REVIEW
The Beginning of ChatGPT Ads
Computers and Society · Submitted 5 Aug 2026 · 35 references
Pith's verdict: First useful baseline of ChatGPT ads, but the headline income gradient doesn't survive contact with the study's own timing data.
T0 · AI review
Why Pith says this
Highlight any text to comment on it.
Referee report
The paper
The paper claims
ChatGPT’s first ad rollout concentrates ads on lower-income accounts.
The reading
What carries the argument
The mechanism that carries the argument is a demographically stratified sock puppet audit. The researchers created 91 synthetic ChatGPT accounts, each bound to a U.S. ZIP code selected from census data to represent one of nine cells formed by crossing three racial/ethnic groups (Black, Hispanic, White) with three income terciles (low, medium, high). All traffic for an account was routed through a residential proxy matched to that ZIP code, and each chat session began with location-signaling prompts. Every day each account received the same random 30-prompt subset drawn from a 335-prompt corpus, and the full HTML of every response was captured so that ads could be identified by the “Sponsored” label. This pipeline converts demographic signals into observable differences in ad delivery and is designed to be replicable on other platforms or later time periods.
Extended reading notes
Core claim
The paper’s central empirical discovery is that in the first weeks of ChatGPT’s ad rollout, ad exposure was concentrated on lower-income accounts. A logistic regression of ad exposure on median household income of the assigned ZIP code yielded an odds ratio of 0.98 per $1,000 increase (95% CI [0.96, 1.00], p = 0.0438), and the association strengthened when three high-leverage accounts were removed (OR = 0.975, p = 0.0053). No significant association between race and ad delivery was found, though the authors caution that with 16–19 exposed accounts per racial group the study is underpowered to detect small effects. The paper also documents that ads began appearing on March 8, a median of 14 days after account activation, and that among exposed accounts ads appeared in a median of 24% of prompts after the first ad. In this early phase, ads were labeled “Sponsored”, visually separated from the model’s text, and concentrated in retail and information sectors, with Target, Top10.com, Preply, and Advance Auto Parts among the most frequent advertisers.
Load-bearing premise
The income gradient in ad exposure is only meaningful if it reflects income-based targeting rather than the study’s own timing—specifically, accounts were created in batches from late January to mid-March, ads began on March 8, and the analysis does not adjust for account creation date or days active during the ad window.
Editorial extensions
If this is right
- If ad delivery concentrates on lower-income accounts, the new LLM ad channel begins with a distributional pattern that places a higher advertising burden on users with less income.
- OpenAI’s stated exclusions for health, mental health, and political content appear to be working in this early window, since those prompt categories show near-zero ad rates.
- The current ad format is clearly labeled and separate from model responses, giving researchers a stable baseline for tracking how disclosure and integration evolve.
- Among exposed accounts, ad delivery is persistent rather than sporadic, with a median of 24% of prompts producing an ad once delivery begins.
- The methodology of ZIP-stratified sock puppets can be reused to audit future LLM ad systems and other closed platforms.
Reading between the lines
- If the income gradient persists as ad targeting matures, ChatGPT advertising could create a regressive pattern in which lower-income users bear more commercial messages—a dynamic the paper’s three-week window cannot yet confirm.
- The consistent 14-day gap between account creation and first ad suggests a warm-up or anti-abuse mechanism; future audits of LLM ads should explicitly model this lag rather than treating it as noise.
- The released ad archive makes a natural longitudinal follow-up possible: re-running the same audit after the platform’s May 2026 international pilot would show whether the income gradient and content exclusions hold across markets.
Original abstract
This paper presents the first empirical study of advertising content being rolled out in the user-facing online interfaces of large language models (LLMs). We systematically examine possible demographic differences in ad content shown to U.S. users of ChatGPT using a sock puppet audit methodology. We create and deploy 91 sock puppets in a 3x3 factorial design, using geolocation cues (account IP proxies and location-signaling prompts) to signal three racial/ethnic groups (Black, Hispanic, and White) and three income terciles (low, medium, and high). We conduct data collection starting in February 2026, collecting over 3,000 advertisements from 186 unique advertisers in response to 335 prompts on a range of realistic user queries. We find that accounts begin receiving ads 14 days after account creation, and that lower-income accounts, regardless of race, are more likely to receive ads. In this first phase of ChatGPT ads, the ads themselves skewed heavily towards consumer goods, directed users to a specific advertiser rather than a particular product, and were clearly separated from the LLM's response text, observations we anticipate will change as ads continue being integrated into LLM chat interfaces. We release a public, searchable archive of all collected advertisements. Finally, we discuss the implications of our findings, and conclude with methodological and theoretical recommendations for future empirical studies of LLM advertisements.
The compiled record
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.
The abstract, sentence by sentence
This paper presents the first empirical study of advertising content being rolled out in the user-facing online interfaces of large language models (LLMs). We systematically examine possible demographic differences in ad content shown to U.S. users of ChatGPT using a sock puppet audit methodology. We create and deploy 91 sock puppets in a 3x3 factorial design, using geolocation cues (account IP proxies and location-signaling prompts) to signal three racial/ethnic groups (Black, Hispanic, and White) and three income terciles (low, medium, and high). We conduct data collection starting in February 2026, collecting over 3,000 advertisements from 186 unique advertisers in response to 335 prompts on a range of realistic user queries. We find that accounts begin receiving ads 14 days after account creation, and that lower-income accounts, regardless of race, are more likely to receive ads. In this first phase of ChatGPT ads, the ads themselves skewed heavily towards consumer goods, directed users to a specific advertiser rather than a particular product, and were clearly separated from the LLM's response text, observations we anticipate will change as ads continue being integrated into LLM chat interfaces. We release a public, searchable archive of all collected advertisements. Finally, we discuss the implications of our findings, and conclude with methodological and theoretical recommendations for future empirical studies of LLM advertisements.
background contribution method result interpretation main claim
| # | sentence | role | the authors' own work | states the main conclusion | strength |
|---|---|---|---|---|---|
| s0 | This paper presents the first empirical study of advertising content being rolled out in the user-facing online interfaces of large language models (LLMs). | contribution 1.00 | 0.83 | 0.04 | asserted 1.00 |
| s1 | We systematically examine possible demographic differences in ad content shown to U.S. users of ChatGPT using a sock puppet audit methodology. | method 0.67 | 0.93 | 0.03 | asserted 0.83 |
| s2 | We create and deploy 91 sock puppets in a 3x3 factorial design, using geolocation cues (account IP proxies and location-signaling prompts) to signal three racial/ethnic groups (Black, Hispanic, and White) and three income terciles (low, medium, and high). | method 1.00 | 0.97 | 0.02 | asserted 0.98 |
| s3 | We conduct data collection starting in February 2026, collecting over 3,000 advertisements from 186 unique advertisers in response to 335 prompts on a range of realistic user queries. | method 1.00 | 0.97 | 0.02 | asserted 1.00 |
| s4 | We find that accounts begin receiving ads 14 days after account creation, and that lower-income accounts, regardless of race, are more likely to receive ads. | result 1.00 | 0.98 | 0.64 | asserted 1.00 |
| s5 | In this first phase of ChatGPT ads, the ads themselves skewed heavily towards consumer goods, directed users to a specific advertiser rather than a particular product, and were clearly separated from the LLM's response text, observations we anticipate will change as ads continue being integrated into LLM chat interfaces. | result 0.87 | 0.91 | 0.12 | asserted 0.78 |
| s6 | We release a public, searchable archive of all collected advertisements. | contribution 0.98 | 0.93 | 0.02 | asserted 0.98 |
| s7 | Finally, we discuss the implications of our findings, and conclude with methodological and theoretical recommendations for future empirical studies of LLM advertisements. | future work 1.00 | 0.51 | 0.06 | unclear 0.54 |
What was examined, and what was not
- arXiv listing abstract v1
- full text
- figures and tables
- references and citing sentences
- relations to other papers
- P abstract
- S sections
- T body sentences
- C citations
- R cross-paper relations
Figures
Showing 10 of the paper's 12 figures. See all figures in the PDF ↗ Figures reproduced from arXiv:2608.05008 by the authors.
References
How the literature uses it
What this paper does with the arXiv papers it cites
Discussion
Disagree with the review, or know something it missed? Leave a comment, highlight any passage above to comment on it, or challenge this review formally.