AI Answers Name One Set of Vendors and Read Another. In 27 Namings, 4 Had Their Own Page Cited.

Every AI visibility tool reports two things and most buyers read them as one: whether the answer said your name, and whether the answer used your page. We measured the distance between them. Across 7 software categories on two of Google’s AI surfaces, answers named a vendor from our frozen lists 27 times. In 4 of those 27 the vendor’s own domain appeared in the citation list. Being recommended and being read are not the same game, and in this market almost nobody wins both.

Disclosure and method: EchoWi sells AI visibility measurement, so this measures a distinction our own product has to get right. 48 vendor observations across 8 usable cells and 7 categories, United States, English, 20 August 2026, on AI Overview and Gemini. Every brand named below belongs to somebody else. The design, the statistic and three predictions were frozen before the first counted call, and all three are reported below whether they held or not. A second arm then repeated the identical design in Spain, in Spanish, adding 24 observations across 4 cells. The rows are in our measurement register, one per vendor observation, so you can recount them.

The short version

1. Answers name vendors constantly and read vendor pages almost never. Of 27 observations where an answer named a vendor from the frozen list, 4 also cited that vendor’s own domain. That is 15 per cent, with a 95 per cent interval of 6 to 32.
2. The gap runs almost entirely one way. 23 observations are named but not cited. Only 2 are cited but not named.
3. Every vendor domain that did get cited is a comparison page, never a product page. All 6 are a vendor’s own listicle, blog roundup or comparison table.
4. In 4 of 8 cells not one vendor domain was cited at all. The answer named three or four vendors and built itself entirely from publishers, forums and video.
5. The one thing a citation does buy is a name. Of the 6 observations where a vendor’s domain was cited, 4 also named the vendor.

What was measured, and the two traps avoided first

The unit is one vendor in one answer on one surface. For each software category we froze a list of well known vendors before making the call, so the measurement could not quietly select the vendors that happened to appear. Then, for each vendor, two independent readings: did the answer’s prose say the name, and did the vendor’s own registrable domain appear in the answer’s citation list.

Two traps had to be closed before any counting, and both would have manufactured the finding rather than measured it.

The first is the answer body’s own links. An answer’s prose carries markdown links, so a raw text search for “Zoho” matches zoho.com/bigin inside a URL and credits a name the answer never said out loud. URLs are stripped from the prose before the name is looked for.

The second is worse, and it removed two surfaces from the study. On AI Mode, this route returns every citation with its domain wrapped in a Google redirect, so the vendor’s real domain is never visible in the field a count would read. On ChatGPT, no citations come back at all. In both cases “cited” would be false by construction, so every observation would have landed in the named but not cited box and produced exactly the headline this study was looking for. A finding that your instrument cannot fail to produce is not a finding. AI Mode and ChatGPT are therefore absent by design, and that is a real limit on what follows, not a footnote.

The result

Read across the top row: 4 of 27, or 15 per cent, with a Wilson 95 per cent interval of 6 to 32. The interval is wide because 27 observations is a small sample, and quoting the point estimate without it would be the sort of precision this data does not support.

Read down the first column instead and the picture inverts: of the 6 observations where a vendor’s own page was cited, 4 also named the vendor. Getting read is rare, and when it happens it usually comes with the name attached.

The corner that matters commercially is the 23. Those are vendors an answer recommended to a buyer while sourcing that recommendation from somebody else entirely.

The cells that cited no vendor page at all

The cleanest cell is web hosting. The answer names Hostinger, SiteGround and Bluehost, and builds itself from Reddit, CNET, Forbes, TechRadar and three independent blogs. Not one hosting company’s own domain appears. The set of vendors recommended and the set of pages read are completely disjoint.

Project management does the same thing: Asana, Trello and ClickUp are named, and the sources are Forbes, PCMag, Zapier, a consultancy blog and two YouTube videos.

This is the part that should reorder a vendor’s thinking. In these categories the page that decides whether you are recommended is not your page. It is a review site’s comparison table, and your name is a row in it.

Where a vendor page does get cited, it is always a comparison page

Six observations have a vendor’s own domain in the citation list. Every single one is the same kind of page:

• An accounting and payroll company’s roundup of the best payroll providers, which is the page the answer used to recommend a competitor.
• An email platform’s guide to the best email marketing platforms, used to recommend three of its rivals.
• A password manager’s blog post ranking the best business password managers, used to recommend a different one.
• A password manager’s own comparison table, a second password manager’s blog, and a help desk company’s list of the best help desk software.

Not one is a product page. Not one is a pricing page. The pages that get read are the pages where a vendor writes about the whole category, and the reward for writing one is that the answer quotes your page while recommending somebody else.

That is the mirror image of a pattern we have measured before, that ranked lists on page one put their own publisher first. The publisher does come first in its own list. What this study adds is what happens next: the assistant reads the list and does not necessarily pass the ranking along.

The two vendors that were cited and never named

Only 2 of 48 observations are cited without being named, and both are worth looking at because they are the state most people assume is common.

A password manager’s blog post ranking business password managers was cited in an answer that never once said that company’s name, while naming three of its competitors. And a help desk company’s own listicle was cited, on its Dutch domain, in an answer to a United States question that named four other products and not that one.

In both cases the company paid for the content and the answer took the content without the credit. This is a real failure mode. It is simply not the common one here, and a tool that reports it as your main problem is describing a different market than the one these seven categories are in.

What this changes for a vendor

If your visibility report says you are mentioned in 60 per cent of answers, that number is telling you about the third party pages the assistant read, not about your site. Improving your own pages moves a lever that, in these categories, was pulled 6 times out of 48.

Three practical consequences follow, and the second is the uncomfortable one.

Mentions and citations need separate targets. They are moved by different work. Mentions are moved by what publishers and forums say about you. Citations are moved by whether you have a page worth reading on the whole category.

When a vendor’s page is cited, it is the category page and not the product page. That held 8 times out of 8 across both markets. What it does not tell you is that writing one earns the citation, and the section below is us checking that and finding it does not.

Ask your vendor which of the two they are reporting, and whether the number is a rate or a single draw. It is the fourth question worth asking a visibility tool, after whether its runs skip the cache, how many phrasings it tracks per intent, and what happens to its stability figure when the run count goes up.

It replicates in a second market

The first limit this piece published was that it measured one market. So the same design ran again in Spain, in Spanish, with fresh vendor lists frozen per category before each call. 24 observations, 4 usable cells, 3 categories.

Spain lands inside the American interval, and the asymmetry is if anything sharper: 13 named but not cited, against 0 cited but not named. The two markets are reported side by side rather than collapsed into one number, because two markets that agree is a replication and a single pooled figure would hide that there were two.

The pattern about which pages get cited held again. Both Spanish vendor domains that appear in a citation list are category comparison pages: one platform’s list of the best CRMs for business, and another’s guide to the best small business CRM. Across both markets that is 8 of 8, and not one is a product page.

The clearest illustration came from a vendor that was not even on a frozen list. In the Spanish invoicing answer, the single most cited domain belongs to a business management vendor, cited three times, plus its YouTube channel. All three cited pages are its own roundups: best ERP for small companies, best point-of-sale programs, and best software for heating and plumbing firms. That last one was cited in an answer about invoicing software.

Correction: we checked the base rate, and it retires our own advice

The observation that every cited vendor page is a category comparison page was selection on the outcome. If most vendors publish such a page anyway, the observation carries no advice at all. So we measured the base rate over the whole population of 33 vendor domains the study ever saw named, by reading each vendor’s own sitemap and counting URLs whose slug matches a category-list shape.

Among the vendors that were never cited even once, 18 of 22 readable sitemaps publish category list pages. That is 82 per cent, interval 61 to 93. The never-cited group publishes them almost as universally as the cited group, which is 5 of 5. At a stricter threshold the two move together rather than apart.

So the honest reading changes. What survives is which page gets cited: when an answer does cite a vendor’s own domain, it is the category page rather than the product page, 8 times out of 8. What does not survive is the implication we published earlier in this piece, that the comparison page is one most vendors refuse to write. Most of them write it. Only 4 of the 22 readable never-cited vendors publish nothing of that shape.

Three things about this check, because it is weaker than it looks:

• The measure is a slug shape in a sitemap, which is a loose proxy. A URL containing “best” need not be a category comparison at all. And the direction of that looseness matters: a loose detector inflates both groups and pushes toward finding no difference, which is what was found. A stricter measure could widen the gap, and we have not built one.
• 6 of the 33 domains served no readable sitemap and sit in their own bucket rather than counting as an absence. Scoring “we could not look” as “it does not have one” is a mistake worth avoiding.
• This cannot separate the obvious confound. Larger vendors publish more of every kind of page, and size plausibly tracks being cited on its own. Nothing here is causal, and the association is not even statistically distinguishable.

The reason it is published rather than dropped is that the design said it would be, whichever way it came out, and it came out against us.

What this does not show

2 markets, two languages, one day each, and one draw per question. These answers were read once each. We widened the sample by asking different questions rather than by repeating one, which answers a different question: not how often a given answer cites a vendor, but what fraction of questions of this type do.

The vendor lists are lists of well known names, chosen before each call precisely so the measurement could not select its own universe. That choice has a cost: the figure describes well known vendors. A brand an assistant does not already know may be more likely to need a citation in order to be mentioned at all, which would pull the true rate up. We have not tested that and it is the obvious next arm.

Two surfaces of four, and both belong to Google. AI Mode and ChatGPT were excluded on instrument grounds explained above. Anything here may look different on a surface we could not read.

The citation list is not the whole answer. In one cell two builder brands were linked inside the answer’s body while being absent from its citation list. Counting those inline links as citations moves the headline from 15 per cent to 22 per cent, interval 11 to 41. The direction of the finding does not change and the size of it does, so both are published and the headline uses the stricter reading.

A vendor’s own content is not always on a vendor’s own domain. Three Spanish observations are scored as not cited although a page that vendor wrote was cited, because it sat somewhere else: one hosting company’s tools site under a different name, another’s own first-person promotional copy hosted on a chamber of commerce domain, and a software vendor’s YouTube channel. They are scored strictly by the frozen rule. The direction of that choice is knowable rather than argued: any wider definition of a vendor’s own content can only add to the numerator while the denominator stays put, so the published rate is a floor.

And the direction is not universal. In a consumer marketplace category we have looked at separately, the common failure was the opposite one: pages cited constantly and the brand named rarely. Two markets, opposite gaps. Nothing here should be read as a fact about AI answers in general, only about business software questions asked this way on these two surfaces.

Frequently asked questions

Does being cited make you more likely to be named?

In this sample, yes, though the sample is small. Of 6 observations where a vendor’s own domain was cited, 4 also named the vendor, against 27 namings overall. With 6 observations that is a suggestive ratio and not an established effect, which is why it is reported as counts.

Why exclude ChatGPT and AI Mode?

Because on both, “cited” cannot be true through this route. ChatGPT returns no citation list at all, and AI Mode returns every citation with its domain wrapped in a Google redirect, so the publisher is not in the field a count reads. Including them would have added observations that could only land in the named-but-not-cited box, inflating the exact number the study set out to measure.

Is that rate low, or normal?

There is nothing to compare it against yet, which is why the interval matters more than the point. What can be said without a benchmark is the shape: 23 observations named but not cited against 2 cited but not named. The asymmetry is the finding, not the percentage.

Does this mean a vendor’s own pages do not matter?

No. It means they are not what these answers were built from. Your pages still decide what a buyer sees after the answer sends them looking, and they still decide what the comparison sites write about you. What the data argues against is treating your own site as the main lever on whether an assistant recommends you in these categories.

How would I check this for my own category?

Ask your category’s buying question on AI Overview and Gemini, write down which vendors the prose names, then write down which domains the citation list holds. Compare the two lists. If they barely overlap, the pages deciding your category are not yours. Decide your vendor list before you look, or you will end up measuring the vendors that happened to show up.