OpenAI has given advertisers the clearest look yet at how ChatGPT Ads decides what appears next to a conversation. The company says the platform hit a $1 billion annualized revenue run rate less than 200 days after launch, with more than 100,000 advertisers running campaigns. The growth is real. The more useful detail is the mechanics underneath it.
Ads stay separate from answers
OpenAI says the system that chooses ads runs on different models and code from the system that writes ChatGPT’s answers. Both can read the conversation, but the ad system’s decisions do not feed back into the text ChatGPT generates. If ad selection is slow or fails, the answer still ships. A late ad simply misses the response.
Sponsored placements carry a label below the response. OpenAI also writes a short introduction connecting the ad to the discussion, and that introduction is kept separate from the advertiser’s headline, description and call to action. This separation is OpenAI’s own description, not independent proof that ads never influence answers.
Conversation history changes matching
ChatGPT Ads treats sessions as “stateful,” meaning each new message does not start from scratch. The system can use details from earlier in the conversation to decide what a person needs now. OpenAI’s example: someone comparing project management tools says they run a small agency, need client approvals and want to avoid paying for guest accounts. A later question about pricing carries more context than a simple search for “project management software pricing.”
For advertisers, that is a shift. A product feature such as free guest access can become the ad’s hook even when the latest message is about pricing. Personalized ads can also use nonsensitive information from past conversations, but OpenAI says the model must separate current intent from old information and hypothetical questions.
Inside the orchestrator
OpenAI calls the ad-selection model the “orchestrator.” It converts relevant conversation parts into search queries, turns those queries into embeddings, and can run several queries as needs develop. Candidates then pass through eligibility checks for location, policy, budget and pacing, before being scored on fit, past performance and the likelihood of a view, click or conversion. The orchestrator sends candidates to a separate auction tool that handles final ranking, pricing and selection.
- No potential ads found in search
- Failed eligibility checks such as location or policy
- Lost the auction on ranking or pricing
- Won the auction but failed a post-auction relevance check
- Missed the delivery window for that response
Winning isn’t showing
Even after winning an auction, an ad can be blocked by a final fit check or arrive too late to appear. That means low impressions may have several causes, and raising a bid is not always the right response. OpenAI says internal logs track these stages, but it does not explain how much advertisers can see in reporting.
The $1B question remains unanswered
A $1 billion run rate projects a full-year pace; it does not mean OpenAI has collected $1 billion. The advertiser count also doesn’t show how many remained active, returned or increased spend. There are no CPA, ROAS or conversion benchmarks, and no controlled studies showing whether ChatGPT Ads added incremental purchases. Advertisers should treat the scale numbers as adoption signals, not proof of profitable returns.
Source: Search Engine Journal
