ChatGPT served ads on 25.94% of commercial prompts and 14.35% of those ads were semantically unrelated to the prompt that triggered them, according to an SE Ranking study reported by Search Engine Land on 10 August 2026. The study analysed more than 50,000 commercial prompts across 20 niches to measure how often ChatGPT shows ads, where they sit in the response, and how relevant they are to what the user actually asked.
This is the first large scale independent read on the quality of ChatGPT Ads inventory. The findings that matter to a media buyer are the mismatch rate, the near total disconnect between being cited in an answer and being advertised beside it, and a category spread that runs from 2.6% mismatched ads in one niche to more than 50% in two others. SE Ranking also found that advertisers currently have no visibility into which specific prompts triggered their ads.
What SE Ranking measured, and on what sample
SE Ranking analysed more than 50,000 commercial prompts across 20 niches, as reported by Search Engine Land on 10 August 2026. The study set out to answer three questions: how often ChatGPT serves an ad against a commercial prompt, where the ad appears relative to the generated response, and how semantically related the ad is to the prompt that produced it. It also compared sensitive categories against Google AI Mode to see whether the two systems treat the same subject matter differently.
On format, the study found each ad appeared below the generated response as a single sponsored offer, with no competing advertisers shown in the same slot. That one sentence carries more strategic weight than any of the percentages, and it is treated separately further down.
The headline numbers from the SE Ranking study
Five figures define what the study found about ad frequency, relevance and the relationship between advertising and citation. They are set out below exactly as Search Engine Land reported them.
| What SE Ranking measured | Figure |
|---|---|
| Commercial prompts that returned an ad | 25.94% |
| Served ads that were semantically unrelated to the triggering prompt | 14.35% |
| Advertisers appearing as a cited source in the answer above their own ad | 3.63% |
| Exact advertised URL appearing in the response citations | 0.09% |
| Advertiser brands mentioned anywhere in the response | 4.44% |
Two of these numbers describe supply, and three describe the relationship between paid placement and the answer itself. Read together they say that roughly one commercial prompt in four carries an ad, that about one ad in seven does not match the question, and that in the overwhelming majority of cases the advertiser is a stranger to the text sitting directly above its own placement.
What a 14.35% mismatch rate does to the cost of a qualified click
This section is analysis. SE Ranking reported the mismatch rate. It did not report cost per click, spend, or conversion data for ChatGPT Ads, and none of those figures should be attached to this story.
The mechanism the mismatch rate implies is a gap between the impressions an advertiser pays for and the impressions that could ever have converted. If roughly one in seven served ads is semantically unrelated to the prompt behind it, then a share of delivery is going to users who were asking about something else at the moment they saw it. Those users are not a poorly qualified segment of the target audience. They are outside the intent the ad was written for.
The arithmetic follows without needing a single invented number. Whatever the effective cost of a qualified click turns out to be for a given advertiser, it is higher than the raw reported cost by the proportion of delivery that never had qualifying intent behind it. A mismatch rate in the mid teens is a mid teens tax on the denominator before any of the normal qualification loss inside the matched portion is counted. On a search results page an advertiser can find that waste in a search terms report and negate it. That option does not exist here, which is the point the study makes at the end and the reason the mismatch figure is not merely an interesting statistic.
What the source does not settle is whether the mismatch is a ranking artefact that improves as the system matures, a consequence of thin advertiser coverage in some categories, or a stable property of matching ads to conversational prompts. The study measures the rate. It does not diagnose the cause.
One sponsored offer with no competing advertisers is not a search auction
SE Ranking found that each ad appeared below the generated response as a single sponsored offer, with no competing advertisers shown in the same slot. Structurally that is a different market from a search results page, and the difference runs deeper than the number of slots.
On a search results page, several advertisers appear against the same query at the same time. The user compares them. Relative position, offer, and message are all live variables an advertiser can work on, because a competitor's ad is visible right next to their own. Underperformance shows up as lost impression share against named rivals, and the lever is incremental: bid more, write better, extend the offer.
A single slot with no visible competitor removes the comparison from the user's view entirely. The user sees one sponsored offer or none. For the advertiser who wins the slot that is an unusually clean placement with no adjacent price war. For everyone else, the outcome for that impression is absence rather than a lower position. The mechanism this implies is a winner takes the impression dynamic rather than a graded one, which tends to reward whatever the system uses to pick the single offer far more sharply than a ranked auction does. What the source does not say is how that selection is made, whether it is priced by bid, quality, coverage, or something else, and how frequently the same advertiser wins repeated prompts in a niche. Anyone claiming to know the auction mechanics from this study is going beyond the data.
Being cited and being advertised are almost entirely separate
The three citation figures are the most under discussed part of the study. Only 3.63% of advertisers appeared as a cited source in the answer positioned above their own ad. The exact advertised URL showed up in citations just 0.09% of the time. Advertiser brands were mentioned anywhere in the response 4.44% of the time.
Taken together, these say that paying for the slot does almost nothing to put a brand inside the answer, and that appearing inside the answer is a separate achievement earned through a different mechanism. An advertiser buying visibility in ChatGPT and expecting the answer above the ad to reinforce it is, on this data, buying one thing and hoping for another. The 0.09% figure is the sharpest of the three: the specific page an advertiser is paying to send users to is essentially never among the sources the model cited when composing the answer directly above it.
The practical read is that earned presence in AI answers and paid presence beneath them are two separate programmes with two separate methods. Getting a brand cited in generated answers is the work of ChatGPT SEO and content structure, not of media budget. This study offers a rough measure of how little the paid channel does for the earned one, which is a data point marketers have not previously had.
The category spread is the finding to act on
Irrelevance varied enormously by category in the SE Ranking study. Pets showed only 2.6% mismatched ads. Relationships and news and politics both exceeded 50% mismatched ads. That is a range of roughly twenty times between the cleanest and the messiest categories in the sample.
A single average across 20 niches is close to meaningless as a planning input when the underlying spread is that wide. The 14.35% headline figure is a portfolio number. No individual advertiser buys the portfolio. They buy their own category, and the honest conclusion for a marketer is that the study says very little about any specific niche other than the ones it named.
The reasoning that does transfer is directional. Categories with dense, well defined commercial vocabulary and clear product intent appear to match better. Categories where the prompt is conversational, emotional or discursive appear to match far worse. A prompt about a pet product names the product. A prompt about a relationship or a news event names neither a product nor a purchase, so the system is matching an ad to a conversation that has no commercial anchor in it. If a category sits closer to the second description than the first, the sensible assumption is that its mismatch rate is above the average rather than at it, until measurement says otherwise.
The brand safety reading of the healthcare comparison
SE Ranking compared sensitive YMYL categories against Google AI Mode. ChatGPT served ads on 28.69% of healthcare prompts against 2.64% in Google AI Mode. On news and politics prompts, ChatGPT served ads on 28.76% against 6.8% in Google AI Mode. Those are the figures the study reported.
The gap is roughly eleven times on healthcare and roughly four times on news and politics. This is a difference in commercial policy posture rather than a difference in technology. One system is monetising sensitive question types at close to its overall rate. The other is holding them back sharply. Neither posture is stated in the source as a policy; the study observed the outcome, not the rule behind it.
For brand safety the reading is straightforward and it cuts both ways. A healthcare advertiser gets far more inventory in ChatGPT than in Google AI Mode, and gets it against prompts where a user may be asking about a symptom, a diagnosis or a treatment. Ad density that high in a category that sensitive raises the odds of a placement landing next to a question the brand would not have chosen to appear against. Combined with the 14.35% mismatch rate and the absence of prompt level reporting, a healthcare advertiser cannot currently verify which questions their ads appeared beside. That is the situation the study describes.
The honest limit: you cannot audit any of this in your own account
SE Ranking noted that advertisers currently have no visibility into which specific prompts triggered their ads. This single limitation is what turns every other finding from a manageable problem into an unmanageable one.
On a search results page, the equivalent problems all have a workflow. Irrelevant matches show in a search terms report and get negated. Category level waste shows in segment reporting. Sensitive placements can be excluded. Every one of those workflows depends on knowing what the user actually asked. Without prompt level reporting, an advertiser has no way to reproduce the study's finding inside their own account, no way to identify which prompts are generating mismatched impressions, and no lever to exclude them.
What the source does not say is whether prompt level reporting is planned, whether any aggregated or categorised version of it exists for some advertisers, or what the pricing model attaches to. Those are open questions, and treating them as answered in either direction would be inventing detail the study does not contain.
What this means for Thai marketers
The SE Ranking study contains no Thai data. It covered more than 50,000 commercial prompts across 20 niches, and Search Engine Land's report says nothing about Thailand, Thai language prompts, or Southeast Asian markets. Everything in this section is reasoning from the reported findings, not a finding about Thailand.
The reasoning that matters locally is the YMYL overlap. Thai clinics, hospitals, insurers and financial services sit in exactly the categories the study found carry the highest ad density in ChatGPT and the widest gap against Google AI Mode. If a Thai healthcare or insurance brand is considering this channel, the questions to settle before spending are the ones the study leaves open: what proportion of delivery is matched to actual intent in that category, and what evidence the platform can provide about it.
The second local read is about language. The study does not report whether mismatch rates differ by language, and there is no basis in the source for assuming Thai language prompts behave like the English sample. A category that matches cleanly in English may not match cleanly in Thai, and nothing in this data answers that. Until there is a way to see prompt level delivery, a Thai advertiser entering this channel is running a test with a limited feedback loop, which is an argument for small controlled budgets and for keeping the earned side of AI visibility as the primary investment. Structuring content so that AI systems cite it is the more measurable half of AI search visibility work today.
FAQ on the SE Ranking ChatGPT ads study
How often does ChatGPT actually show ads?
On 25.94% of commercial prompts, according to the SE Ranking study of more than 50,000 commercial prompts across 20 niches. That figure covers commercial prompts specifically, and the study did not report an ad frequency for non commercial prompts.
Are ChatGPT ads relevant to what the user asked?
Often, but 14.35% of served ads were semantically unrelated to the prompt that triggered them. Relevance varied hugely by category, from 2.6% mismatched ads in pets to more than 50% in both relationships and news and politics, so the average is a poor guide to any individual niche.
Does buying a ChatGPT ad get my brand mentioned in the answer?
Almost never, on this data. Only 3.63% of advertisers appeared as a cited source in the answer above their own ad, the exact advertised URL appeared in citations 0.09% of the time, and advertiser brands were mentioned anywhere in the response 4.44% of the time. Paid placement and earned citation are separate outcomes.
Can I see which prompts triggered my ads?
No. The study states plainly that advertisers currently have no visibility into which specific prompts triggered their ads. That means the mismatch rate the study measured cannot be reproduced or acted on inside an advertiser account today, and the source does not say whether that reporting is coming.
Is any of this specific to Thailand?
No. The source contains no Thai data and makes no statement about Thailand, Thai language prompts, or regional markets. The relevance to Thai advertisers is that clinics, hospitals and insurers sit in the YMYL categories the study found carry the highest ad density, which is reasoning rather than a reported finding.
The short version
SE Ranking analysed more than 50,000 commercial prompts across 20 niches and found ChatGPT served ads on 25.94% of them, with 14.35% of served ads semantically unrelated to the triggering prompt, category mismatch running from 2.6% in pets to over 50% in relationships and news and politics, and healthcare ad density of 28.69% against 2.64% in Google AI Mode. Advertisers were cited above their own ad 3.63% of the time and the advertised URL appeared in citations 0.09% of the time. Search Engine Land reported the study on 10 August 2026, and the full write up is at Search Engine Land.
The honest summary for a media buyer is that this is early inventory with real reach, uneven quality by category, and no reporting layer to manage it with. That combination argues for controlled tests rather than budget shifts, and for putting the weight of AI visibility work on the earned side where the output can be measured. Relevant Audience helps Thai and regional brands build that earned presence and evaluate paid AI placements without guessing at the numbers. If you are being pitched this channel, the first question to ask is what prompt level reporting you will get.
