How we measure AI visibility

We count mentions in the answers we receive and calculate metrics and estimates of their uncertainty. Here is how.

Where the answers come from

We ask the questions to the public versions of ChatGPT, Google AI Mode, Gemini and Microsoft Copilot, logged out, the way a new customer uses them. That's not the same as the providers' developer APIs, which often return different answers and different sources.

Only Google AI Mode lets us set a country (we use your business's country). ChatGPT, Gemini and Copilot answer as if asked from the US, which is why the place is written into every question. Perplexity is not included at the moment.

When a business counts as named

A language model reads each answer and lists the businesses, brands and providers it names as options, in the order they first appear. ChatGPT tags businesses itself; we take those tags as they are. Your business counts as named when the answer names it, including under another name you gave us or your domain. Position is its place in that list.

The numbers

• Named in k of n answers: answers naming you divided by all answers received. Answers that didn't come back aren't counted.
• Likely range: a 90% Wilson interval. AI answers are samples: the same question rarely gets the same list twice. With 24 answers the range is wide. More comparable answers per check generally narrow it; weekly tracking lets you compare checks over time.
• Average position: only over the answers that name you.
• Statistical evidence of change: a two-sided pooled two-proportion z-test (no continuity correction) at the 90% level between two checks, i.e. |z| > 1.645. If a change doesn't pass it, we say it's within normal variation. Example: 3 to 6 of 24 answers stays below, 3 to 8 of 24 passes.
• Limits: the range and the test treat every answer as independent. Answers to the same questions are related, so read both as a good guide rather than an exact cut-off. Under these assumptions a single comparison flags a change about 1 time in 10 when nothing changed.
• Sources: websites an answer cites, counted once per answer. The grouping uses domain lists and the names of the competitors found. "Articles and other sites" is a catch-all that also includes unclassified websites.
• Share of all mentions: your mentions divided by all mentions of businesses in the answers, each business once per answer. Names you mark as "not a competitor" are left out.
• Recommended: answers that name you and also recommend or praise you, as the language model reads the tone of the answer. A mention on its own is not a recommendation.

What the assistants say about your business

Besides the customer questions, we ask each assistant once, directly, what your business is, what it offers, for whom, where and at what price. These answers don't count towards the numbers above. A language model picks out concrete statements and compares them with the text of your homepage and up to three other pages (such as About, Pricing or the legal notice): they match, the website says otherwise, or the website doesn't say. A statement that isn't on your website isn't necessarily wrong.

Whether the linked pages mention you

After the check we open up to 15 of the most-linked pages and look for your name or domain, and for the competitors found in the answers. We don't read pages behind a login, videos or social networks. A page we couldn't read is shown as not checked.

Structured company details to copy

From your homepage we put together a schema.org "Organization" block: name, address, logo, description, contact details and links to your profiles, using only what the page shows. If your page already has such a block, we only add what's missing. Check every line before you use it; the block doesn't guarantee being named.

The action plan

The suggestions are written by a language model from this check's measured numbers only. Every action names the fact it's based on. Nobody can guarantee you'll be named afterwards.