AI visibility: what the number measures, and what ours is made of.

Ask five vendors for your AI visibility and you will get five numbers that do not agree, because each one asked different questions of different engines and weighted the answers its own way. That is not dishonesty; it is the absence of a standard. But it means a score you cannot inspect is a number you are taking on faith. AI visibility, as a metric, is simple to define and worth defining precisely — and then, because we sell one of those scores, this page prints ours. Not a description of it: the formula, transcribed from the code that runs it.

The definition

AI visibility is the extent to which AI assistants name, cite and feature your business when someone asks a question you should be the answer to. It is observed, not inferred: you ask the engines a fixed set of buyer questions and record what comes back. Three observations do most of the work — whether you were mentioned, whether your site was linked, and where in the answer you appeared — and a fourth, who was named instead, tells you what you are up against. The free report leads with exactly those:

Everything else — a score, a trend, a competitor gap — is a way of compressing those observations. The observations are the thing.

Why the scores disagree

Three choices sit under every AI visibility score, and vendors make them differently. Which questions:a fixed panel, a keyword list, or questions generated from the business's own profile. Which engines: the category meters engines as a product feature, so a score on three engines and a score on six are not the same measurement. How each engine is read: through a consumer interface, an API, or a model asked what it believes — and those produce different answers to the same question. Here is how we read each one, stated the way the product states it:

• ChatGPT — scanned through the live consumer interface, daily
• Gemini — scanned through the live consumer interface, daily
• Perplexity — answers captured daily through its API
• Google AI Overviews — read from live Google results
• Microsoft Copilot — read from Bing's answer layer
• Claude — model mode: how the model understands your brand long-term (labeled, never counted as a live capture)

The last one matters: a model asked what it thinks of a brand is not the same as a live answer a buyer would see, and we label it that way rather than counting it as a capture. A score that blends the two without saying so is measuring two different things and calling it one.

Our formula, from the code

What follows is transcribed from the scoring function in the scan engine, current as of this page's last update. Six components carry weight in the overall score; a seventh is measured and reported but excluded, for a reason given below. Each component is 0–100, the overall is the weighted sum, rounded and capped at 100.

The one we report but exclude: citation rate

Citation rate — the share of questions where the engine linked to your site — is recorded on every scan and shown in the report, and it carries no weight in the overall. The code's own comment gives the reason: it is too sparse. Engines mention businesses far more often than they link to them, so for most businesses the rate sits near zero and would swing the headline on noise. A link is still the strongest single signal an engine gives, which is why it is never hidden. It just does not get to move the number by itself.

The rule that stops a blank form from being worth 15 points

Competitor pressure is worth 15% of the overall, and it can only be measured if the project has competitors configured. With none configured every answer shows zero competitors, which would score a perfect 100 for leaving a field blank. So the code drops the term and renormalizes the remaining 85% to a full weighting instead. It is a small rule and it is the kind of thing a formula you cannot read would never tell you.

Why the score is stored in pieces

A single number hides which lever to pull. A 40 made of strong brand mentions and no coverage is a business that owns its own name and is absent from every problem and comparison question; a 40 made of thin mentions everywhere is the opposite problem with the same headline. So the seven components are stored separately on every scan, the scan records which engine version and which question set produced them, and the formula can change without destroying the comparability of the history — the components from a year ago can always be re-weighted. That is an architectural rule of the product, not a reporting preference.

How to read your own

• Components before the total. Which of the six is lowest tells you the work; the total tells you almost nothing.
• The verbatim answers before the components. Three answers, word for word, with every name highlighted, will tell you in thirty seconds whether the problem is that you are unknown, outnumbered, or mentioned late.
• Only compare against yourself. Same questions, same engines, next month. A score from another tool is a different measurement of a different thing.

And the reason it is a standing measurement rather than a one-time audit: 40–60% of the sources AI cites change every month (eMarketer, 2026). A score is a snapshot of a thing that moves every month.