If you have run a free audit on WebValet you have seen two numbers: a human score and an AI score. The human score is familiar. It covers speed, titles, links and whether the page works on a phone. The AI score is newer and people ask about it most. This post explains what it measures, how it is calculated and what you can do about it this week.
The short version
An AI visibility score answers one question: when a customer asks an AI assistant something you should be the answer to, are you the answer? Not "could you be", not "are you indexed". Are you actually named, linked or quoted in the reply.
That sounds simple, but it depends on three separate things: whether the assistant can read your site at all, whether it understands what you do, and whether it trusts you enough to recommend you over someone else. The score blends those three.
Part one: can the assistant read you?
AI assistants get information in two ways. Some of it is baked into the model from training data. The rest comes from a live search at the moment you ask, which is how Perplexity, ChatGPT with browsing and Google's AI Overviews work. For the live part, the assistant has to fetch your page and make sense of it quickly.
This is where most small business sites lose points without knowing it. A page that loads its text with JavaScript after the fact may look fine to you and empty to a crawler. A page with no clear heading structure reads like a wall. A page that blocks unknown bots in robots.txt is invisible on purpose.
- Text in the HTML, not injected later. View the page source. If your main copy is not there, that is the first fix.
- One H1 that says what the page is, and H2s that read like questions a customer would ask.
- A robots.txt that does not block GPTBot, PerplexityBot, ClaudeBot or Google-Extended unless you have decided to.
- An llms.txt file at your root. It is a plain-text summary of your site written for assistants. WebValet hosts one for you on the free plan.
Part two: does it understand what you do?
Reading is not understanding. An assistant that fetches your homepage and finds "Welcome to a new way of working" learns nothing. One that finds "Acme Studio: a photography studio for rent in Dubai Marina, from AED 250 an hour" learns everything it needs to recommend you.
The check here is what we call a clear entity test. Within the first screen of your homepage, can a reader tell your name, what you sell, who it is for and where you are? If a human needs to scroll to find that, an assistant will too, and it may give up before it does.
Structured data helps a lot. This is a small block of code, usually JSON-LD, that states facts about your business in a format machines parse directly: name, address, opening hours, services, prices, FAQs. Google has used it for years for rich results. Assistants use it for the same reason: it removes guesswork.
WebValet checks for Organization, LocalBusiness, Product, Service and FAQ structured data on the free audit, and Autopilot can add the missing ones on Pro.
Part three: does it trust you?
This is the hard part, and it is the part no tool can fake. When an assistant chooses between two businesses, it leans on the same signals Google does: other sites that mention you, reviews, consistent facts across the web, and content that answers the question better than anyone else.
The practical measurement is a mention check. WebValet asks ChatGPT, Perplexity, Claude, Gemini and Google AI a set of questions your customers would ask, in your market, and records who gets recommended. If you are named, you get credit. If you are cited with a link, more. If a competitor is named instead, that is logged as an answer gap.
Answer gaps are the most useful output of the whole exercise. They tell you the exact question where you are losing, which means you know exactly which page to write or improve. On Pro, WebValet drafts that page for you.
How the score is put together
The free audit's AI score is a readiness score. It is based on what we can see on your site right now: whether an llms.txt exists, whether there is structured data, whether the homepage passes the clear entity test, how many headings are phrased as questions, and an estimate of citation likelihood from those signals.
On paid plans the score adds the live mention check, weekly on Starter and daily on Pro. That part is not a prediction; it is a record of what the assistants actually said. Over time you can see the line move as fixes go in, which is the only honest way to know if any of this is working.
- Readable: text in HTML, headings, no bot blocks (about 30% of the readiness score).
- Understandable: clear entity, structured data, llms.txt (about 40%).
- Answerable: question-style headings and FAQ coverage (about 30%).
- Mentioned: share of test questions where you are named or cited (paid plans only).
What moves the score fastest
If you only do three things, do these. First, rewrite the top of your homepage so a stranger knows what you do in one sentence. Second, add an llms.txt and FAQ structured data; both are an hour of work or one click in WebValet. Third, write one page that answers the single most common question customers ask you, in full, with a plain heading that matches the question.
Those three usually lift a readiness score from the 30s into the 60s within a crawl or two. Mentions take longer, typically four to eight weeks, because the live search results have to pick up the new pages first.
What the score does not measure
It does not measure brand sentiment, which is whether the assistant says nice things about you. It does not measure training-data presence, because that cannot be checked reliably from outside. And it does not promise traffic. AI answers often resolve the customer's question without a click, so the win is being the business they call, not the site they visit.
Treat the score as a checklist with a scoreboard attached. The checklist is fully in your control. The scoreboard tells you if the checklist worked.