AI visibility score: what it can measure and how ours is calculated

An AI visibility score is a weighted average of checks run against your pages, and it measures readiness rather than citations. Ours runs 36 checks worth 87 weight points, grouped into eight categories, and every weight is published below. A score built this way tells you whether an assistant that reaches your page can read it. It cannot tell you whether any assistant reached it.

The two halves, and why most scores are one of them

There are two genuinely different questions hiding under one phrase.

The first is whether your pages are legible to an assistant: can its crawler fetch them, does the content exist without JavaScript, is there a date, an author, a clean answer it can lift. That question is answerable from the page itself, deterministically, in a few seconds.

The second is whether assistants actually name you when somebody asks a question you should be an answer to. That is answerable only by asking them, repeatedly, and recording what comes back.

Almost every product sold as an AI visibility score answers the first question. It is the useful and cheap one, and the trouble starts when the number gets described as though it answered the second. A page can score 98 and be named by nobody, because being named depends on how much of the web talks about you, which no page level audit can see.

The weight table, in full

Every check returns a score between 0 and 1 and carries a fixed weight. The overall score is the weighted mean, expressed as a percentage. Checks are grouped into eight categories, and the category totals below are what actually determine how much each area moves the number. Read from the engine on 6 August 2026.

Category	Checks	Weight	Share of the score
Crawlability and indexing	7	19	21.8%
AI and GEO	6	19	21.8%
On page SEO	7	16	18.4%
Performance	6	12	13.8%
Content	4	7	8.0%
Security	2	5	5.7%
Media	3	5	5.7%
Structured data	1	4	4.6%
Total	36	87	100%

The individual weights are more revealing than the categories. Three checks carry a weight of 5, the heaviest in the engine: whether the page is reachable at all, whether it is indexable, and whether the AI crawlers are allowed to fetch it. Five checks carry a weight of 4: the title tag, the render gap, structured data, Core Web Vitals, and chunk quality, which measures whether a clean passage can be lifted out of the page.

Everything else sits at 1, 2 or 3. Caching headers, image loading, readability and text ratio are each worth a single point out of 87, which is a deliberate statement that they are worth knowing and not worth a project.

Three decisions inside the arithmetic that change the number

A check that cannot run is excluded, not failed. If the Core Web Vitals check has no data, its weight leaves the denominator entirely rather than scoring zero. The alternative punishes a site for our missing measurement, which produces a lower number that is not about the site. It also means two pages can be scored out of different totals, which is worth knowing when comparing them closely.

The bands are set so an A is a floor. 90 and above is an A, 75 to 89 a B, 60 to 74 a C, 40 to 59 a D. When we ran this engine over 98 agency websites in this category, the median across the ones we could fetch came out at 87 and nothing landed below 75. An ordinary competent site sits in the B band, which means an A is what a site with no real defect looks like rather than a distinction.

Access checks dominate on purpose. Reachability, indexability and crawler access together are 15 of 87 points, and they are close to binary in practice. A site that blocks GPTBot loses more from that one line in a robots file than it could gain from every content check in the engine combined. That ordering is a claim about how citation actually works, and it is the weighting most worth arguing with.

The arithmetic on one page

Worth doing once by hand, because it shows how insensitive a weighted mean is and stops you reading small movements as progress.

Take a page that scores perfectly on everything, then break one thing. Score zero on structured data, the only check in its category, and the page loses 4 of 87 points and lands at 95. Score zero on both freshness and attribution, the two checks this category most often fails, and it loses 4 points and lands at 95 again. Neither moves the grade. A page can be undated, unattributed and unmarked up and still show a 95 in a report, which is why the per check list matters more than the total.

Now break access instead. Score zero on reachability, indexability and AI crawler access, which is 15 points, and the same page falls to 83 and a B. Lose the whole AI and GEO category, 19 points, and it lands at 78, still a B.

Two things follow. A single check almost never changes a grade, so chasing a total is a poor use of a week. And nothing in the engine can drop a page into a D except a failure to fetch or index it, which is the ordering we intended: everything else is a matter of degree, and being unreachable is not.

What a score cannot tell you

It cannot tell you your competitors are worse. Ours is a page level audit with no comparative element in the number at all.

It cannot tell you what an assistant will say about you. That depends on the answer side, and the two correlate loosely at best: a technically excellent site with no third party coverage is invisible, and a mediocre site that a popular roundup lists gets named constantly.

It cannot tell you the score will hold. Every one of these numbers is a measurement of a page at a moment. Ours carry the date they were taken for that reason, and a score published without one should be read as an anecdote.

How to measure the other half

Ask the assistants directly, and record it in a form that survives to next month. The AI visibility audit prompt is the version of this we use: it makes each assistant answer a buyer question normally, then transcribe what it named, in what position, and on what basis, into fixed fields you can put in a sheet.

Two rules make the difference between a measurement and an impression. Use a fresh conversation for every question, because an earlier answer becomes context and contaminates the next one. And run every question at least twice, because these systems are not deterministic and the same question twenty minutes later can name a different set of companies.

Do both halves and the numbers start to explain each other. A high readiness score with no mentions means the problem is authority rather than pages. A low readiness score with mentions means you are being described from somebody else’s page, which works until that page changes.

Score your own page

The engine described here is the one behind our free audit, and the weights above are the weights it uses. Run a page through it and you get the per check results rather than only the total, which is the part worth reading: audit a page, or check whether an assistant can lift a clean answer out of it with the AI content readiness check.