Can advertisers see which prompt triggered their ad?
No. And one impression in seven is no closer to its prompt than a random pairing, across 50,006 tested. What that leaves you.
No. And that answer is stable enough to plan around, because it is not a gap waiting to be filled in the next release — there is no search terms report in ChatGPT advertising, and advertisers who have run campaigns on it say there is no way to know which prompt an ad appeared inside. What makes this worth more than a one‑word answer is the second number, which is what the invisibility is actually hiding.
What you can see
The reporting is conventional performance data and it is not thin: impressions, clicks, spend, click‑through rate, cost per click, cost per thousand, conversions, return on ad spend, device and country. Conversions come through a pixel and a conversions API, in the way an advertiser already understands.
What is absent is the layer underneath. No prompt or query report. No demographics. No reach and frequency. No log‑level data and no scheduled export. No incrementality testing, no third‑party verification. View‑through sits on a fixed one‑day window you do not set.
Targeting works the same way round. There are no keywords to bid on. An advertiser supplies context hints at ad‑group level and the system decides what that matches, in a relevance‑weighted auction. So you do not choose the prompts, and you do not get told which ones you got.
You cannot pick the query, and you cannot see the query. The middle of the funnel is a closed box with a receipt attached.
The second number
This is the finding that makes the first one matter. SE Ranking ran 50,006 commercially oriented prompts across twenty US niches on 23 July 2026 — roughly 2,500 prompts a niche — and recorded what came back.
Ads appeared on 12,974 of them. Just under 26 per cent, so about one commercial prompt in four carries an ad. That part is unremarkable and arguably low.
Then they measured relevance properly. Every prompt, ad and generated answer was embedded, and similarity measured against a baseline built by pairing ads with unrelated prompts at random. An impression counted as off‑target only if its closeness to its real prompt fell inside the range typical of those random pairings. On that test, 14.35 per cent of ad impressions were no more related to the prompt they appeared on than an ad matched to a random question. A manual review of flagged cases suggested the true rate is higher, not lower.
1in 7
Share of ChatGPT ad impressions no more related to their prompt than a random pairing, across 50,006 commercial prompts.
Set the two findings beside each other and the problem is not that reporting is immature. It is that roughly one impression in seven was served against something the advertiser would not have chosen, and the one report that would reveal which ones does not exist. The spend is visible. The mismatch is not.
You cannot ask the model either
The obvious workaround is to ask ChatGPT where your ads ran. AdExchanger tried it, got an answer, and OpenAI confirmed the answer was wrong. Which is the expected result rather than a scandal — a model describing its own ad serving is generating plausible text about a system it has no reporting access to, and that is what it did.
The same study notes that the same prompt can return a different ad, or none at all, on a repeat run. So manual spot‑checking does not substitute for a report. Observe a prompt a second time and you may be looking at a different auction.
Advertisers have asked for the missing layer. The trade press through late 2026 has been fairly consistent on it: the asks are prompt‑level reporting, prompt matching, brand safety controls and incrementality tracking. Those are the requests of buyers who are spending and cannot audit, which is a reasonable position to hold and not yet an answered one.
Why this is structural rather than temporary
It is worth separating two explanations, because they lead to different decisions.
If the absence is a product gap, the right move is to wait, spend small, and expect a search terms report within a few quarters. If it is structural, waiting is the wrong move.
The case for structural is straightforward. A prompt is not a query. People type things into an assistant they would never put into a search box — medical details, financial positions, drafts of things they have not sent, a situation described in full because the assistant needs it. A keyword report hands the advertiser the user's words. For a meaningful share of prompts, those words are not disclosable, and no aggregation threshold makes a long sentence safely anonymous. A report that excluded everything sensitive would also exclude most of what made the prompt useful.
So the honest expectation is that something arrives, and that it is a categorisation rather than a transcript — topics, intents, buckets. Useful, and not the thing being asked for.
What that leaves you
Three consequences, and none of them is a bidding tactic.
The ad buy cannot be optimised the way search is. Search advertising improves by reading the query report and cutting what does not convert. Remove the query report and that loop is gone. What is left is creative, offer, landing page and total spend — the levers of a brand campaign, not a performance one. Budget it as the former and the measurement you have is adequate. Budget it as the latter and it is not.
The organic side of the same surface is the part you can actually influence. The assistant reads pages to answer the question, and what it reads and cites is observable. That is a different discipline from buying placement next to the answer, and it is the one with a feedback loop attached. We have written on how that changes what a studio publishes.
Being the answer beats being beside it. If one impression in seven is mismatched and you cannot tell which, the reliable position is to be the thing the assistant reaches for when the question is asked, rather than the thing placed alongside whatever it says. That is slower, it compounds, and it does not require a report nobody is going to publish.
Where this came from
The prompt volumes, the 25.94 per cent ad rate and the 14.35 per cent mismatch figure are from SE Ranking's ChatGPT ads study, 50,006 prompts collected 23 July 2026 and published 10 August 2026; it is also our source for the statement that advertisers “cannot see the specific queries or conversations where their ads appeared”. The attempt to ask the model directly, and OpenAI's correction of it, is reported by AdExchanger. The buyer‑side asks are covered by eMarketer and Adweek. One study, one market, one day of collection — the mismatch figure is the best available measurement of this and should not be treated as a constant. The argument in the section above about why prompt reporting may never arrive is ours, not theirs. Checked November 2026.