Five things we tested that do not predict whether an AI assistant names you.

Key findings

• Across 112 cited pages and 128 uncited pages answering the same queries, cited pages had fewer words, fewer numbers, fewer lists and less structured data than the pages that were not cited.
• Structured data was less common on cited pages than on uncited ones: 70% against 80%, with FAQ markup at 20% against 23%.
• A cited page's best-matching sentence beat an uncited page's in only 10 of 26 comparisons - 38%, worse than a coin flip.
• Three separate keyword families totalling roughly 1,160 trade-specific phrasings returned no measurable search volume: 387 of 387 in one, 276 of 276 in another.
• Among 9 companies an assistant named, site size ranged from 24 to 565 pages, and only 3 of the 8 that could be judged had a page for the trade they were named for.

How this was measured

When

Five separate tests run between 29 and 31 August 2026.

Instrument

Each test has its own: page fetching and feature extraction for structure, sentence-embedding similarity for text matching, keyword volume data for the phrasing families, published sitemaps for site shape, and third-party estimates for authority.

Surface

Two surfaces, kept apart: Google's AI answers for the citation tests, logged-in ChatGPT for the naming tests. Results from one are never presented as results from the other.

Sample

112 cited and 128 uncited pages; 26 sentence comparisons over 5 queries; ~1,160 keywords; 9 companies.

Pre-registration

The authority test had its thresholds fixed in writing beforehand. The others were run as open comparisons against a matched control.

1. Page structure

We fetched the pages cited in AI answers and, for the same queries, a control set of pages that ranked but were not cited. Then we compared them on every structural feature people are told to optimise.

The cited pages were not better built. They were shorter (median 2,207 words against 2,707), carried fewer numbers (18 against 22), fewer lists (18 against 19), and virtually identical heading counts. Structured data - the thing most often recommended - appeared on 70% of cited pages and 80% of uncited ones. FAQ markup: 20% against 23%.

1 feature moved in the expected direction: cited pages had more question-shaped headings, a median of 4 against 3. That is the single positive signal in the whole comparison, and it is not enough on its own to explain anything.

2. Sentence-level text matching

The intuitive theory is that an assistant lifts the passage that most closely matches what it wants to say, so the page with the best-matching sentence wins.

We measured it. For each query we took the best-matching sentence from a cited page and the best-matching sentence from an uncited control page and compared them against the answer text. The cited page won 10 of 26 comparisons - 38%. The control pages' median similarity was slightly higher than the cited pages'.

Writing sentences that mirror the answer does not get a page cited. It performs marginally worse than chance in our data.

3. Per-trade phrasing and architecture

If owners searched for their trade attached to their problem, a page per trade would be the obvious build. Three separate keyword families were tested for that pattern, roughly 1,160 phrasings in total.

“AI visibility for ”: 276 of 276 registered nothing. “Why is my not showing up on Google”: 387 of 387 registered nothing. A third family covering lead generation per trade came back at negligible national volume.

The same owner who generates no searches at all for his trade attached to his problem generates thousands for his trade attached to a service he is shopping for. He does not describe his problem by trade. The honest caveat: a zero here means below the reporting floor, not proof that nobody ever types it - and it says nothing about what an assistant asks itself internally.

4. The size and shape of the site

We read the published sitemaps of 9 companies an assistant had actually named, to see what the winners are built like.

They are built like nothing in particular. The smallest had 24 pages; the largest 565 - a range of more than 20 times. The company named first in one answer had 24 pages and no page at all for the trade it was named for; its only trade pages are for a different profession entirely. Another was named for a trade while having no industry pages of any kind.

Of the 8 where it could be judged, only 3 had a page for the trade they were named for. Having the matching page is close to a coin flip.

5. Off-page authority

The last page-side explanation was that none of this is about pages at all - that assistants simply name domains that are already strong.

Among the same 9 companies, off-page strength ranged from one referring domain to 2,095. The weakest had 1 referring domain, 1 backlink and 20 organic keywords, and was named anyway.

That is set out in full, with the pre-registered thresholds, on our page about off-page authority.

What survived

5 explanations, all page-side or domain-side, all negative. What the cited pages did have in common was subject matter, not shape: each one held the specific factual content the question required.

On one query about pricing, every cited source was a page containing actual figures - a pricing survey, a cost calculator, a rates directory. Not the best-designed pages. The pages that owned the number the answer needed.

That is why this site publishes measurements instead of optimising a template, and it is the only hypothesis in this project that has not yet failed a test.

What this means if you are buying

• If a vendor is selling schema markup, FAQ blocks or answer-shaped paragraphs as the route into AI answers, our data does not support it - on two of those, the cited pages were slightly worse than the uncited ones.
• Building a page for every trade you serve is not what the companies who get named have done, and the search demand that would justify it is not measurable.
• The work that remains is unglamorous and hard to fake: own a specific, checkable fact that an answer needs, and be the page that states it plainly.

Honest limits of this measurement

• These are negative results at modest sample sizes. Absence of an effect at this scale is not proof that no effect exists.
• Each test measures one surface. The structure and text-matching tests were run against Google's AI answers; the site-shape and authority tests against a chatbot. We do not carry a conclusion from one to the other.
• The sentence-matching test rests on 26 comparisons across 5 queries. It is the thinnest of the five.
• A sitemap says what a company built, not why a model reached for it. 9 companies, no traffic data, no age, no revenue.
• Zero measured search volume means below the reporting floor of the data source, not proof that no human types the phrase.

Questions about this research

More measurements from the same instrument

• One of the companies an AI assistant named has a single referring domain
• An AI answer to a business owner cites 8.7 sources
• All seven measurements →

We run this measurement for a living. If you want it run on your own category and your own city, the free scan is the same instrument on one question, and the report is the full set. Nothing here is behind a form.

Every category we measure has its own page, with the questions customers put to an assistant in that category and what a reading costs. See all industries we measure →

Published 31 August 2026. Every figure on this page is recomputed from raw observation files by a single script, so any of them can be traced back to the readings behind it. If you are named here and believe a reading is wrong, tell us: we will run the measurement again and publish what it returns, whichever way it goes.