The first systematic empirical audit of advertising in ChatGPT found that synthetic accounts signaling lower income were more likely to receive ads than higher-income accounts, regardless of the racial or ethnic identity signaled by the researchers.
The study, The Beginning of ChatGPT Ads, was submitted to arXiv on August 5, 2026. It examined 91 synthetic U.S. accounts and collected more than 3,000 advertisements from 186 unique advertisers across 335 realistic prompts.
91 synthetic accounts tested demographic differences
The researchers used a sock-puppet audit methodology with a 3×3 factorial design. Accounts were configured using geolocation cues and prompts to signal three racial or ethnic groups — Black, Hispanic and White — and three income terciles: low, medium and high.
Data collection began in February 2026, allowing the researchers to compare advertising exposure across controlled account profiles rather than relying on self-reported user experiences.
Lower-income accounts were more likely to receive ads
The paper's abstract reports that lower-income accounts were more likely to receive advertisements regardless of race. This is an observed disparity in the audit; it does not establish that OpenAI intentionally targets lower-income users or that income is a declared advertising-targeting criterion.
That distinction is important because an external audit can measure outcomes but cannot, by itself, identify the internal mechanism or intent that produced them.
Ads generally appeared after about 14 days
The researchers also found that accounts began receiving ads 14 days after account creation. The advertisements collected during this initial phase skewed heavily toward consumer goods and generally directed users to a specific advertiser rather than to a particular product.
The paper additionally notes that the advertisements were clearly separated from ChatGPT's response text during the observed period.
Why the audit matters for AI advertising
Advertising inside conversational AI creates a measurement environment different from traditional search or social advertising. The interface is personalized, conversational and potentially shaped by accumulated account context, making systematic external auditing especially important.
If different synthetic personas experience different probabilities of receiving commercial content, researchers need to distinguish between the content of ads, the advertisers selected and the probability that an ad appears at all.
The NetContentSEO view
The study introduces another visibility layer for AI platforms: commercial exposure itself may vary across user profiles. For marketers, publishers and AI visibility researchers, measuring only which brands appear will not be enough. The probability of an advertisement being shown — and how that probability varies across controlled personas — becomes part of the visibility model.
At the same time, the findings should be described carefully. This is an independent academic audit and not evidence of a publicly stated OpenAI policy to target advertising based on income. What the researchers established is a measurable difference in observed exposure among their synthetic accounts.