PaperlayerView exact revision on arXiv

Source-linked AI summary

From Scroll to Sale: Exploring the Impact of Interaction Type and Device Price on TikTok Advertisements

Nazanin Sabri, Cat Mai, Haodi Zou, Isha Varada, Damon McCoy, Deepak Kumar, Kristen Vaccaro

arXiv:2610.00684v1cs.CYcs.HCcs.SI

TL;DR

Evidence on income-based ad targeting and advertising on short-form video platforms is limited, despite concerns about how dynamic pricing may affect low-income shoppers. The authors audit TikTok with paired automated-account studies, varying interaction type, device price, age, and gender. They find greater ad exposure with liking and sharing, and lower-priced devices receive larger discounts while the most expensive device has the lowest ad load.

  • Problem

    Research on targeted advertising has focused mainly on demographics, and short-form video platforms have rarely been the subject of advertising audits, leaving limited evidence on income-based targeting there.

    Source: §1 · p. 1

  • Method

    Two paired automated-account studies varied interaction type and device price as an income proxy, along with age and gender, to audit TikTok ad frequency and content.

    Source: §1 · p. 1, §8 · p. 9

  • Results

    Liking and sharing significantly increased ad exposure, commenting did not increase ad load, and lower-income accounts received higher discount percentages while the least expensive device received more ads.

    Source: §8 · p. 9

  • Takeaways & Limitations

    Users who cannot or do not want to purchase expensive devices may face a higher advertisement burden, while interactions may offer everyday users more power over what they see.

    Source: §1 · p. 2

    The absence of observed gender differences may reflect TikTok’s not using names as a gender proxy; the platform may still infer gender from other signals.

    Source: §6.1 · p. 9

Abstract

from arXiv · show

Companies and brands increasingly use dynamic pricing, including targeting social media ads to users based on their income. In this work we audit TikTok's feed using 56 automated accounts, which collect data on over 80,000 videos, across two studies. We test the impact of device price on ad load and ad types, using 12 phones of low ($0-$250), medium ($400-$650), and high ($750-$1,000+) price as a proxy for income. We also test the impact of interaction type (i.e., like, comment, share), age, and gender on the frequency and content of ads. Overall, the ad load was 29.4%, but liking and sharing content increased ad load significantly. We also found that as accounts spend more time on TikTok, the ad load steadily increases. We found some evidence that device price impacts both ad load and content -- more expensive devices were targeted with fewer ads, while the least expensive devices received more discounts.

1 Introduction

The paper examines whether TikTok ad delivery varies with interaction, device price as an income proxy, age, and gender. Across its two studies, ad exposure rose with liking and sharing, while lower-priced devices received larger discounts and the most expensive device had the lowest ad load.

1 Introduction

  • The audit addresses limited evidence on income-based ad targeting and on advertising audits of short-form video platforms.Prior TikTok ad-targeting work examined age and gender; this study measures income-proxy effects on ad volume, content, and prices.Source: §1 · p. 1
  • The study asks how scrolling, liking, commenting, and sharing affect ad number and content, and how income, age, and gender affect ad number and content.Device price is experimentally varied as a proxy for income.Source: §1 · p. 1
  • Liking and sharing increased ad load to as much as 37.6% of videos, while commenting did not produce the same increase.Study 1 collected 42,819 videos using 32 automated accounts; overall ad load was 29.4%.Source: §1 · pp. 1–2
  • The most expensive phone had the lowest ad load, while the least expensive device received the largest discounts.Study 2 tested 12 phones priced from $80 to $1,000, along with age and gender conditions.Source: §1 · p. 2
  • The authors frame differential ad delivery as a potential burden for shoppers with less ability or desire to buy expensive devices, while interactions may give users more influence over what they see.The motivation notes that dynamic pricing could leave low-income shoppers facing higher prices or lower-quality goods.Source: §1 · pp. 1–2

2 Related Work

Prior research has audited algorithmic feeds and ad delivery, but TikTok studies have mainly focused on recommendations and other platform features. This paper extends interaction analysis to advertising and examines device price and behavior together.

2 Related Work

  • Ad-delivery audits find that optimization can make an ad’s final audience differ from the advertiser’s intended audience.The cited studies include Lambrecht and Tucker 2019, Asplund et al. 2020, and Ali et al. 2023.Source: Audits of Ad Delivery · p. 2
  • Prior work links ad targeting to demographics and user value, but had not examined device price and behavior together in TikTok advertising.Related findings include higher-priced hotel rankings for Mac OS users and higher ad load for less-private users.Source: Audits of Ad Delivery · p. 2
  • TikTok audits have used sockpuppet accounts, data donations, and network traffic analysis, with most focusing on recommendations rather than advertising.Other audited features include autocomplete search suggestions and the accuracy of explanations for recommended videos.Source: Audits of TikTok · p. 2
  • Prior TikTok work found age-related differences in ads shown to adult and child accounts, alongside evidence of limited safeguards against youth exposure to non-child-directed content.The cited work reports that 83% of videos shown on a child’s recommendation feed were not child-directed.Source: Audits of TikTok · p. 2
  • Earlier work examined how interaction type shapes content recommendations; this study extends that analysis to advertisements and adds further interaction types.Boeker and Urman (2022) found follow, like, and watch rate to be the most influential interactions for recommendations.Source: Audits of TikTok · p. 2

3 Experimental Design

The authors ran two paired TikTok audits that varied interaction behavior, device price, age, and name-proxied gender, then recorded and analyzed feed content. Their design used duplicated accounts and concurrent collection to reduce randomness and time-varying campaign effects.

3 Experimental Design

  • Across two studies, the audit varied interaction behavior and income-proxy and demographic conditions, collecting 80,169 videos, including 24,255 advertisements.Data collection ran from March 12 to April 25, 2026.Source: §3 · p. 3

3.1 Independent Variables

  • Study 1 isolated scroll, like, comment, and share conditions, while Study 2 combined interactions using a more realistic like/comment/share mix.Study 2 used 93% likes, 5% comments, and 2% shares, based on prior data-donation statistics.Source: Interaction Behaviors · pp. 3–4
  • Device price represented income across three price bands, ages 26 and 46 represented younger and older adults, and account names served as gender proxies.The price bands were $0–$250, $400–$650, and $750–$1,000+; the accounts used Jacob/Emily and Michael/Jennifer.Source: Income · p. 4

3.2 Dependent Variables

  • The analysis measured ad frequency and content using screenshots; it excluded TikTok’s ad explanations because of prior accuracy concerns and the share-menu access path.Accounts ran for 3–4 days or until at least 1,000 screenshots had been collected.Source: §3.2 · p. 4, System Outputs · p. 4
  • Ad detection combined red-button image processing with OCR keyword matching, and the method classified 24,255 of 80,169 screenshots as ads.It achieved 99.6% accuracy on a balanced sample and 100% on a random screenshot sample.Source: System Outputs · pp. 4–5
  • Ads were assigned to one of 49 categories using detected logos or image labels, while price extraction identified explicit dollar amounts and percentages for manual review.Ads without logos or labels were classified as “Unknown.”Source: Advertisement Content Categorization · p. 5, Price Detection · p. 5

3.3 Summary of Experimental Conditions

  • Study 1 used 32 accounts and Study 2 used 24; each setting had two copies, with data collected simultaneously where possible to control randomness and campaign timing.Age and gender effects were analyzed using Study 2 because Study 1’s interaction behaviors were artificial.Source: Study 1 · p. 5, Study 2 · p. 5

3.4 Auditing System Implementation

  • The audit automated TikTok interactions on Android devices, reset app storage and advertising IDs between accounts, and kept the collected dataset request-access only because it could contain identifiable user information.Account creation and login remained manual because TikTok bot detection made automation unreliable.Source: Software · p. 5, Hardware · p. 5, Switching Between Accounts · p. 5

4 RQ1: How does interaction type affect the number and content of ads?

Liking and sharing were associated with higher ad loads than scrolling, while ad load also rose over account lifetimes and with interactions with ads. Commenting showed a later, slower increase.

4.1 Liking and sharing increased ad load

  • 35.2% of videos in the share condition were ads, versus 30.8% for like, 25.7% for scroll, and 25.9% for comment.Source: §4.1 · p. 6, Table 2 · p. 7
  • 88.5% of ad-count variance was explained by the model; liking and sharing significantly increased counts, while commenting’s positive association was nonsignificant, including with demographic controls.Source: §4.1 · p. 6, Table 3 · p. 7

4.2 As accounts viewed more content, they were shown more advertisements

  • Ad load rose over accounts’ first 1,000 videos across interaction types, but commenting’s increase was delayed and slower.Source: §4.2 · p. 6, Figure 1 · p. 6
  • The authors offer two interpretations: sustained scrolling may signal engagement, or TikTok may pace early sessions lightly before monetizing more aggressively.Source: §4.2 · p. 6

4.3 More interactions with ads increased ad load

  • Ad load correlated with the number of ads interacted with (Pearson r=+0.759, p<0.001), suggesting both interaction type and target mattered.Source: §4.3 · p. 6, Figure 2 · p. 6

5 RQ2: How does income affect the number and content of ads?

Device price showed a directional relationship with ad load, while age was the only significant demographic predictor in the regression. Ad categories were largely similar across devices, but interactions with non-ad content better predicted category-specific ads; the cheapest phone received the highest discounts.

5.1 More expensive devices received fewer ads

  • 33.9% of ads appeared on the cheapest device, compared with 31.5% on the midpriced device and 29.1% on the most expensive device.The results support a directional relationship between device price and ad exposure.Source: §5.1 · p. 7
  • Device price had a negative but nonsignificant relationship with ad count, while younger profiles received approximately 89 more ads per account, a 5% increase in ads.Age was the only statistically significant demographic variable in the OLS model.Source: §5.1 · p. 7, Table 5 · p. 7
  • Travel ads appeared less often on the cheapest Pixel 3a, while pairwise tests found no significant differences in category rankings across devices.Entertainment, Electronics, and food were among the most frequent categories; healthcare ads may have been missed because ambiguous visual categories were removed.Source: §5.2 · p. 8, Figure 3 · p. 8
  • Interactions with non-ad content predicted ad categories more strongly than device price, although some categories, including Restaurants, Furniture, and Arts, had few or no matching interactions.The ad-category and interaction-category rankings differed significantly (W=299.00, p<0.001); further study is needed to test whether some sectors advertise to uninterested users.Source: §5.2 · p. 8, Figure 4 · p. 8

5.3 The cheapest phone was shown the highest discount percentages

  • 14.4% of Study 2 ads listed exact prices; the cheapest phone received the highest discount percentages, while the mid-range phone saw the highest-cost ads.Discount percentages do not necessarily mean cheaper deals, because many discount ads omit product costs.Source: §5.3 · p. 8, Table 6 · p. 8

6 RQ3: How do age and gender affect the number and content of ads?

Younger accounts received more ads, and age was the only statistically significant demographic predictor of ad count. Ad-category rankings did not differ significantly by age or gender.

6.1 Younger accounts received more ads

  • Younger users and men received more ads, but age was the only statistically significant predictor of ad count.The reported pattern for age was consistent with prior work on age differences in content recommendation; gender differences were not significant there.Source: §6.1 · p. 9

6.2 Ad categories were consistent across age and gender

  • Ad-category rankings showed no statistically significant differences by either age or gender.Source: §6.2 · p. 9

7 Discussion

Liking and sharing were associated with substantially more ad exposure, giving users some influence over the ads they see but potentially discouraging engagement with favored content. Device price had a directional relationship with ad load, though the authors characterize it as a weak and ambiguous signal.

7.1 A focus on behavior rather than demographics

  • Likes and shares substantially increased ad exposure, giving users some influence over ads but potentially making them wary of supporting content they enjoy.The authors connect this concern to users’ fear of receiving excessive volumes of similar content.Source: §7.1 · p. 9
  • With interaction held at 10%, ad percentages rose over time; the authors suggest longer use or feedback from interacting with ads as possible explanations.They call for testing different engagement patterns with ads and non-ad content, as well as behavioral changes over time.Source: §7.1 · p. 9

7.2 Device price as a weak and ambiguous signal

  • The cheapest device had the highest ad load, which may burden users unable or unwilling to buy pricier devices; observed differences were small, and device price may be an ambiguous income proxy.The authors call for further study of income proxies, including features such as Spending Power.Source: §7.2 · p. 9

8 Conclusion

The conclusion reports that active liking and sharing significantly increased ad exposure, while commenting did not; the cheapest device showed a directional increase in ad load and lower-income accounts saw higher discount percentages. The supplied appendix passages describe TikTok’s advertising platform and the researchers’ ad-identification process.

8 Conclusion

  • Across two studies, 56 accounts collected 80,169 videos and 24,255 ads; liking and sharing significantly increased ad exposure, commenting did not, and the cheapest device had a directional increase in ad load.Lower-income accounts were shown higher discount percentages.Source: §8 · p. 9

A TikTok Advertisement Platform

  • TikTok’s advertising platform lets advertisers select user characteristics for targeting and estimates the audience eligible to see an ad based on those criteria and recent ad delivery.The Available audience estimate refers to accounts shown ads on TikTok in the past 30 days.Source: A TikTok Advertisement Platform · p. 12

B.1 What do ads look like?

  • Ads may be signaled by labels, dollar amounts, percentage discounts, or a red call-to-action button, but those cues are not always present or mutually exclusive.A location icon alone does not necessarily indicate an ad, and a button can remain visible after other ad tags disappear.Source: §B.1 · p. 12
  • The identification method was designed to exclude videos that are ads but not officially classified as ads by TikTok.Source: §B.1 · pp. 12–13

B.2 B.2 Identification Process

  • The ad-detection process used image and OCR outputs, first checking for a red box and then matching keywords if none was found; it identified 24,255 ads among 80,169 screenshots.Figure 9 reports the tag or method used to identify each ad, counting each ad once even if it had multiple tags.Source: B Ad Detection · p. 12, §B.2 · p. 13, Figure 9 · p. 14

B.3 Performance Evaluation

Ad-identification performance was checked against manually verified samples, and the evaluation describes a false positive involving “Shop.” The following logo analysis used automated logo detection on identified ads.

B.3 Performance Evaluation

C Analyzing Products Through Logo Detection

C.1 Logos Removed From our Analysis

The analysis excluded TiVo after logo detection repeatedly mistook another logo for it.

C.1 Logos Removed From our Analysis

  • TiVo was removed from the analysis after it was detected in 5,244 ads, or 45.0% of them.The authors hypothesize that the “Live” logo caused the false detections because it resembles the TiVo logo.Source: §C.1 · p. 15
  • The authors attribute the mistaken TiVo detections to the visually similar “Live” logo.This is presented as a hypothesis, not a confirmed explanation.Source: §C.1 · p. 15
  • The detected TiVo logos were excluded from subsequent analysis.The exclusion followed the high rate of mistaken detections.Source: §C.1 · p. 15

C.2 Statistics of Logo Detection Across Study 2

The study categorized ads using detected logos and image labels, then assigned each ad a single category. It also reports how content was interacted with under the two study conditions.

C.2 Statistics of Logo Detection Across Study 2

  • The reviewed taxonomy contained 49 categories, including “Sports” and “Nonprofit & Advocacy.”“Labor & Unions” was merged into “Nonprofit & Advocacy,” and three sports-related categories were merged into “Sports.”Source: §C.2 · p. 16

C.3 Assigning Categories to Images

  • Each ad received the most frequent category among its detected logos, with ties broken randomly.An ad with multiple detected logos could therefore receive one category based on the frequency of those logos.Source: §C.3 · p. 16, Figure 12 · p. 16

D Categorizing Image Labels

  • Image labels were extracted from ad screenshots after cropping the top 9% and bottom 20%, then mapped to the 49-category taxonomy plus “Others.”The 2,866 unique labels were mapped, and a researcher reviewed labels assigned to “Others.”Source: D Categorizing Image Labels · pp. 16–17

E.1 Single Interaction Condition

  • Study 1’s single-interaction condition is summarized by counts of ad and non-ad content interacted with.Table 14 presents those counts.Source: §E.1 · p. 17

E.2 Multiple Interaction Condition

  • In Study 2, a combination of interactions within each account programmed any interaction at 10% probability and sharing, conditional on interaction, at 2%; sharing never occurred.Table 13 displays the number of content interacted with.Source: §E.2 · p. 17

F Study 1: Regression Model Including Age & Gender

The Study 1 regression included age and gender alongside interaction types and videos watched. Liking and sharing remained significantly associated with increased ad count.

F Study 1: Regression Model Including Age & Gender

  • Liking and sharing remained statistically significant predictors of increased ad count when age, gender, and total videos watched were included.Commenting had a positive but non-significant relationship with ad count.Source: & Gender · p. 17, Table 15 · p. 18
  • The model coded gender as 1 for the Jacob and Michael accounts and 0 for the Emily and Jennifer accounts, with age 26 as the reference group.The passage identifies these as the regression’s gender coding and age reference.Source: & Gender · p. 17
  • Tables 14 and 15 respectively report Study 1 interaction counts and the regression predicting ad count.Table 15’s caption notes that commenting was positive but non-significant.Source: Table 14 · p. 18, Table 15 · p. 18
Loading 2610.00684v1…