OpenAI explains how ChatGPT ad auction works

ChatGPT ad auction: how OpenAI ranks sponsored placements

OpenAI says the ChatGPT ad auction runs separately from the system that writes answers. A custom language model orchestrator reads the conversation, identifies relevant ads, and sends candidates to a separate auction tool. OpenAI also says ChatGPT Ads reached a $1 billion annualized revenue run rate in less than 200 days, with more than 100,000 advertisers having run ads.

Key takeaways

• OpenAI says the ad system and the answer system use separate models and code, so ad selection does not feed back into ChatGPT’s organic response, according to Search Engine Journal.
• The orchestrator can use earlier turns in a conversation, turn the context into embeddings, and filter ads by eligibility before they enter auction, according to Search Engine Journal.
• OpenAI says an ad can win the auction and still fail a later fit check or miss the response window, so low impressions can come from several different bottlenecks, according to Search Engine Journal.
• OpenAI says ChatGPT Ads reached a $1 billion annualized revenue run rate in less than 200 days, and more than 100,000 advertisers have run ads, according to Search Engine Journal.

Source: Search Engine Journal

How does the ChatGPT ad auction turn conversation into bids?

OpenAI calls the model that manages ad selection the orchestrator. It reads the conversation, creates search queries from relevant parts of the exchange, and converts them into embeddings, numerical representations of meaning that help the system match ads without exact wording. The orchestrator can update those searches as more messages arrive, so a later question can change the ad candidates.

After finding possible ads, the system checks eligibility, including location targeting, policy restrictions, budget, and pacing. It then scores fit, past performance, and the likelihood of a view, click, or conversion. Only then does it hand candidates to a separate auction tool that handles final ranking, pricing, and selection.

Source: Search Engine Journal

Why can a winning ad still fail to show?

Because OpenAI says delivery has more than one gate. ChatGPT does not wait for the ads system to finish. If the ads system is slow or stops working, the answer still appears, and an ad that arrives too late will not show with that response. OpenAI also says a selected ad can fail a post-auction fit check if the conversation changes.

That means an impression can be lost after the auction, not just before it. OpenAI says its internal logs track no-match searches, eligibility failures, and placements that appeared and received an interaction. For advertisers, that matters because a bid problem and a delivery problem are not the same thing.

Source: Search Engine Journal

What do the revenue and advertiser counts really mean?

OpenAI’s headline figures show adoption, but they do not show advertiser economics. A $1 billion annualized revenue run rate is a projection based on the pace at that point. It does not mean OpenAI had already collected $1 billion from ads. The more than 100,000 advertiser figure also does not show how many remain active or how many increased spend.

The post does not give CPA, ROAS, or conversion-rate benchmarks, and it does not show controlled studies on whether ChatGPT Ads created incremental purchases. For a CFO or commercial leader, that is the missing piece. A platform can grow quickly and still leave buyers guessing about return on spend.

Source: Search Engine Journal

Why does memory matter in ChatGPT Ads?

OpenAI says the ads system can use details from earlier turns and, in personalized ads, nonsensitive information from past conversations. The point is to keep context across a session. That helps the system distinguish current intent from old information and hypothetical questions. In OpenAI’s example, someone can start with general training questions and later ask about shoes.

The practical effect is that a later query may carry more commercial intent than the wording suggests. OpenAI’s project management example shows that a buyer who mentions a small agency, client approvals, and guest accounts can trigger a more specific offer than a generic pricing query. That is useful for relevance, but it also means advertisers need to design offers that survive the full conversation, not just the last prompt.

Source: Search Engine Journal

What to do now

For companies testing ChatGPT Ads, the first move is not simply to raise bids. OpenAI’s own framework suggests you need to know where the ad dropped out, whether before eligibility, before auction, after auction, or because it missed the delivery window. The better answer is usually tighter targeting, clearer offers, and cleaner measurement.

Source: Search Engine Journal

• Map the customer questions that actually signal buying intent, not just the broad keyword someone might type at the start of research.
• Check whether the offer matches the conversation. A feature such as free guest access may matter more than a generic productivity promise.
• Review landing pages and creative against the same requirements you expect the orchestrator to detect.
• Ask for reporting that separates no-match searches, eligibility failures, auction losses, and late delivery.

Frequently asked questions

How does OpenAI decide which ChatGPT ad to show?

OpenAI says a dedicated orchestrator reads the conversation, turns relevant parts into search queries and embeddings, checks eligibility, and then sends candidates to a separate auction tool. That system uses context from earlier messages, not just the latest prompt. The final ranking and price come from the auction tool, not from the answer model.

Why can an ad win the auction and still not appear?

According to OpenAI, an ad can miss a later fit check or arrive after the response is already being sent. ChatGPT does not wait for ads if the ad system is slow. So an auction win is not the same as an impression.

What should advertisers watch before increasing bids?

They should first identify where delivery is breaking. OpenAI says internal logs can show whether the campaign found no ads, failed eligibility, won the auction, or appeared and got an interaction. Without that visibility, a higher bid might fix the wrong problem.