Media Morpheus

Explainerby Media Morpheus Research

October 23, 2026 · Last fact-checked: October 23, 2026

How Does ChatGPT Choose Which Ad to Show, and What Does OpenAI's Ad Stack Description Mean for Advertisers?

Through a separate service from the one writing the answer, OpenAI says: a model it calls the orchestrator turns the conversation into ad searches and nominates candidates, and an auction picks the winner and sets the price. OpenAI does not disclose how bids are weighted against relevance.

ChatGPT AdsAI AdvertisingPaid PlacementDisclosure

Key Findings

  • OpenAI says ad selection runs “separately from ChatGPT’s answer generation, in its own service,” and that “we don’t delay chat responses to wait for ads to load” (OpenAI).
  • A post-trained language model OpenAI calls the orchestrator “interprets conversations and runs the entire ads flow in an agentic loop,” but “doesn’t directly decide which ad will be served, nor does it set a price”; the auction does (OpenAI).
  • Ads that arrive in time appear below the response, “clearly labeled as sponsored”; if a user chooses to personalize ads, the orchestrator “also considers available nonsensitive information about the user’s past conversations” (OpenAI).
  • OpenAI now says “more than 100,000 advertisers have run ads on ChatGPT since launch,” up from “tens of thousands” in its August 31, 2026 post; it repeats that ChatGPT Ads reached a $1 billion annualized revenue run rate less than 200 days after launch (OpenAI).
  • Search Engine Journal, which reported the post on October 22, 2026, notes OpenAI did not disclose the weight given to each signal and gave no cost-per-acquisition, return-on-ad-spend or conversion-rate benchmarks (Search Engine Journal).

What Did OpenAI Publish?

An engineering description of the system that picks ads in ChatGPT. The post, “How we built ChatGPT’s novel ad stack for user trust and relevance,” carries no visible date or named author and refers to “our engineering team” (OpenAI). OpenAI’s site returned an access error to our direct request, so we read the post’s text through a text-reader service and checked each point against Search Engine Journal’s October 22, 2026 report of it, by Brooke Osmundson (Search Engine Journal). The two agree on every point below.

OpenAI frames the design around keeping ads out of answers: the ads process runs separately “to keep ads from influencing answers and maintain ChatGPT’s neutral voice” (OpenAI). Search Engine Journal points out that this separation is OpenAI’s own description, not independent verification.

How Does OpenAI Say ChatGPT Ads Are Chosen?

By a pipeline that reads the conversation, retrieves eligible ads and sends them to an auction. OpenAI says the orchestrator turns conversations into text summaries, which are converted into semantic embeddings, numerical representations of meaning, used to retrieve ads, so an ad need not match the user’s exact words. The orchestrator “maintains a stateful session, meaning the ads session runs for the duration of an eligible turn,” rather than starting from scratch with each message (OpenAI).

Candidate ads are then filtered. OpenAI lists location (“which locales a campaign is allowed to operate in”), “policy restrictions,” “budget availability” and pacing, meaning whether a campaign should spend now or later in the day. Search Engine Journal adds that the remaining candidates are assessed on fit with the conversation, past performance and the likelihood of a view, click or conversion (Search Engine Journal).

Who Picks the Winner, the Model or the Auction?

The auction. OpenAI says: “The orchestrator doesn’t directly decide which ad will be served, nor does it set a price.” It nominates candidates, and the auction handles final selection and pricing. The post does not name the auction type or say how bids are weighted against relevance and predicted engagement (OpenAI; Search Engine Journal).

A win does not guarantee an impression. “After the auction, there are two ways an ad might fail to deliver,” OpenAI says: the orchestrator can block an ad that seems out of place in the conversation, and an ad that arrives too late is dropped for that turn (OpenAI).

StageWhat OpenAI saysNot disclosed
SeparationAds run in their own service; chat responses are not delayed for adsIndependent verification
MatchingOrchestrator summarizes the conversation; embeddings retrieve adsRelevance thresholds
EligibilityLocation, policy restrictions, budget availability, pacingFull policy list
AuctionAuction selects and prices; orchestrator only nominatesAuction type and signal weights
DeliveryBlocked if out of place, dropped if lateHow often either happens
DisplayBelow the response, “clearly labeled as sponsored”Label design rules
LoggingRecords ineligible turns, empty retrievals and delivered adsWhat advertisers can see

Does ChatGPT Use Past Conversations to Place Ads?

Only if the user opts in to personalized ads, OpenAI says. In that case the orchestrator “also considers available nonsensitive information about the user’s past conversations” (OpenAI). The post does not define “nonsensitive,” and we do not guess at a definition.

How Big Is ChatGPT Advertising Now?

OpenAI gives two figures. It says ChatGPT Ads reached “$1 billion in annualized revenue run rate” less than 200 days after launch, and that “more than 100,000 advertisers have run ads on ChatGPT since launch” (OpenAI). An annualized run rate projects recent revenue across a year; it is not revenue collected, as Search Engine Journal notes, and the post does not give a launch date.

The run-rate figure is not new: OpenAI’s August 31, 2026 post already gave it, and said ChatGPT Ads was “used by tens of thousands of advertisers.” The advertiser count is the new number, well above both that and the 60,000 an analyst read off an OpenAI slide in early October, which we covered in our guide to the 60,000-advertiser figure. It counts every advertiser that has run ads since launch, not those active now.

What This Means for Anyone Buying a Placement

A ChatGPT ad is advertising, shown beside an answer, not part of it. Payment does not buy a recommendation in the model’s reply; our guide on whether you can pay to appear in a ChatGPT answer covers the difference, and our guide to creative enhancements covers what OpenAI may rewrite. Our recommendations, not OpenAI rules:

  • Ask for performance data, not totals. OpenAI gave no CPA, ROAS or conversion-rate benchmarks. Do not plan spend around the $1 billion run rate or the advertiser count.
  • Ask what you can see. OpenAI logs ineligible turns and empty retrievals internally; ask whether advertisers get any of that reporting.
  • Test relevance, do not assume it. Bid weights are undisclosed, so the price of relevance has to be measured in your own campaigns.
  • Promise nothing. A winning bid can still be blocked or arrive too late, so no ad buy guarantees impressions, clicks or conversions.

Methodology and Limitations

Four limits apply. First, openai.com returned an access error to direct requests, so we read the post through a text-reader service and cross-checked it with Search Engine Journal’s report; we could not see any date or author metadata. Second, everything here is OpenAI’s description of its own system; no independent audit has been published. Third, we did not test the ad system. Fourth, there is no chart, because OpenAI publishes no comparable figures beyond the two totals above.

Conclusion

OpenAI’s design keeps the ad decision outside the answer and leaves price to an auction, and it is OpenAI’s own account. The uncomfortable conclusion for an advertiser is that the weights deciding who wins remain undisclosed, so relevance and price are things to test, not assume.

Frequently Asked Questions

Does the ad system decide what ChatGPT says?

OpenAI says ad selection runs separately from answer generation, in its own service, to keep ads from influencing answers. That is OpenAI’s own description, not independent evidence.

How does ChatGPT decide which ad wins?

OpenAI says a model it calls the orchestrator nominates candidate ads, and an auction picks the winner and sets the price. It does not disclose how bids are weighted against relevance.

Are ChatGPT ads labeled?

Yes. OpenAI says ads appear below the response, “clearly labeled as sponsored.”

Does ChatGPT use my past conversations for ads?

Only if you choose to personalize ads, OpenAI says; then it considers “available nonsensitive information” about past conversations. The post does not define nonsensitive.

How many advertisers use ChatGPT ads?

OpenAI says more than 100,000 advertisers have run ads on ChatGPT since launch, and that ChatGPT Ads reached a $1 billion annualized revenue run rate less than 200 days after launch.

Can paying for an ad get a company mentioned in ChatGPT’s answer?

No source we read says so. The ad is shown below the answer and labeled sponsored, and OpenAI says the ads system is kept separate from answers.

Sources and Methodology

This Explainer is built from OpenAI’s engineering post on its ad stack, read through a text-reader service because openai.com returned an access error, and Search Engine Journal’s October 22, 2026 report of it, both read on October 23, 2026. Recommendations are house advice. We did not test the ad system.

  1. official OpenAI, How we built ChatGPT’s novel ad stack for user trust and relevance (undated). Separation, orchestrator, embeddings, eligibility filters, auction, delivery checks, label, personalization, logs, run rate and advertiser count. Read October 23, 2026, via a text reader.
  2. trade press Search Engine Journal (Brooke Osmundson), Inside ChatGPT’s Ad Auction: How Relevance, Bids, And Memory Decide Placement, October 22, 2026. Cross-check of each point; undisclosed weights and missing benchmarks. Read October 23, 2026.
  3. official OpenAI, August 31, 2026 post on ChatGPT Ads. Earlier $1 billion run-rate and “tens of thousands of advertisers” figures, as quoted in our October 16, 2026 guide. Read October 16, 2026.