How we measure whether AI assistants name you

What one measurement is

We take a set of purchase-intent questions, ask each one to four engines — ChatGPT, Gemini, Claude and Perplexity — with live web search switched on, and record every answer with the URLs it cited. Then we count how often your business is named, at what position, and who is named instead. Nothing is estimated from a smaller sample: the score is computed from the stored answers, so it can be recomputed without asking again.

The engines are the same everywhere; the number of questions and runs is what changes.

Where the questions come from

The questions are written from your own site — what you sell, where, and to whom — and, where we can get them, mixed with queries people actually type, taken from search autocomplete in your market.

They are spread across the buying funnel: about 30% early questions from someone who has a problem but no solution in mind, 40% from someone comparing approaches or providers, and 30% from someone ready to choose. Your brand name is never in the question. That is the one rule we do not break, because a question that contains your name measures whether the engine can look you up, not whether it recommends you.

Separately, a few questions do name you, to find out whether an engine knows who you are. They explain a gap when there is one, are reported on their own, and never enter the score.

What we count

• Appearance rate: the share of runs, across engines, in which you were named.
• Position: where in the answer you were named, when you were.
• Variance band: the lowest and highest appearance across the runs of the same question. A wide band means the result is close to a coin flip.
• Who was named instead, so a gap has a name attached to it.

Each question then gets one label. Defended means you appear in at least two thirds of runs and, on average, in the first three names. A consistency gap means you appear sometimes. A visibility gap means you never appear while a competitor does. An educational question is one where nobody is named at all — that is not a failure, so it is not counted as one.

If an engine is down or rate-limited on the day, it is excluded rather than counted as a miss. Every row of a report says how many engines actually answered.

How the score is built

The GEO score is out of 100 and has two parts. The AI Visibility Index, up to 60 points, comes from the live answers: how often and how prominently you are named. The Generative Readiness Index, up to 40 points, comes from crawling your site: whether AI crawlers may read it, whether the sitemap works, and whether the pages carry answers in a form an engine can lift.

Two things are recorded but deliberately not scored: your llms.txt file, and page-speed field data. We show them; they do not move the number.

What the method cannot tell you

• It is not what one person sees. Answers come from the engines’ APIs with web search on. They are close to, but not identical with, the consumer apps, where personalisation and memory differ.
• It is a point in time. Engines change without notice, which is why the full audit asks every question three times and monthly runs exist.
• One run shows whether you appear, not how reliably. The free audit asks once and says so.
• It guarantees nothing. Nobody controls these engines, and no measurement can promise a mention.

Our own crawler obeys your robots.txt, including the paths it disallows. If it blocks us entirely, the report says that instead of working around it.