OpenAI has finally pulled back the curtain on how advertising works inside ChatGPT. In a technical breakdown published this week, the company explains how a dedicated language model reads conversations, finds relevant ads, runs auctions, and decides what actually appears below a response. It is the clearest picture yet of an ad business that reached a $1 billion annualized revenue run rate less than 200 days after launch, with more than 100,000 advertisers testing the format.
For marketers, this matters because ChatGPT ads do not behave like search ads. There is no keyword and no guarantee that a winning bid buys you an impression. Understanding the pipeline matters more than optimizing any single lever.
Ads and Answers Run on Separate Tracks
The first thing to know: the system that picks ads runs separately from the system that writes ChatGPT’s answers. Each uses its own models and code. As ChatGPT composes a response, the ads system checks whether a relevant ad could appear below it. The ad choice never feeds back into what the assistant writes.
ChatGPT also does not wait for the ads system. If the ad model is slow or fails, the answer still reaches the user, and an ad that arrives too late simply never appears. This architecture supports OpenAI’s promise that advertisers cannot pay to influence what the assistant says. That is OpenAI’s description of its own system, not independent verification, and it is worth holding that distinction in mind.
When an ad does appear, it carries a sponsored label below the response, plus a short introduction written by the ad system explaining how the ad connects to the discussion. That introduction stays separate from the advertiser’s headline, description, and call to action.
To pick a relevant ad, the system can look beyond the latest question. Details shared earlier in the conversation feed the decision, which means context accumulates across messages.
Your Customer’s Intent Builds Across Messages, and So Does Ad Matching
OpenAI uses an unusual example to explain this: race training. A user might open with general training questions and only ask about shoes several messages later. The ads system tracks relevant details as the discussion develops, in what OpenAI calls “stateful” sessions. Each new message does not force the system to start from scratch.
For advertisers, the project management example is more useful. A buyer might mention they run a small agency, need client approvals, and want to avoid paying for guest accounts. A later question about pricing carries far more context than a search for “project management software pricing” ever could. An advertiser offering free guest access could message around that specific detail, and the landing page would need to back up the offer.
When users opt into personalized ads, the system can also draw on nonsensitive information from past conversations. That adds power and risk. The model must separate current intent from outdated details and hypothetical questions. A perfectly matched ad could reach someone researching a purchase months away. Relevance, in other words, does not equal readiness to convert.
Inside the Orchestrator: How Ads Get Found, Checked, and Auctioned
OpenAI calls the model managing ad selection the “orchestrator.” It reads the conversation, calls tools to find potential ads, checks relevance, and submits candidates to an auction.
The pipeline works roughly like this. First, the orchestrator turns relevant parts of the conversation into search queries. Those queries become embeddings, numerical representations of meaning that let the system find ads without requiring the same wording. A conversation can generate several queries when the person has different needs, and the orchestrator refines its searches as more information arrives.
Next, the system checks whether candidate ads are eligible to run: location targeting, policy restrictions, available budget, and pacing. Then it assesses each ad’s fit with the conversation, its past performance, and the likelihood of a view, click, or conversion.
Finally, candidates go to a separate auction tool that handles ranking, pricing, and selection. The language model helps decide which ads enter the auction. It does not directly pick the winner or set the price.
One honest limitation: OpenAI did not disclose how much weight each signal carries, so the breakdown offers limited guidance on improving an ad’s ranking. And even winning the auction is not the end of the story.
Winning the Auction Does Not Guarantee an Impression
Here is the part every advertiser should read twice. After the auction selects an ad, OpenAI runs another check on whether the ad fits the conversation. An ad can clear every policy and still be judged inappropriate for a particular discussion.
Auction winners can also miss the delivery window. Remember the timing problem from the first section: if the ad arrives after the response is ready, there is no impression. So low impressions could reflect failed relevance checks, slow delivery, eligibility issues, or several other chokepoints. Before you raise bids or rewrite creative, you need to know where your ads actually stopped progressing.
OpenAI says its internal logs track each stage: searches that found no potential ads, ads that failed eligibility checks, and placements that appeared and earned interaction. That visibility could explain why a campaign is underdelivering. A campaign excluded by location targeting needs a very different fix than one that enters auctions but rarely wins.
The catch: OpenAI did not explain how much of this information advertisers can actually see in their reporting. Without that visibility, advertisers may be forced to test changes blind, never knowing which stage limited their impressions.
The Billion-Dollar Question: What Did Advertisers Get for It?
The headline numbers are impressive: a $1 billion annualized run rate, achieved in under 200 days, from more than 100,000 advertisers. But neither figure tells you what advertisers earned. The revenue figure is a projection of a full year based on current pace, not money already collected. The advertiser count says nothing about how many stayed active, returned after a first test, or increased spending. Repeat spending would be the real signal that advertisers found value. That number is missing.
Also missing: CPA, ROAS, and conversion-rate benchmarks, including how results vary by industry or campaign objective. OpenAI disclosed engineering mechanics without disclosing advertiser outcomes.
There is also the incrementality question. A campaign can claim credit for purchases from customers who would have bought anyway. The post includes no controlled studies showing whether ChatGPT ads generated additional purchases. The figures establish that advertisers are spending. They do not establish that advertisers are acquiring customers profitably, or how consistently.
That gap is worth watching because early-adopter averages can hide huge variance between accounts. Who is winning on this platform, how much they spent, and how conversions were measured all remain open questions.
What Marketers Should Do Now
First, think in requirements, not keywords. The project management example shows the power of a specific detail like free guest access surfacing mid-conversation. Map the concrete requirements your customers voice: the constraints, the deal-breakers, the must-haves. Customer research and your best-converting search terms are the raw material. You do not need a campaign for every conversational variation, but your ads and landing pages should speak the language of real buyer requirements.
Second, if you are already testing ChatGPT Ads, demand funnel visibility before you spend more. Low impressions are ambiguous by design. Ask whether your ads entered auctions, failed relevance checks, hit delivery timing problems, or never found eligible candidates. Change bids only when you know bidding is the problem.
Third, hold the platform to outcome standards. A $1 billion run rate is OpenAI’s story. Yours is cost per acquisition, incrementality, and repeat profitable spend. Run ChatGPT ad tests with control groups and clean attribution so you know whether new demand was created or existing demand was harvested.
Finally, keep the verification honest. OpenAI’s separation of ads and answers is a described architecture, not an audited one. Treat it as the current best explanation, not settled fact.

FAQ
What is ChatGPT’s LLM ad stack?
It is the advertising system inside ChatGPT, coordinated by a dedicated language model OpenAI calls the orchestrator. The orchestrator reads conversations, converts context into search queries, finds eligible ads, and submits candidates to an auction that runs separately from the models that write answers.
How are ChatGPT ads different from Google search ads?
There are no keywords. Ad selection is driven by the full conversation context, embeddings match meaning rather than exact wording, and winning an auction does not guarantee an impression. The ad system also cannot influence the assistant’s written answers, according to OpenAI.
What does “winning the auction doesn’t guarantee an impression” mean?
After the auction picks an ad, OpenAI runs a final fit check on the conversation, and the ad must also arrive before the response is delivered. An ad can win the auction and still never appear.
Did OpenAI share advertiser performance data?
No. The company disclosed a $1 billion annualized revenue run rate and 100,000+ advertisers, but no CPA, ROAS, or conversion benchmarks, and no data on how many advertisers stayed active after testing.
Should my business test ChatGPT Ads now?
Test with a budget you can afford to learn from, set up clean attribution with control groups, and measure incrementality rather than reported conversions. Go in to learn what the platform can do, not to scale on day one.