Digital Advertising Operations
The First Big ChatGPT Ads Audit Found an Income Gap
A new independent study gives ad operations teams an early baseline for delivery fairness, disclosure, context, and brand safety.
The August 5 preprint found lower-income simulated accounts were more likely to receive ads. Treat the result as an important early signal—not a final verdict—and demand ongoing delivery, exclusion, and disclosure audits.
Researchers created 91 simulated U.S. accounts across income and race-associated location signals, then ran a large browser-based audit of early ChatGPT advertising. The work covered 127,801 conversations and thousands of confirmed impressions. Ads appeared as clearly separated Sponsored units beneath responses rather than secretly inside the answer.
The setting matters. Conversational systems can feel helpful, personal, and unusually persuasive. They also receive sensitive context that a search box may never see. An ad below a recipe is familiar; an ad beside advice about health, money, or relationships raises a different class of operational questions—even when the label is wearing its best shirt.
Lower-income accounts were significantly more likely to receive advertisements. The researchers did not find a significant association with race, but explicitly cautioned that the sample was too small for strong conclusions on racial differences. Retail and software represented roughly 57% of impressions, and product, fitness, beauty, and cooking prompts drew higher ad rates.
The study also found delivery near zero for categories OpenAI said it would initially exclude, including politics, mental health, and medical conditions. That is encouraging, but it is a snapshot of an early rollout using simulated behavior—not proof that every user, language, market, or future model will behave identically.