There is no shared canon: 81.6% of the domains AI engines cite are cited by exactly one engine

Method, stated before the numbers

Twelve pinned buyer questions — “What are the best AI visibility tools for small businesses?”, “How can I check what ChatGPT says about my business?”, “Is there a one-time AI visibility audit instead of a monthly subscription?” and nine more — each run three times against four engines (OpenAI, Perplexity, Google Gemini, Anthropic Claude). That is 144 answers, all from the production engine a paying customer's audit runs on. The run completed 2026-07-28; no answers failed.

The citations are read from each engine's returned citation array, not from a substring search of the answer text. That distinction is the whole reason this cut exists, and we will say plainly why: our own previous instrument tested for a domain string inside the prose of the answer. That test could not see a citation an engine attached as structured metadata rather than typing into its reply, and on three of our four engines that is exactly how citations arrive — so the old number was blind on three engines and we reported it anyway. If you are reading somebody's AI-visibility number, ask which of the two things they measured.

One more thing an honest method line has to carry: the four engines do not return citations in the same shape. OpenAI, Perplexity and Anthropic return full URLs — 1,020 of them here. Gemini's grounding metadata returns bare domains (`salesforce.com`, not a deep link) — 445 entries. So the domain-level analysis below runs on all 1,465 entries, and any URL-level or page-level analysis can only run on the 1,020. A domain is counted once per answer no matter how many times that answer links it, so the unit is answers-citing-domain, never links.

How much do the four engines actually agree?

Pairwise, the overlap is just as thin. Measured as a Jaccard index over each pair's distinct-domain sets, the most similar pair of engines in this run is Claude and Perplexity at 0.189 — meaning fewer than one domain in five appears on both. The least similar is Gemini and ChatGPT at 0.062. Every pair: Claude×Perplexity 0.189, Gemini×Perplexity 0.164, Claude×Gemini 0.092, ChatGPT×Perplexity 0.093, Claude×ChatGPT 0.084, Gemini×ChatGPT 0.062.

A single measurement against a single engine is therefore not a measurement of your AI visibility. It is a measurement of your visibility in one engine, and on this evidence it predicts almost nothing about the other three.

The ten domains every engine cited

In alphabetical order: aiclicks.io, airops.com, alhena.ai, g2.com, layer3labs.io, rankability.com, searchengineland.com, semrush.com, therankmasters.com, useomnia.com. Two of the ten are what you would expect — a review marketplace (G2) and a large SEO platform (Semrush) — and one is a trade publication (Search Engine Land). The other seven are AI-visibility startups, most of them younger than the category. Being cited by all four engines is not a function of size or age. It is, on this sample, a function of having published a page that directly answers the question.

Per-engine breadth, and how many answers carry citations at all

Gemini cites the widest field — 280 distinct domains, nearly twice Perplexity's 159 — while returning fewer entries than Perplexity does, because it repeats itself less. Claude is the narrowest on both counts and is the only engine that frequently answered with no citations at all: 11 of its 36 answers carried none, against 0 for ChatGPT and Perplexity. If you are invisible to Claude, part of the reason is that Claude cites less often in the first place.

The most-cited domains

The single most-cited domain in the entire run is reddit.com, on 34 of 144 answers — more than any vendor, any publication and any tool. YouTube is second. Neither is a place a business can publish a specification-grade page; both are places its customers talk. That is a finding with a cost attached, and we will not pretend otherwise: the two surfaces the engines reach for most are the two an ordinary business controls least.

Our own number: 0 of 144

askedabout.com is cited in zero of the 144 answers and appears in none of the 1,465 citation entries. Across the same corpus the engines named Peec AI in 34 answers, Profound in 23, Semrush in 22, AthenaHQ in 16 — and Otterly, counted across every spelling variant it appears under, in 61 of 144, more than any other tool in the category and more often than the engines named ChatGPT itself (42). We publish our own zero every week, and this week we also did the thing that makes a zero worth publishing: we ran a positive control against the exact code path that produces it — the literal URL formats an engine returns, upper-case hosts, `www.` prefixes, deep paths with query strings, our domain buried among twenty others, and three look-alike negatives such as `askedabout.com.evil.co`. It passes on every case, and a deliberately broken copy fails. So the sensor can fire. It fired on nothing for us. That is a fact about our visibility, not about our tooling — and the difference between those two sentences is the only reason the number is worth reporting.

What a business should take from this

• Measure per engine, not in aggregate. With 81.6% of cited domains appearing on exactly one engine, a single blended “AI visibility score” averages away the only variable you can act on.
• Ask what any tool actually reads. A number derived from the presence of your brand name in the answer text is a different measurement from one derived from the engines' citation arrays, and the two disagree systematically — ours did.
• Repeat the measurement. Three samples of the same question against the same engine do not always return the same domains. One run is an anecdote.
• Expect the shape, not the ranking, to be stable. Which specific tool sits at position three moves week to week; that four engines barely share a canon has held every time we have measured it.

The full ranked domain list behind this post — all 538, with the engine set for each — is a first-party cut we are happy to share with anyone who wants to check it. You can run the same measurement on your own business in 60 seconds, or read which engine recommends what for the vertical-level version of the same data.