AI visibility: what it means and how it is measured

Updated 2026-09-02

"AI visibility" is used to mean whether an AI assistant can read you, whether it knows who you are, whether it cites you, and whether it recommends you — four different questions, answered by different mechanisms, measured by different methods, and routinely sold as one number. This page separates them so that a claim about any of them can be checked against what was actually observed.

The short version: an assistant answering a question either recalls from training or searches, fetches pages and reads them. Being readable is a property of your page. Being found is a property of the indexes. Being recognised as the right entity is a property of how consistently you are described across sources. Being cited is a property of one answer. Being recommended for a question that does not name you is the rarest event of all and depends on all of the above plus the competition. Each is worth knowing; none is a proxy for the others.

Five things people mean by AI visibility

1. Technical readiness

Can a fetcher get substantive content from your page? This is the only one of the five that is entirely yours to control and entirely deterministic: robots.txt policy for the search and user-tier crawler tokens, an HTTP status below 400, readable text in the response without JavaScript, a title and description, no soft 404, Person or Organization JSON-LD, a consistent name, a sitemap, and no noindex on the page. What measures it: fetching the page the way a plain crawler does and applying written rules — which is what the AI readiness check does, with every rule published. What does not measure it: asking an assistant whether it knows you, counting crawler hits, or opening the page in a browser, which runs the JavaScript most crawlers will not.

2. Search and index discovery

Is the page in the indexes answer engines draw on at answer time? The assistants that search use their own indexes (OpenAI's, Anthropic's, Perplexity's) or a partner's; being crawlable is necessary for inclusion but is not inclusion. What measures it, partially: whether the search-tier crawlers (OAI-SearchBot, Claude-SearchBot, PerplexityBot) actually fetch the page — Hilyt logs those reads for its profiles — and whether an assistant with search enabled retrieves it for a query that should surface it. What does not: Google's index. Ranking in Google proves Googlebot, which runs JavaScript, could read the page; it says nothing about crawlers that do not, as indexed by Google but invisible to AI sets out. OpenAI and Anthropic publish no API to their answer-time indexes; Perplexity's Search API (September 2025) is the one exception, so for the others "is it indexed" can only be inferred from fetches and from retrieval.

3. Entity recognition

Does the assistant know which entity you are, without conflating you with someone else of the same name, and does it describe something current? This is a property of consistency across sources rather than of any one page. What measures it: asking, with and without search, who a named entity is, and checking the answer against known facts — the name of the current venture, the right field, no merged biography. Hilyt's benchmark does this for one entity, its founder, as a keyword test on the answer, and says so. What does not: a closed-book answer that sounds confident. A model can recall a same-name person fluently, and the readiness check's model observation is reported as exactly that — closed-book recall, outside the score — because it cannot tell the two apart. AI is confusing you with someone else is the diagnosis when this fails.

4. Retrieval and citation

When an assistant searches for a question, is your page among the results it pulls, and does the reply name it as a source? Retrieval is the pool; citation is the subset the answer used. What measures it: an API surface with a search tool attached, a fixed prompt, and the retrieved and cited URLs recorded per answer — with the caveat that each vendor exposes a different view (OpenAI's Responses API shows only the sources the answer used, so there retrieval equals citation by construction). What does not: crawler hits. A read by a search-tier crawler means the page was fetched into an index; a read by a user-tier agent means it was fetched during an answer; neither says the answer cited it. Training crawls are further still: a GPTBot or ClaudeBot fetch is a corpus event, not an answer-time event, and cannot be traced to any reply. How AI decides which sources to cite explains the four filters a page passes to be cited.

5. Recommendation share

For a question that does not name you — "best tool for X", "who does Y in the UK" — are you in the answer, and where? This is the event people usually want and the one furthest from anyone's control: it depends on readiness, discovery, recognition and retrieval all holding, and then on what else is in the pool. What measures it: unbranded prompts sent to a surface repeatedly, with mentions and ranked positions recorded and, just as importantly, the products and sources named instead. What does not: branded prompts (asking "what is <your name>" tests recognition, not recommendation), a single run (answers vary between runs with nothing changed), or any number that pools this with the four above. Recommendation share is also close to meaningless for a person entity, where recognition is the question that matters.

What Hilyt measures, and for whom

Hilyt measures the first of the five for any site, and observes the third, fourth and fifth for its own brand only. The two are kept apart because they are different kinds of evidence — one deterministic and reproducible, the other observed and noisy — and a single score mixing them would be precise about nothing.

The AI readiness check is the readiness measurement. It fetches the submitted URL, robots.txt, llms.txt and sitemap.xml as a plain fetcher, identifying itself as HilytVisibilityCheck/1.0, and scores ten checks in five dimensions out of a denominator that is always 100. A check it cannot verify — an edge that challenged the fetcher, a host that did not answer — earns nothing and is reported as unverified rather than dropped. Alongside the score, and never counted in it, one model is asked closed-book who the submitted name is; the answer is quoted with the warning that it may describe someone else. Every weight, rule and limitation is written out at how the AI readiness score works, including the one condition the checker cannot equalise: hilyt.it recognises its own fetcher as a crawler and serves it the prerendered page, which most sites do not.

The AI visibility benchmark is the observation. A frozen set of nine prompts — category and problem questions that do not name Hilyt, plus one entity prompt that names the founder — is sent to four developer API surfaces with a search tool attached: the OpenAI Responses API with web_search, the Anthropic Messages API with web_search, the Gemini API with google_search grounding, and the Perplexity API's sonar model. For each answer it records whether Hilyt was mentioned, ranked in the requested top list, cited as a source, or retrieved into the pool, and which domains and products were named instead. Misses are published as misses; a surface whose search did not fire in a run is marked "not measured" rather than shown as zero; runs are appended and never overwritten. These are API surfaces, not the consumer apps, and with few runs the page is a snapshot, not a trend. We do not yet benchmark other entities — the observation exists so that Hilyt's own claims about the category can be checked against what assistants actually return, losses included.

What tends to make an entity resolvable

Recognition, retrieval and recommendation all rest on an assistant being able to resolve a name to one entity and trust what it finds. No one outside the vendors can state the rules, and they change; what follows is what tends to correlate, hedged accordingly and drawn from the guides rather than from any single test.

The order is roughly the order of effort, and the last item is the slow one. A page can pass every readiness check on the day it is published; corroboration accrues over months, from other people, and no host can grant it.

What Hilyt does not claim

Where to go next

AI-readable profiles · AI visibility · Run the AI readiness check · How the score works · Benchmark · Guides · hilyt.it