The Shop Counter
Picture a shop assistant handing you a leaflet that ranks five products. You are about to read it when they lean in and add: “Just so you know, the company in first place printed that.” You would still read it. You would read it very differently. And you would wonder what else they had printed.
That is what ChatGPT has started doing, in writing, in some of its answers. The quiet trick of publishing your own league table and letting an assistant read it out as advice now comes with a footnote, and some of those footnotes name the vendor. It does not publish a list of offenders; it adds the caveat one answer at a time. Below is what was seen, who saw it, the exact words, and the question every marketer will ask first: is that us?
What was observed, and by whom
One distinction before the detail. The company that publishes a ranking and the company ranked first are not always the same. A vendor can write its own list, pay a review site for placement, earn commission from an affiliate publisher, or simply be ranked well by someone independent. Those are different degrees of involvement, and the caveats below reflect that range.
On 19 October 2026 Lily Ray, an SEO practitioner who tracks AI answers closely, published on X what she had seen in her own monitoring, with two screenshots of the export that readers can inspect. She had been tracking a long list of B2B software prompts in a monitoring tool. Of roughly 700 ChatGPT answers in the previous week, 82 (about 11.7 percent) included some kind of disclaimer, and 19 (about 2.7 percent) specifically warned the user about who writes the rankings and reviews that ChatGPT draws on. The denominator is answers, not vendors or pages, so the figures say how often the caveat appeared, not how many companies it touched. She did not publish the model or mode used, whether prompts were repeated, or how a disclaimer was classified; treat the numbers as a practitioner’s count, not a study.
The wording is the useful part, so here it is as she quoted it:
“Because this ranking is published by the vendor itself, I would treat it as one input rather than an independent verdict.”
“Keep in mind that many ’best tools’ lists are published by vendors, so treat rankings cautiously and compare a few sources.”
“Check whether the author has tested the tools or is primarily summarizing vendor claims.”
“One important caveat: many comparison articles are published by companies selling AI visibility products or related services.”
Her summary of what ChatGPT was cautioning about: vendors ranking themselves in their own comparison pages; publishers affiliated with a brand they review; whole categories where most comparison articles are written by companies selling the product; sponsored placements on review sites; recommendations that exist mainly to carry affiliate links; and taking vendor claims at face value.
Her own reading was that this is “a temporary bandaid” while AI companies work out better ways to surface objective answers, and that it gives OpenAI cover if ChatGPT recommends poor products in categories full of self-promotion. Her shorter verdict, posted on X the same day: these listicles are mostly just votes for your competitors at this point.
What this does and does not show
This is one practitioner’s observation, over one week, in one category, B2B software. It is not an OpenAI announcement. Nobody outside OpenAI knows the rule that produces these sentences, whether it applies to every category, or whether it will be there next month. Lily Ray herself wrote “appears to be”.
Treat it the way the disclaimer tells you to treat a vendor ranking: as one input. It is a strong input because it is a visible example of the warning moving from behind the scenes (where search engines have handled sponsored and manipulated content for years) into the answer the buyer actually reads, sometimes with the publisher identified.
This did not come from nowhere
Two earlier, published data points point the same way.
On 25 June 2026 Wil Reynolds and Nick Haigler of Seer Interactive published a study of 617 prompts, plus one agency-selection prompt run 30 times, looking at the sub-queries ChatGPT 5.5 generates behind an answer. Those sub-queries increasingly named specific brands and people rather than generic “top X” terms. Their reading: brand recognition may matter more than listicle volume.
On 17 August 2026 Lily Ray published an analysis of Peec AI data after ChatGPT 5.6 became the default. Listicles and comparison pages were losing share of retrievals, product pages had moved ahead of them (16.39 percent of retrieved pages), and the site: operator had gone from about 0.3 percent of ChatGPT’s sub-queries to about 23 percent, which she read as working “at least in part” as a spam filter in favour of trusted domains. The page types losing share were, in her words, “the formats most heavily spammed for GEO over the past year or two.”
These observations measure different things: how ChatGPT searches, which pages it retrieves, and what qualifications appear in its answers. The measurements belong to Peec AI and Seer; the readings belong to the practitioners. Together they raise a worthwhile question about the future of vendor-written rankings. They do not prove that ChatGPT has introduced a penalty, that one mechanism caused all three changes, or that comparison content has stopped influencing recommendations.
One more figure bears on the same question from a different angle. Semrush’s 2026 AI Visibility Index for the UK and US, published on 13 October 2026 and reported by Search Engine Land the next day, found that, when the top 100 mentioned brands were matched against the top 100 cited domains, fewer than half appeared on both lists, on average, in each month measured. Brands and domains are different entity types, so this is Semrush’s matching of one list against the other, not a measure of whether each brand’s own website was cited. The data illustrates why mentions and citations need to be measured separately. Whether a brand is actually recommended is another question again, and that figure does not answer it. It is Semrush’s own data, published on a site Semrush owns, and it still makes the narrower point.
A conversation I have had more than once this year
The setting changes, the words do not.
A marketing lead, under pressure, shows me a competitor’s “best tools for X” article. The competitor ranks itself first. The article is being cited in AI Overviews and in ChatGPT answers. “They are doing it and it is working. We need one of these.”
I understand the pressure completely. In many categories clicks are down because the answer now sits on the results page, budgets are tight, and the board wants to see something. Here is a competitor apparently being rewarded for a tactic you could copy by Friday. That is the same in B2B and in B2C; the only difference is the size of the invoice. My concern is not that comparison pages are wrong. It is that copying the format is easy and copying the credibility behind it is hard, and a citation on its own says remarkably little about whether that competitor is winning customers.
Here is what I say, and it is not “don’t”.
You are looking at one screen. What an AI answer shows you depends on whether you are logged in, your location and IP address, your history with that assistant, the model version serving you that day, and the exact wording of the prompt. Your competitor’s citation may be real and may not appear for your buyers at all. The only honest way to know is to run the same questions repeatedly, from clean sessions, over weeks, and log what comes back. Seer ran one prompt 30 times before drawing a conclusion from it; most screenshots in a Monday meeting were run once.
It changes faster than you can copy it. Semrush’s AI Visibility Index, reported by Leigh McKenzie in Search Engine Land on 24 September 2026, found that 40 to 60 percent of the sources cited across 2,500 prompts changed from one month to the next. A citation today proves that a source appeared in that answer. It does not tell you how consistently it appears, whether the brand is recommended, or whether the visibility produces enquiries.
Being cited is not being recommended. Wil Reynolds of Seer Interactive described, on 2 October 2026, the experience of being visible and anti-recommended: named in an AI answer only for the reader to be told to look beyond you. A self-ranked listicle is a plausible candidate for that treatment, and in the answers above the assistant says why, next to the publisher’s name.
The copy is the weakest version of the page. When you copy a competitor’s list you inherit its structure, its claims and its conflict of interest, without the authority that may actually have earned the citation: their reviews, their case studies, their named people, their presence on sites they do not control. You get the leaflet without the shop.
So the advice is not “never publish a comparison”. It is: publish the one that survives the disclaimer.
Four signals to record from every AI answer
Visibility is not one thing, and these are not four alternatives: a brand can be cited, mentioned, recommended and qualified in the same answer. Record each separately, because the fourth sits quietly next to the third until you read to the end.
| Signal | The question to answer |
|---|---|
| Citation | Was your content used as a source? |
| Mention | Was your brand named? |
| Recommendation | Were you positively recommended for the buyer’s need? |
| Qualification | Was the recommendation or the source accompanied by a credibility caveat? |
A monitoring routine that counts mentions and citations and stops there will miss the fourth signal entirely. This is the point Ranked Is Not Named made about search results, now with a fourth row.
Three things to do instead
1. Disclose before the assistant does it for you. If you compare products you sell against others, say so in the first paragraph, in plain words, on the page. “We make one of these. Here is how we compared them and what we would say against ourselves.” In the answers Lily Ray quoted, the assistant treats a publisher that discloses its commercial relationships as a point in that publisher’s favour. Make it true of you.
2. Put the evidence where someone else holds it. Reviews on platforms that verify the reviewer. Case studies with a named client, a date and a number the client approved. People with names and roles on the page. These give a buyer, and an assistant, a way to corroborate your claims against something you do not control, and they are the things a competitor cannot copy by Friday. This is the Corroboration layer of the Entity SEO work, and it is slower than a listicle, which is exactly why it holds.
3. Measure properly before you imitate anyone. Decide the ten questions your buyers actually ask. Run them weekly in controlled sessions and record the prompt, platform, date, market, the answer, who is named, who is cited, which of the four signals above each brand got, and whether any caveat appeared anywhere in the answer. Repeated, documented observations beat a single screenshot of a competitor’s good day. If their listicle still shows after a month, you have learned something. If it does not, you have saved a quarter. The routine is set out in How We Measure AI Visibility.
Check your own comparison pages: the Comparison Credibility Audit
Five questions about the page, then one about how assistants present it. Answer them honestly before an assistant does it for you. This is an editorial checklist, not a measure of how ChatGPT decides anything, and the questions do not carry equal weight.
The three that matter most. If any of these are missing, readers have less information with which to judge the ranking. The more consequential the claim, the stronger the evidence should be:
- Ownership disclosed. Does the page say who publishes it and which products they sell, in the first paragraph?
- Ranking method explained. Are the criteria, weighting and inclusion rules on the page?
- Testing evidence. Are there first-hand tests, dates, data or demonstrations, or only claims?
The two that separate a useful comparison from a defensible one:
- Fair treatment of competitors. Does the page give real reasons a buyer might choose an alternative?
- External corroboration. Can the important claims be checked on sources you do not control?
And one that is not about the page at all, but about what happens to it:
- Presentation in AI answers. When the page appears, which of the four signals above do you get: cited, mentioned, recommended, qualified?
A comparison that is honest about what it is remains fine to publish, as long as you are not relying on an assistant to mistake it for research. If you want the audit run against your pages and your buyers’ questions, that is what the AI Visibility Audit does.
What we are not claiming
We are not claiming OpenAI has announced a policy, that these disclaimers appear outside B2B software, that they will persist, or that they affect which sources ChatGPT chooses. We are not claiming listicles never work; some comparison pages are useful and some are cited for good reasons. The observation does not prove that ChatGPT has changed how it ranks or retrieves comparison content. It does show that, in some answers, the assistant is explicitly drawing attention to the commercial interests behind rankings. That gives marketers another reason to distinguish between getting a citation and earning a credible recommendation, and it is why the pressure to copy a competitor’s list is understandable and still a poor bet.
Evidence ledger
site: in about 23 percent of sub-queries.