We asked 79 AI visibility tools which engines they cover. Eleven never say it on the page where they sell

A buyer comparing tools in this category has one question before any other: which AI engines does this thing actually watch. We read the page each of 79 vendors points its own catalogue entry at, on 27 August 2026, against a list of seventeen engine names frozen before the first request. 11 of the 77 pages we could read name none of them. 5 more name them only in the navigation menu, so the answer is in the chrome and not in the pitch.

Disclosure: we sell AI visibility measurement, so we compete with most of the vendors below, and we are in the catalogue ourselves. Here is the whole method, for all three arms: the page is the one each catalogue entry already points at, the engine list was frozen before the first request, navigation and header and footer are stripped before the body is read, the two extra word lists and the control group were written down before the arm that uses them ran, and the plan counts come from the catalogue on the day each vendor’s price was verified. Every row of every arm is in our measurement register. Anyone can repeat it against us.

The short version

1. 11 of 77 readable pages name no AI engine in the body that sells, which is 14 per cent of a category whose product is watching AI engines.
2. 6 of the 77 name none anywhere in the document, menu included. The other 5 put the names in the menu and not in the pitch.
3. Vendors who publish a price name no engine 13 per cent of the time; vendors who do not publish a price, 23 per cent. That is a 10-point gap on 64 and 13 pages, which is a direction and not a result at that size.
4. ChatGPT is the most named engine, on 64 pages. The median page names 5.
5. Our own first detector got this wrong, and the way it got it wrong is the reason the menu column exists at all.
6. We attacked the count with the list it depends on and it held: 0 of the 11 name any engine outside our frozen seventeen.
7. The page is describing the expensive plan, not the cheap one. Against the entry plan the median page names 2 engines more; against the top plan the median gap is 0.

What we asked, and what counts as an answer

The catalogue behind this site records, for every vendor, the URL we verified its pricing against. That is the page a buyer lands on, and it is the page this study reads. Nothing here is a crawl of a whole site: one page per vendor, one reading, on one day.

The engine list was written down before the first request and not touched afterwards: ChatGPT, OpenAI, GPT, Gemini, Google AI Overview, AI Mode, Perplexity, Claude, Copilot, Bing, Grok, DeepSeek, Llama, Meta AI, Mistral, Qwen and Doubao. Seventeen names. 16 of them are named by at least one vendor; the seventeenth is named by nobody in the sample.

Corrected the same day: the first version of this paragraph listed Kimi and left out GPT. The instrument always used the list above, so no figure on this page moves, and the correction is written here rather than quietly applied because a correction that changes no number is the easiest one to keep to yourself.

Four outcomes, kept apart, because folding any of them together is how a study like this invents its own headline. A page we read is a page whose body came back. A page whose body is under 200 characters is a shell, not an absence: it exists, we just cannot see it without executing scripts. A bot challenge is its own bucket, because that is a fact about the edge and not about the vendor. A host that never answered is a fourth. Of the 79 we asked, 77 were read, 1 answered a bot challenge and 1 never answered, and only those 77 are in any proportion on this page.

The vendors who name engines in the menu and not in the pitch

The first version of this detector read the whole document. It reported one vendor as naming three engines, which contradicted our own published review of that vendor, which says it names none. Both could not be right.

Opening the page settled it in a minute. The three names were in one mega-menu sentence, the kind of navigation that lists every product a company sells. Nothing in the body that pitches the product mentioned an engine at all.

So the detector strips navigation, header and footer before it reads, and what the menu names is kept in its own column instead of being folded in. That distinction turns out to carry a finding of its own, and it is one we did not predict: of the 11 pages that name no engine where they sell, 5 do name them in the menu. The information exists on the page. It is just not in the part that argues for the product.

That matters for more than tidiness. A retrieval layer reading a page for what it covers has no reason to weight a navigation list the way a human eye does, and a buyer skimming a pitch will not find in the menu what the pitch declined to say.

The price cross, and why we are publishing it as a direction

The design froze one cross before the data existed: does a vendor that publishes a price behave differently from one that does not. The answer is that the vendors who publish a price name no engine 13 per cent of the time, and the vendors who do not, 23 per cent.

10 points, on 64 pages and 13 pages. The honest way to report that is as a direction, because at those sizes one vendor rewriting one page moves the smaller figure by most of a point of ten. Our frozen design predicted a gap and named a threshold for calling it real, and 10 points fell between the two, in a band the design never named. That is a defect in the design and not a finding about the world, and it is the third time in this series that a prediction and its retirement threshold failed to cover the whole range between them. We publish the number, we say which band it landed in, and we do not choose the reading now that we can see it.

What the direction does suggest, and what we would want a bigger sample to test, is that the two silences travel together: a page that will not tell you what it costs is also more likely not to tell you what it watches.

The thin-page check

A shell page is excluded by definition, but the 200-character line is arbitrary and a page can be thin without being a shell. So the same figures were recomputed over only the pages with a body of 1,200 characters or more. That leaves 73 pages, and 8 of them name no engine, which is the same band as the headline.

That check matters because the alternative explanation for a zero is boring: a page with almost nothing on it names nothing. Restricting to pages with a substantial body removes that explanation and the finding survives.

What else we have asked this catalogue

This is not the first question put to the same population, and the answers do not line up the way a single reading would suggest. Asked whether they block AI crawlers, not one vendor in the catalogue prohibits a single one of the fourteen named crawlers we check. Asked whether they publish an llms.txt, a clear majority do. Asked whether they serve markdown to a client that requests it, a small minority do. Asked whether they name an engine where they sell, one in seven does not.

Put together, those four say something more useful than any of them alone: what this category implements is not one posture but a set of separate decisions, and the cheap ones are made far more often than the ones that cost a build step or an edit to the pitch. The convention that costs nothing but a file is widely adopted. The sentence that costs a rewrite of the page that sells is the one 11 vendors have not written.

We attacked our own number, and it held

The weakest thing about the figure above is that the list of seventeen is ours. A vendor naming an engine we did not ask about reads here as naming fewer, so the eleven could be an artefact of our own choosing. That is cheap to test and it is the first thing we tested: eleven page reads, no measurement calls, with the second word list and the control group written down before the first request.

We re-read the 11 pages that named no engine, and as a control the 11 that named the most, the same day with the same instrument. Against a list of engine names the frozen seventeen never asked about, from Kimi and Ernie to Poe, Baidu, DuckDuckGo and Genspark, 0 of the 11 name any of them. The count is not an artefact of the list.

The second question was what those pages say instead. 7 of the 11 use generic phrasing, “AI search”, “answer engines”, “LLMs”. That leaves 4 pages selling AI visibility that do not refer to engines at all, by name or otherwise.

And the control refuted the prediction that came with it. We expected generic phrasing to be what a page uses instead of names, so it should be commoner in the group that names none. It is the opposite: the 11 pages naming the most engines use generic phrasing on 11 of 11, four pages more than the group naming none. Generic wording does not substitute for naming, it travels with it. The pages that say “AI search” the most are the same pages that tell you which AI they mean.

The page names the expensive plan

There is a second reading in the catalogue that nobody has crossed with this one: for every vendor, how many engines the entry plan actually includes, read on the vendor’s own page on the day its price was verified. Put the two side by side and the question stops being about honesty and starts being about which product the page is describing.

60 vendors publish a number for their entry plan and were read. 40 of the 60 name more engines on the page than that plan grants, 67 per cent, with 9 naming exactly as many and 11 naming fewer. The median page names 2 more than the cheap plan includes, and 11 vendors name at least 5 more.

The obvious reading is that pages overpromise. The control says something more precise. On the 52 vendors where the top plan is also published, the median gap is 0 and only 19 pages name more than the top plan grants. The page is not exaggerating: it is describing the expensive plan. What the entry buyer gets is two engines short of what the page they read was about.

That is a different problem and a fixable one. The gap is not in the claim, it is in which plan the claim belongs to, and one line under the price would close it.

And the tail runs the other way too. 6 vendors name no engine at all while their entry plan includes some, which is the same set of pages the arm above found saying nothing about engines: their plan is more specific than their pitch.

The two sides are different units and the piece counts them as such. One is names printed in the body that sells, the other is engines included in a plan, so a positive gap is a distance and not a lie. The two halves were also read on different days, the naming on 27 August 2026 and each plan on the day that vendor’s price was verified, so a vendor that changed its tiers in between appears here with a gap made of two dates.

What this does not say

It does not say those 11 products do not cover any engines. It says their selling page does not print the names. A vendor can watch ten engines and describe them as “the major AI assistants”, and this study counts the printing, not the product. That is the right thing to count for a buyer comparing pages and for a retrieval layer reading them, and it is the wrong thing to count for a procurement checklist.

It is one reading of one page per vendor, so it is a proportion over vendors and never a rate over readings. A page rewritten tomorrow is not in it.

The engine list is ours and it is frozen, and the arm above tests exactly that: none of the 11 names anything outside it. What that arm cannot rule out is an engine absent from both of our lists, and a list drawn up six months from now would have different names on it.

And the population is a catalogue we maintain, which means it is the set of vendors we found and verified, not a census of the category. We add vendors to it when a measurement of ours turns one up, which is how it grew past eighty.

What a vendor should take from this

The fix is not a longer feature table. It is one sentence in the body of the page that says which engines the product watches, in the words the engines are called by, near the claim it supports. 5 of the 11 already have that sentence somewhere on the page: it is in the menu, where it does the least work.

If you want to see whether your own pages answer that question the way a retrieval layer reads them rather than the way you remember writing them, that is the sort of thing our product measures, and the register behind this piece has every row we read.

Common questions about engine coverage claims

How many AI visibility vendors do not name a single engine on the page where they sell?

11 of the 77 pages we could read on 27 August 2026, which is 14 per cent. 6 of the 77 name none anywhere in the document, and the other 5 name them only in the navigation menu. The full row-by-row list is in our measurement register.

Which AI engine do vendors name most often?

ChatGPT, on 64 of the 77 pages. The median page names 5 engines out of a frozen list of seventeen, and 16 of those seventeen are named by at least one vendor in the sample.

Does publishing a price go with naming engines?

In this sample, yes, as a direction. Vendors who publish a price name no engine 13 per cent of the time and vendors who do not publish a price 23 per cent, a gap of 10 points on 64 and 13 pages. The design froze a threshold for calling that real and 10 points landed below it, so we report it as a direction and not a result.

Why does navigation get its own column?

Because our first detector counted it and produced a figure that contradicted our own published review of a vendor. The names were in a mega-menu and not in the body that pitches the product. Keeping the two counts apart is what turned an instrument bug into the finding that 5 vendors name engines only in the chrome.

Can a vendor cover an engine without naming it?

Yes, and several probably do. This study measures what a page prints, which is what a buyer comparing pages and a retrieval layer reading them both have to work with. It is not a claim about product capability, and we say so wherever the figure appears.