Four academics built 91 fake ChatGPT accounts, ran 127,801 conversations through them between February and May 2026, and captured 3,602 ads. “The Beginning of ChatGPT Ads,” posted to arXiv on August 5, 2026 by Emma Lurie, Ro Encarnación, and Danaé Metaxa at the University of Pennsylvania with Sorelle Friedler of Haverford College, is the first independent read on advertising inside an LLM interface. The accounts ran identical prompts and differed only in the income and racial profile of the ZIP codes they were proxied through. Any difference in what came back – by study design – should point at the platform’s targeting capabilities.
No account saw an ad in its first seven days. The median gap from activation to first ad was 14 days, with an interquartile range of 13 to 16 across 53 exposed accounts. Once ads started they kept coming: 5.61% of all interactions carried one, but measured per exposed account from its first ad onward the median rate was 24% of prompts, and the heaviest-served account hit 37.1%. The daily prompt set deliberately re-ran queries that had already produced ads, so the corpus overrepresents ad-friendly prompts by design.
Exposure fell as income rose. The odds of receiving any ad at all dropped roughly 2% per $1,000 of ZIP-code median income. Frequency once ads had begun showed no income pattern, and the study cannot say why the exposure gradient exists. It found no delivery difference across the Black, Hispanic and White account groups, and cautions that with 16 to 19 exposed accounts per group it could only have caught a large effect.
Retail Trade and Information sectors took roughly 57% of the impressions, spread across 191 advertisers. Target was the most frequent at 130 sessions, with Best Buy, The Home Depot, Peloton, HelloFresh, DoorDash and Disney among the top ten, a more mainstream retail picture than Similarweb’s May panel data, which skewed to HubSpot, Fiverr and Cursor. Every ad is now in a public, searchable library with advertiser, prompt text and screenshots attached. The window was three weeks.
Read more at LinkedIn or find the source paper here.