Why an AI visibility scan checks 19 signals, not fifty

An AI visibility scan checks 19 signals because 19 is the number of capabilities that generative AI actually uses and that a crawler can establish from a site's public surface, with no account and nothing installed. Past that point, a list grows by guessing. A fifty-line report is not three times more useful than a nineteen-line one: it mixes what was verified with what was assumed, and never tells you where the line falls.

What makes a signal verifiable from the outside?

A signal enters the Biencité reference list on three conditions, and leaves it the moment one fails. First, it names a capability that a generative AI uses to read, cite or act on a site. Second, a crawler can establish it in read-only mode, from the home page and the public files at the domain root. Third, its absence is explained by the site itself, not by a failure of the measurement.

The third condition disqualifies more candidates than the other two combined. Plenty of criteria that circulate in GEO (Generative Engine Optimization) vocabulary are perfectly real and perfectly unobservable from the kerb: how fresh a catalogue genuinely is, whether a piece of writing is any good. A tool that claims to score those from outside is offering an opinion, not a measurement, and an opinion has no place in a number you intend to compare against itself six months from now.

Why does the list not grow with every fashionable new file?

Because a new file only becomes a signal once an engine or an agent actually reads it, and most are not there yet. Google Search Central's documentation on AI features, updated on December 10, 2025, is blunt about this: appearing in AI Overviews and AI Mode carries no additional requirement beyond ordinary search optimisation, no special schema.org markup to add, and no machine-readable file to create. The condition is still to be indexed and eligible to show with a snippet.

Google's position does not disqualify the files at the domain root. It puts them where they belong. A file such as llms.txt, or a capability card for agents, is not a way into Google's index: it serves the engines and agents that fetch a URL directly, without going through a search index. So a report that presents a root file as a Google requirement is selling you a worry that Google had already denied in writing.

What is a fine-grained signal worth if an AI cannot read your page?

Nothing, which is why a useful reference list stays short and ranked rather than long and flat. The measurement published by Vercel and MERJ on December 17, 2024 under the title "The rise of the AI crawler" found that none of the major AI crawlers execute the JavaScript on the pages they download: OpenAI's crawler fetched JavaScript files on 11.50% of its requests and Anthropic's on 23.84%, without ever running them. Gemini is the exception because it rides on Googlebot's infrastructure, and AppleBot because it uses a browser-based crawler.

For a site whose content only appears after JavaScript runs, the consequence is abrupt: the "content readable directly by AI" signal fails, and everything downstream becomes theoretical. A site can ship flawless structured data, and a crawler that sees an empty page still has nothing to quote. That is what the "invisible to AI" verdict describes, and it holds whatever the rest of the score says.

What does this reference list refuse to measure?

Four things, and naming them beats letting you assume they were covered.

• The editorial worth of your writing. A scan establishes that content is structured to be extracted, not that it deserves to be.
• Your reputation and inbound links. Both weigh heavily on whether you get cited, and neither can be read off a home page.
• What AI already says about you. That is the job of citation tracking, which asks ChatGPT, Claude, Perplexity, Gemini and Le Chat the same fixed questions every week.
• Your other pages. The scan reads the domain root and the public files, not a catalogue of a thousand product pages.

That separation is why a strong score and zero citations sit together without contradiction, as the article "A good score and no AI citations: what the gap tells you" sets out. A reference list that blended technical eligibility with reputation would produce a single number: easier to read, and impossible to act on.

Why three axes instead of one score?

Because "visible", "citable" and "ready for agents" describe three different problems, and three different places to spend money. A site can be perfectly readable by AI and remain uncitable, for want of a structure that lets a short answer be lifted out of it. Another can be readable and citable while exposing no capability at all to AI agents. A single overall score does not tell you which of the three you are.

The verdict does not follow from the score alone either. It follows a cascade, and a site that AI cannot read stays invisible even when its other two axes look good. The weight of each signal is not published, and that choice has a direct consequence for you: a Biencité score is comparable to itself over time, never to another site's. The 19 signals are public, the scale stays on the server, because a public scale invites you to optimise the number rather than the visibility.

Why does the report say how many signals were actually checked?

Because a failed measurement is not an established absence, and conflating the two manufactures problems that do not exist. An unreachable file or a server error drops the signal out of the calculation instead of scoring it zero, and the number of signals genuinely verified appears in the report. The article "Not measured is not failed: how to read an AI visibility report" covers how to read that coverage, and the full reference list is published at /en/methode.

The same logic applies to your platform. Shopify, Wix, Webflow and Framer lock certain files at the domain root. Those signals are marked impossible on your platform rather than counted against your site, and they never enter your action plan. A fifty-line reference list held by a tool that ignores this constraint would send you off to fix files you will never be allowed to write.

When will 19 signals, or fifty, be no use to you?

In at least three situations, all of them common. If your site holds very little content, making that content readable gives nothing more to quote: the spending that pays is writing the pages that answer your customers' questions. If your market runs on referral, and nobody is looking for you inside an AI assistant, the subject can wait a year without costing you anything. If you run a local business whose map listing is barely filled in, an hour spent on that listing will outweigh any file at your domain root.

Then there is the limit that holds whatever the number of signals: a reference list measures eligibility, not an outcome. It describes the conditions without which a citation is highly unlikely. No tool, Biencité included, can promise that an AI will cite your company.