Low impressions, higher bids
The reflex in every service business is the same. Impressions are low, so the bid goes up, then the creative gets rewritten. On the platforms you already know, that reflex has some logic, because your ad is mostly competing for a slot against other bidders.
In ChatGPT there is no slot waiting for you. Your ad has to be found by a model reading the conversation, pass eligibility, fit what the person is working through right now, and arrive before the answer is finished. The bid enters at the end, among the candidates that survived everything before it. More than 100,000 advertisers have run ads there since launch, so the volume is real.
A model reads along
OpenAI calls the model managing ad selection the orchestrator. It reads the conversation and calls tools to find potential ads, check their relevance, and submit suitable options to an auction. It turns the relevant parts of the conversation into search queries and then into embeddings, numerical representations of meaning that let it find ads without requiring the same wording.
It also remembers. OpenAI describes stateful sessions, so each new message does not require starting from scratch, and the system uses past conversation context to tell current intent apart from old information and hypothetical questions. Candidates then face eligibility checks including location targeting, policy restrictions, budget, and pacing, and the system weighs how well each ad fits the conversation, its past performance, and the likelihood of a view, click, or conversion. OpenAI did not disclose how much weight each signal receives.
Winning is not appearing
After the auction picks a winner, OpenAI applies another check to decide whether that ad fits the conversation, so even an auction winner may not be shown. There is also a clock. The ads system runs separately from answer generation, ChatGPT does not wait for it, and an ad that arrives too late will not appear with that response.
When your ad does appear, it sits below the response with a sponsored label and a short introduction explaining how the ad connects to the discussion, kept separate from your headline, description, and call to action. That introduction is the system saying, in its own words, why you belong in that conversation. The clearer the need your offer answers, the easier that sentence is to write.
What changes this week
Start by naming, in one sentence per offer, the customer requirement it answers. Not the service, the requirement, the way a roof inspection before an insurance deadline is a different need from financing that lets a renovation start this month. Whoever decides marketing can do this in an afternoon with the sales notes and the search terms that already convert.
Then change the order of the diagnosis. Before answering low impressions with a higher bid, find out whether your ads entered auctions, failed relevance checks, or ran into another delivery issue, because each one calls for a different move. OpenAI says it tracks those stages internally without saying how much advertisers can see, so plan on testing. By Friday you have stopped paying more for the lever that moves last, and every offer says out loud which need it answers.