OpenAI says its ads system differs from conventional advertising built around a single search or page-view event because user intent can develop across multiple turns and conversations. It describes the stack as a custom, post-trained frontier language model that calls tools for ad retrieval, ranking, relevance assessment, auctions and final selection.
The company says the system maintains stateful ad sessions and uses conversation context to update its understanding of intent. Ad selection runs in a separate service from answer generation, with separate models and code; OpenAI says ChatGPT’s organic response continues if the ad system is slow or unavailable, and that returning no ad is an acceptable outcome.
OpenAI says the orchestrator turns conversations into summaries and semantic embeddings, retrieves candidate ads, checks eligibility and relevance, and ranks opportunities. If users choose to personalize ads, it may consider available nonsensitive information from past conversations. The page does not specify how the system performs in practice or provide independent evidence for its claims about user benefit.
For AI builders, OpenAI’s design offers an example of pairing conversation-aware ad selection with a service boundary intended to keep monetization separate from answer generation. The caveat is that these are OpenAI’s descriptions of its own system, not independent validation that the separation prevents influence or that the ads are beneficial.