Advertising inside an AI assistant is supposed to benefit from conversational context. A new analysis suggests that promise remains unreliable: one-third of more than 11,000 advertisements observed inside ChatGPT had no connection to the conversation in which they appeared.
The study came from Searchable, an AI visibility platform. It examined ads shown in real ChatGPT conversations between July 4 and August 4, 2026, paired each placement with its surrounding discussion and assigned it to one of three relevance bands. The findings describe that sample and grading process; they are not an independent audit of every ad served by OpenAI.
Contextual relevance is not immediate intent
Searchable classified 27% of the ads as direct matches for the product a user was currently discussing. Another 40% were contextual matches, connected to something mentioned earlier in the thread but not to the question being asked at that moment. The remaining 33% were judged unrelated.
The distinction matters because a long conversation contains many possible signals. A user might mention travel while planning a school project, discuss insurance while writing fiction or troubleshoot software without intending to buy anything. An ad can be thematically connected to the transcript yet still arrive at an unhelpful moment.
Searchable found that 68% of ads appeared in conversations containing no identifiable intent to buy, book, compare or hire at any point. Measured by conversation rather than individual placement, 40% of chats carrying ads included at least one unrelated advertisement. In 28%, every ad was unrelated to the discussion.
Those figures suggest that conversational targeting is currently better at recognising topics than purchase readiness. They do not establish whether the ads were ignored, clicked or converted, because the source provides no corresponding performance data.
Relevance varies sharply between sectors
The mismatch was not evenly distributed. Searchable judged 47% of marketing and business-to-business service ads unrelated, followed by 45% for software and SaaS. Data brokers and background-check services achieved the highest direct-match rate at 50%, while travel reached 43% and automotive 42%.
No sector produced a direct match more than half the time. Insurance behaved differently: only 23% of its ads were unrelated, but just 11% were direct matches. Two-thirds were contextual, often linked to a broader life event rather than the user’s present question.
ChatGPT’s system differs from paid search, where advertisers choose keywords associated with a potential buyer. Advertisers instead provide context hints describing the conversations, situations and subjects in which they want to appear, and OpenAI matches those hints against a live exchange.
That model potentially reaches people before they express a conventional search query. It also creates ambiguity. Conversation is used for studying, writing, troubleshooting and reflection, so a relevant subject does not necessarily indicate commercial interest.
OpenAI’s improvement claim meets a static mismatch rate
Denise Dresser, then OpenAI’s chief revenue officer, said in June that the rate at which users dismissed ChatGPT ads had fallen by half since the advertising business launched in February. OpenAI treats dismissal as a proxy for relevance. That is a company-reported improvement and uses a different measure from Searchable’s manual relevance bands.
Searchable says the unrelated share stayed close to one-third during its month-long sample even as ad volume increased. The observation does not show whether performance improved before or after that period, and a single month cannot establish a long-term trend. The source also does not disclose inter-rater agreement or how borderline examples were resolved.
For advertisers, the practical warning is that placement inside an assistant does not guarantee alignment with the answer or the user’s goal. A misplaced ad competes with a response that may already recommend a different option. Brands therefore need evidence about conversational relevance, purchase intent and outcomes rather than assuming the assistant’s knowledge of the thread makes targeting precise.
Conversational advertising could become more accurate as systems and campaign controls mature. For now, the reported mismatch shows that having context and interpreting it commercially are separate capabilities.
