OpenAI explains how a separate AI system selects ads from ChatGPT conversations
The system follows evolving conversations and can use past chats when users choose personalization. OpenAI says separate models and code keep ad selection outside answer generation.
OpenAI’s October 21, 2026 disclosure details an ad-serving layer built around an “orchestrator,” a separate language model that tracks conversational context and coordinates ad search, relevance checks, auctions, and delivery. The architecture keeps advertising out of answer generation and can return no ad; OpenAI says it also prevents ad-system outages from delaying replies. OpenAI reports a $1 billion annualized revenue run rate in under 200 days and more than 100,000 advertisers since launch, but a separate early-October estimate of about 60,000 active advertisers does not establish retention.
01
With ad personalization enabled, the orchestrator may use available nonsensitive information from past conversations alongside the live conversation.
02
The system is designed to distinguish current buying intent from stale context and hypothetical remarks, rather than treating every product mention as a signal.
03
Ad candidates can be checked for campaign activity, geographic targeting, policy restrictions, and budget before relevance, engagement, and historical performance inform selection.
ChatGPT’s advertising system does more than match an ad to the latest message. In an October 21, 2026 technical disclosure, OpenAI described a separate AI model that follows conversations, interprets commercial intent and coordinates ad selection. The company says this process runs outside answer generation, using separate models and code to keep advertising from influencing ChatGPT’s responses.
Finding a buying signal without treating every mention as one
The central component is a specially trained large language model OpenAI calls the orchestrator. It interprets context and repeatedly calls software tools to retrieve possible ads, rank them, check relevance, run auctions and arrange delivery.
A running conversation might begin with training advice and later turn to shoes. The system retains and updates context across turns. If users choose ad personalization, it can also consider available nonsensitive information from past conversations, OpenAI says.
The orchestrator must distinguish current intent from stale context and recognize hypothetical remarks—not treat every product mention as buying intent. It distills relevant context into queries and numerical representations of meaning, called embeddings, to search the ad inventory.
An auction behind a separate service
Candidates face checks that can cover campaign activity, geographic targeting, policy restrictions and remaining budget. Surviving ads are scored for relevance, predicted engagement and historical performance. Relevance and safety filtering happen before the auction, which handles final selection and pricing.
The separation also protects availability, according to OpenAI: answer generation has no dependency on ad serving, so a slow or unavailable ad system does not delay or interrupt ChatGPT’s response.
Advertisers tried it; how many remain active?
OpenAI also said ChatGPT Ads reached a $1 billion annualized revenue run rate less than 200 days after launch—a yearly revenue pace, not $1 billion collected during that period. More than 100,000 advertisers have run ads since launch, the company said.
That advertiser total is cumulative. AdExchanger cited The Information’s reporting that ChatGPT had about 60,000 active advertisers in early October, based on research by New Street Research analyst Dan Salmon. The figures differ in timing and definition, so they do not establish a retention rate.
Sources
openai.comHow we built ChatGPT’s novel ad stack for user trust and relevance
adexchanger.comChatGPT Ads Is Expanding To New Markets And Exploring New Ad Formats | AdExchanger
Reader comments
Newest comments first. Replies stay oldest first.