OpenAI has designed ChatGPT’s ad selection to operate separately from the system that generates replies, so an advertising-service delay or outage does not hold up a chatbot response.
To find ads, the system turns conversation context into search queries and embeddings, numerical representations of that context, then uses approximate-neighbor indexes to locate possible matches. A post-trained language model interprets the chat and nominates ad candidates. Retrieval and nomination are distinct steps: the system searches for possibilities, and the model puts forward candidates for consideration. An auction tool, rather than the model, determines the placement and price.
Candidates go through separate checks for eligibility, relevance and engagement. Other filters apply targeting and policy rules, manage campaign budgets and pacing, and remove duplicate ads. These checks cover individual candidates and the rules governing ad delivery, while the auction handles final selection and pricing.
At the last check, the system may suppress ads deemed inappropriate or those that arrive too late to display. ChatGPT places sponsored ads and advertiser creative below its replies. A different model writes a neutral introduction for each placement.
The advertising services, models and code are separate from the system that generates responses. That separation means a delay or outage on the ad side does not hold up ChatGPT’s answer. OpenAI also uses Pixel and Conversions API tools for campaign measurement and optimization. The company has said ads do not shape ChatGPT’s answers.
OpenAI says the ad program now reaches 89 countries, following an expansion of ChatGPT advertising to 89 countries. On Aug. 31, OpenAI reported a $1 billion annualized revenue run rate for its advertising business. That figure reflects the pace of revenue at the time; it does not mean the company had recorded $1 billion in advertising revenue over a completed year.
