebookrsquos Advertising Platform In Proc WWW2019

[73] Jeffrey Warshaw Nina Taft and Allison Woodruff Intu-itions Analytics and Killing Ants Inference Literacyof High School-educated Adults in the US In ProcSOUPS 2016

[74] Ben Weinshel Miranda Wei Mainack Mondal EuirimChoi Shawn Shan Claire Dolin Michelle L Mazurekand Blase Ur Oh the Places Yoursquove Been User Reac-tions to Longitudinal Transparency About Third-PartyWeb Tracking and Inferencing In Proc CCS 2019

[75] Craig E Wills and Can Tatar Understanding What TheyDo with What They Know In Proc WPES 2012

[76] Yuxi Wu Panya Gupta Miranda Wei Yasemin AcarSascha Fahl and Blase Ur Your Secrets are Safe HowBrowersrsquo Explanations Impact Misconceptions AboutPrivate Browsing Mode In Proc WWW 2018

[77] Yaxing Yao Davide Lo Re and Yang Wang Folk Mod-els of Online Behavioral Advertising In Proc CSCW2017


What Twitter Knows Characterizing Ad Targeting Practices User Perceptionsand Ad Explanations Through Usersrsquo Own Twitter Data

Miranda Wei4 Madison Stamos Sophie Veys Nathan Reitinger Justin GoodmanMargot Herman Dorota Filipczukdagger Ben Weinshel Michelle L Mazurek Blase Ur

University of Chicago University of Maryland daggerUniversity of Southampton4University of Chicago and University of Washington

AbstractAlthough targeted advertising has drawn significant attentionfrom privacy researchers many critical empirical questionsremain In particular only a few of the dozens of targetingmechanisms used by major advertising platforms are wellunderstood and studies examining usersrsquo perceptions of adtargeting often rely on hypothetical situations Further it isunclear how well existing transparency mechanisms fromdata-access rights to ad explanations actually serve the usersthey are intended for To develop a deeper understanding ofthe current targeting advertising ecosystem this paper en-gages 231 participantsrsquo own Twitter data containing ads theywere shown and the associated targeting criteria for measure-ment and user study We find many targeting mechanismsignored by prior work mdash including advertiser-uploaded listsof specific users lookalike audiences and retargeting cam-paigns mdash are widely used on Twitter Crucially participantsfound these understudied practices among the most privacyinvasive Participants also found ad explanations designedfor this study more useful more comprehensible and overallmore preferable than Twitterrsquos current ad explanations Ourfindings underscore the benefits of data access characterizeunstudied facets of targeted advertising and identify potentialdirections for improving transparency in targeted advertising

1 Introduction

Social media companies derive a significant fraction of theirrevenue from advertising This advertising is typically highlytargeted drawing on data the company has collected aboutthe user either directly or indirectly Prior work suggests thatwhile users may find well-targeted ads useful they also findthem ldquocreepyrdquo [40 42 58 61 70] Further users sometimesfind targeted ads potentially embarrassing [3] and they may(justifiably) fear discrimination [4 15 21 47 57 59 60 73]In addition there are questions about the accuracy of cate-gorizations assigned to users [7 12 21 71 72] Above allusers currently have a limited understanding of the scope andmechanics of targeted advertising [17 22 48 50 54 70 77]

Many researchers have studied targeted advertising largelyfocusing on coarse demographic or interest-based target-ing However advertising platforms like Twitter [63] andGoogle [27] offer dozens of targeting mechanisms that arefar more precise and leverage data provided by users (egTwitter accounts followed) data inferred by the platform (egpotential future purchases) and data provided by advertisers(eg PII-indexed lists of current customers) Further becausethe detailed contents and provenance of information in usersrsquoadvertising profiles are rarely available prior work focusesheavily on abstract opinions about hypothetical scenarios

We leverage data subjectsrsquo right of access to data collectedabout them (recently strengthened by laws like GDPR andCCPA) to take a more comprehensive and ecologically validlook at targeted advertising Upon request Twitter will pro-vide a user with highly granular data about their accountincluding all ads displayed to the user in the last 90 daysalongside the criteria advertisers used to target those ads allinterests associated with that account and all advertisers whotargeted ads to that account

In this work we ask What are the discrete targeting mech-anisms offered to advertisers on Twitter and how are theyused to target Twitter users What do Twitter users think ofthese practices and existing transparency mechanisms A to-tal of 231 Twitter users downloaded their advertising-relateddata from Twitter shared it with us and completed an on-line user study incorporating this data Through this methodwe analyzed Twitterrsquos targeting ecosystem measured partici-pantsrsquo reactions to different types of ad targeting and ran asurvey-based experiment on potential ad explanations

We make three main contributions First we used our231 participantsrsquo files to characterize the current Twitter ad-targeting ecosystem Participants received ads targeted basedon 30 different targeting types or classes of attributes throughwhich advertisers can select an adrsquos recipients These typesranged from those commonly discussed in the literature (eginterests age gender) to others that have received far lessattention (eg audience lookalikes advertiser-uploaded listsof specific users and retargeting campaigns) Some partic-






2 Related Work
















3 Method

















































4 Results



































Purchase Events






















Wiese MEP


Financial Anker


Race Big Lots




Race Verizon










Tailored lists

Mobile

Tailored web

Behavior

Event

Lookalikes

Retargeting

MovieTV

Location

Gender

Keyword

Conversation

Platform

Interest

Age

Language

0 25 50 75 100

General Fair

0 25 50 75 100

General Comfortable

0 25 50 75 100

General Want

0 25 50 75 100

Specific Want

0 25 50 75 100

Specific Accurate

Stronglyagree











Response ρ p
















Control

Creepy

Detailed Text

Detailed Visual

Facebook

Twitter

0 25 50 75 100

Useful

0 25 50 75 100


0 25 50 75 100


0 25 50 75 100

Increased Trust

0 25 50 75 100

Want Similar

Stronglyagree













5 Discussion








Acknowledgments



References






















































































2 Related Work
















3 Method

















































4 Results



































Purchase Events






















Wiese MEP


Financial Anker


Race Big Lots




Race Verizon










Tailored lists

Mobile

Tailored web

Behavior

Event

Lookalikes

Retargeting

MovieTV

Location

Gender

Keyword

Conversation

Platform

Interest

Age

Language

0 25 50 75 100

General Fair

0 25 50 75 100

General Comfortable

0 25 50 75 100

General Want

0 25 50 75 100

Specific Want

0 25 50 75 100

Specific Accurate

Stronglyagree











Response ρ p
















Control

Creepy

Detailed Text

Detailed Visual

Facebook

Twitter

0 25 50 75 100

Useful

0 25 50 75 100


0 25 50 75 100


0 25 50 75 100

Increased Trust

0 25 50 75 100

Want Similar

Stronglyagree













5 Discussion








Acknowledgments



References

























































































3 Method

















































4 Results



































Purchase Events






















Wiese MEP


Financial Anker


Race Big Lots




Race Verizon










Tailored lists

Mobile

Tailored web

Behavior

Event

Lookalikes

Retargeting

MovieTV

Location

Gender

Keyword

Conversation

Platform

Interest

Age

Language

0 25 50 75 100

General Fair

0 25 50 75 100

General Comfortable

0 25 50 75 100

General Want

0 25 50 75 100

Specific Want

0 25 50 75 100

Specific Accurate

Stronglyagree











Response ρ p
















Control

Creepy

Detailed Text

Detailed Visual

Facebook

Twitter

0 25 50 75 100

Useful

0 25 50 75 100


0 25 50 75 100


0 25 50 75 100

Increased Trust

0 25 50 75 100

Want Similar

Stronglyagree













5 Discussion








Acknowledgments



References




























































































































4 Results



































Purchase Events






















Wiese MEP


Financial Anker


Race Big Lots




Race Verizon










Tailored lists

Mobile

Tailored web

Behavior

Event

Lookalikes

Retargeting

MovieTV

Location

Gender

Keyword

Conversation

Platform

Interest

Age

Language

0 25 50 75 100

General Fair

0 25 50 75 100

General Comfortable

0 25 50 75 100

General Want

0 25 50 75 100

Specific Want

0 25 50 75 100

Specific Accurate

Stronglyagree











Response ρ p
















Control

Creepy

Detailed Text

Detailed Visual

Facebook

Twitter

0 25 50 75 100

Useful

0 25 50 75 100


0 25 50 75 100


0 25 50 75 100

Increased Trust

0 25 50 75 100

Want Similar

Stronglyagree













5 Discussion








Acknowledgments



References
















































































































4 Results



































Purchase Events






















Wiese MEP


Financial Anker


Race Big Lots




Race Verizon










Tailored lists

Mobile

Tailored web

Behavior

Event

Lookalikes

Retargeting

MovieTV

Location

Gender

Keyword

Conversation

Platform

Interest

Age

Language

0 25 50 75 100

General Fair

0 25 50 75 100

General Comfortable

0 25 50 75 100

General Want

0 25 50 75 100

Specific Want

0 25 50 75 100

Specific Accurate

Stronglyagree











Response ρ p
















Control

Creepy

Detailed Text

Detailed Visual

Facebook

Twitter

0 25 50 75 100

Useful

0 25 50 75 100


0 25 50 75 100


0 25 50 75 100

Increased Trust

0 25 50 75 100

Want Similar

Stronglyagree













5 Discussion








Acknowledgments



References

































































































4 Results



































Purchase Events






















Wiese MEP


Financial Anker


Race Big Lots




Race Verizon










Tailored lists

Mobile

Tailored web

Behavior

Event

Lookalikes

Retargeting

MovieTV

Location

Gender

Keyword

Conversation

Platform

Interest

Age

Language

0 25 50 75 100

General Fair

0 25 50 75 100

General Comfortable

0 25 50 75 100

General Want

0 25 50 75 100

Specific Want

0 25 50 75 100

Specific Accurate

Stronglyagree











Response ρ p
















Control

Creepy

Detailed Text

Detailed Visual

Facebook

Twitter

0 25 50 75 100

Useful

0 25 50 75 100


0 25 50 75 100


0 25 50 75 100

Increased Trust

0 25 50 75 100

Want Similar

Stronglyagree













5 Discussion








Acknowledgments



References




















































































4 Results



































Purchase Events






















Wiese MEP


Financial Anker


Race Big Lots




Race Verizon










Tailored lists

Mobile

Tailored web

Behavior

Event

Lookalikes

Retargeting

MovieTV

Location

Gender

Keyword

Conversation

Platform

Interest

Age

Language

0 25 50 75 100

General Fair

0 25 50 75 100

General Comfortable

0 25 50 75 100

General Want

0 25 50 75 100

Specific Want

0 25 50 75 100

Specific Accurate

Stronglyagree











Response ρ p
















Control

Creepy

Detailed Text

Detailed Visual

Facebook

Twitter

0 25 50 75 100

Useful

0 25 50 75 100


0 25 50 75 100


0 25 50 75 100

Increased Trust

0 25 50 75 100

Want Similar

Stronglyagree













5 Discussion








Acknowledgments



References












































































































Purchase Events






















Wiese MEP


Financial Anker


Race Big Lots




Race Verizon










Tailored lists

Mobile

Tailored web

Behavior

Event

Lookalikes

Retargeting

MovieTV

Location

Gender

Keyword

Conversation

Platform

Interest

Age

Language

0 25 50 75 100

General Fair

0 25 50 75 100

General Comfortable

0 25 50 75 100

General Want

0 25 50 75 100

Specific Want

0 25 50 75 100

Specific Accurate

Stronglyagree











Response ρ p
















Control

Creepy

Detailed Text

Detailed Visual

Facebook

Twitter

0 25 50 75 100

Useful

0 25 50 75 100


0 25 50 75 100


0 25 50 75 100

Increased Trust

0 25 50 75 100

Want Similar

Stronglyagree













5 Discussion








Acknowledgments



References
































































































Purchase Events






















Wiese MEP


Financial Anker


Race Big Lots




Race Verizon










Tailored lists

Mobile

Tailored web

Behavior

Event

Lookalikes

Retargeting

MovieTV

Location

Gender

Keyword

Conversation

Platform

Interest

Age

Language

0 25 50 75 100

General Fair

0 25 50 75 100

General Comfortable

0 25 50 75 100

General Want

0 25 50 75 100

Specific Want

0 25 50 75 100

Specific Accurate

Stronglyagree











Response ρ p
















Control

Creepy

Detailed Text

Detailed Visual

Facebook

Twitter

0 25 50 75 100

Useful

0 25 50 75 100


0 25 50 75 100


0 25 50 75 100

Increased Trust

0 25 50 75 100

Want Similar

Stronglyagree













5 Discussion








Acknowledgments



References



























































































Wiese MEP


Financial Anker


Race Big Lots




Race Verizon










Tailored lists

Mobile

Tailored web

Behavior

Event

Lookalikes

Retargeting

MovieTV

Location

Gender

Keyword

Conversation

Platform

Interest

Age

Language

0 25 50 75 100

General Fair

0 25 50 75 100

General Comfortable

0 25 50 75 100

General Want

0 25 50 75 100

Specific Want

0 25 50 75 100

Specific Accurate

Stronglyagree











Response ρ p
















Control

Creepy

Detailed Text

Detailed Visual

Facebook

Twitter

0 25 50 75 100

Useful

0 25 50 75 100


0 25 50 75 100


0 25 50 75 100

Increased Trust

0 25 50 75 100

Want Similar

Stronglyagree













5 Discussion








Acknowledgments



References


















































































Tailored lists

Mobile

Tailored web

Behavior

Event

Lookalikes

Retargeting

MovieTV

Location

Gender

Keyword

Conversation

Platform

Interest

Age

Language

0 25 50 75 100

General Fair

0 25 50 75 100

General Comfortable

0 25 50 75 100

General Want

0 25 50 75 100

Specific Want

0 25 50 75 100

Specific Accurate

Stronglyagree











Response ρ p
















Control

Creepy

Detailed Text

Detailed Visual

Facebook

Twitter

0 25 50 75 100

Useful

0 25 50 75 100


0 25 50 75 100


0 25 50 75 100

Increased Trust

0 25 50 75 100

Want Similar

Stronglyagree













5 Discussion








Acknowledgments



References


















































































Response ρ p
















Control

Creepy

Detailed Text

Detailed Visual

Facebook

Twitter

0 25 50 75 100

Useful

0 25 50 75 100


0 25 50 75 100


0 25 50 75 100

Increased Trust

0 25 50 75 100

Want Similar

Stronglyagree













5 Discussion








Acknowledgments



References


















































































Control

Creepy

Detailed Text

Detailed Visual

Facebook

Twitter

0 25 50 75 100

Useful

0 25 50 75 100


0 25 50 75 100


0 25 50 75 100

Increased Trust

0 25 50 75 100

Want Similar

Stronglyagree













5 Discussion








Acknowledgments



References


















































































5 Discussion








Acknowledgments



References


















































































References





















































































































































































Documents

What Twitter Knows: Characterizing Ad Targeting Practices ... · 978-1-939133-17-5 Open access to the Proceedings of the 29th USENIX Security Symposium is sponsored by USENIX. What