How to check your website's AI visibility

What is AI visibility and how is it measured?

AI visibility is the share of answers in which a language model mentions your company when someone asks for a provider in your industry. You measure it by putting the same set of questions to several models and counting how often you appear and in which place on the list. The result is a percentage of mentions, not traffic and not a Google ranking.

The measurement has three parts. Share of mentions says how often a model names you at all. Model coverage says whether one assistant knows you or all of them do. Average position says whether you are the first or the eighth name. A company mentioned by a single model, always in last place, has measurable but weak AI visibility.

This distinction is not academic. A model answers with a few names, not ten links. If you are not among them, the user never learns you exist, and your traffic does not drop in a way analytics would show. What drops is the number of enquiries, and no visitor counter reports that.

Why AI visibility is not the same as a Google ranking

A search engine returns a list where second and third place still exist. A model answer works differently: it names a few providers and stops. There is no second page with a second chance. A site that ranks well on Google can have zero mentions in AI answers, and the other way round.

The reason lies in where the model gets its material. Part comes from the data it was trained on, which is what other people write about you: media, directories, articles, forums. Part comes from reading pages while it answers. You earn the first part with mentions away from your own site, the second with a site machines can read.

That is why the check has two halves. One measures what models say about you. The other measures whether they can read your site at all. A company that blocks AI bots in robots.txt and renders its content only in JavaScript loses the second half entirely, however good its classic optimisation is.

How to run the check by hand in half an hour

The procedure is simple, but it needs discipline or you will measure the wrong thing.

First, write eight questions the way a buyer would ask them, and never put your own company name in them. A useful query is "who in Slovenia services heat pumps for apartment buildings"; a useless one is "what do you know about company X". The second question measures whether the model knows your name, not whether it would recommend you.

Second, ask in the language of the market you are measuring. A question in Slovenian leads to Slovenian providers; the same question in English often leads to international chains.

Third, use a fresh conversation with no history and, where possible, no sign-in. An assistant that remembers your earlier chats will mention you more often than it would for a stranger, and your measurement will be kinder than reality.

Fourth, record every answer in two columns: mentioned yes or no, and the place on the list. Fifth name in an answer counts for less than first.

Fifth, repeat with three assistants, for example ChatGPT, Claude and Gemini. Eight questions times three models is twenty-four answers. Your share of mentions is the number of mentions divided by twenty-four.

What to check on your own site in ten minutes

The second half of the check is technical and needs no tools.

Open your domain with robots.txt appended and look for these bots: GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, Google-Extended and CCBot. Every blocked bot is one model fewer that can say anything current about you. The first two both belong to OpenAI and are not the same thing: GPTBot governs model training, OAI-SearchBot governs whether you appear in search inside ChatGPT. A site that allows the first and blocks the second looks correctly configured at a glance and is absent from ChatGPT search. Read the result carefully: only a 404 or a 410 means the file is absent. A 403 or a 503 means the server will not let you read it, which is a different thing.

Check whether an llms.txt file exists. It is a short index written for language models and saves the model from guessing which page matters.

Turn JavaScript off in your browser and reload the home page. What remains is roughly what a machine reads. If an empty frame remains, you are invisible to models whatever the quality of your content.

Look at the page source and find a JSON-LD block of type Organization or LocalBusiness. That block states your name, activity and address in a form a machine does not have to infer from prose.

Finally open sitemap.xml and confirm that it is readable and lists the pages you want found.

How a score from 0 to 100 is built

The score joins both halves into one number. Forty points go to the technical readiness of the site, sixty to actual presence in answers. The split by component is in the table below.

One detail separates an honest score from a cosmetic one. A component that was not measured is dropped from both the numerator and the denominator. A site that did not respond during the check earns no points for something nobody saw, and takes no penalty either. A component that was measured and came out badly scores zero: a missing llms.txt is a fact, not a broken measurement.

The same holds for position. If the test ran and no model mentioned you, there is no position, but that is a measured fact and scores zero. It would be dropped only if the test had never run.

Because the denominator is the sum of the weights of the measured components, the result always sits on a scale from 0 to 100, even when a check fails. The score also carries its own version number, so six months later it is clear whether your site changed or the formula did.

Why a manual check cannot be repeated

A manual check gives you the order of magnitude and is enough for a first look. The trouble starts the second time.

Models do not answer the same question identically every time. A single mention is an anecdote until you repeat it. To know whether three months of work improved anything, you have to ask exactly the same questions, of the same models, in the same language, and record the outcome the same way. A reworded question makes the comparison useless.

The second reason is time. Twenty-four answers by hand is roughly two hours of work with note-taking, every single time.

The third reason is impartiality. When you grade your own results, the temptation to explain away a bad round is far greater than with a number produced by a fixed set of weights.

What to look for in a tool

If you would rather not keep a spreadsheet, a tool is worth having only when it does four things. It asks the same questions on every run. It records which models answered and in which language. It shows the score broken down by component instead of one opaque number. And it versions the formula, so a change in the score means a change in your site.

Entexia runs exactly this check free of charge. You enter a domain and an industry; the analysis puts eight questions to three models, reads robots.txt, llms.txt, sitemap.xml, JSON-LD and the content available without JavaScript, and returns a score with the breakdown behind it. The form and the report run in five languages: Slovenian, English, Croatian, German and Polish.

What you get back is a list of weighted items, so it is visible which two fixes are worth the most.

What to do about a low score

Work through the fixes in order of weight, not of instinct.

Unblock the AI bots in robots.txt first, if they are blocked. That is one line and it opens the way for everything else. Then make sure the main content is in the HTML without running scripts. If your site stands on a framework that paints text in the browser, switch on server-side rendering.

JSON-LD and llms.txt come next. Both are small changes with a clear effect: the first says who you are, the second says which pages to read first.

Only then turn to content. Write the page that answers the question your buyer puts to a model, and put the answer in the first sentences rather than the conclusion. Add numbers that can be checked. A model would rather quote a page that says something concrete in its opening paragraph than one that works up to a claim through an introduction.

The last step is the slowest and the most valuable: mentions away from your own site. A model that finds one self-praising page cites it cautiously. A model that finds the same claim at three independent sources states it as fact.