Skip to content
How we measure

How to measure AI visibility: how ScanFree measures what AI says about a business

What we ask, why twice, live search and training data, named and cited, what each answer really is, and what we do not know.

Every checkthe rules, in short
  • Questions a customer would ask, never your name
  • Each phrase asked twice on each AI
  • Live search and training data, never added together
  • Named and cited, counted apart
  • Every answer labelled with what produced it

What we ask

We ask AI assistants the questions a customer would ask, such as "Who is the best emergency plumber in Boise?". These phrases never contain the business's own name. Questions that do name it are checked separately, as a reputation question, and are never counted in the figures.

Twice per AI, every check

Ask an AI the same question twice and you often get a different answer. A BrightLocal study of local AI searches (September 2026) found only 20 to 33% average overlap between the result sets of repeat searches. So we ask each phrase twice on each AI at every check, and we count every answer. One ask is not enough to say what an AI "thinks"; two asks per check, repeated over time, show what it tends to say.

Live search and training data

Some answers come from an AI that searched the web while answering: we call these live search, and they come with sources. Others come from what a model learned in training, with no search: we call these training data. They can be out of date and have no sources. We report the two separately and never add them together.

Named, and your site cited

These are different things. Named means the answer says your business's name. Your site cited means your own website is among the sources the answer listed, or is linked inside the answer (including Google's official links). An answer counts once either way; under the figure we show how many cited you as a source and how many linked to you, and where in the answer the link sat (top, middle or end) when the AI's reply shows it. An AI can name you without citing your site (it may have read about you on a review site), and it can cite your site without naming you. We show both figures, each with how many answers it is based on: "Named in 38% of live answers (15 of 40)".

For live answers we list three things separately. Sources cited are the sources the AI showed for its answer, each website counted once. Links in the answer are links written inside the answer text, often a business's own website. Google listings are Google's own pages, such as Maps or a Business Profile. Only the first counts as a source. Ads are never shown.

With the keyword package, search volume (United States, the last 12 months), cost per click and competition come from Google Ads data; keyword difficulty, related keywords and your Google position come from a commercial keyword index that keeps its own copy of Google's results (the top 100, organic only), so a position is not a live search. Each run is dated and happens once per purchase. It is Google search data, not AI: we show it beside what AI says and never put it into an AI figure.

With the backlinks add-on, link counts come from DataForSEO's backlink index: which websites link to your site and to the competitors you chose, and which of the sites AI cited for your phrases are among them. Every figure is dated, and your figures and your competitors' come from the same index on the same day. An index is a sample of the web: a listing with no link found may still list you, and a nofollow link is still a link. We don't know how much links move AI answers, so we show them side by side with where AI looks, never as a cause.

What each answer really is

Every answer is labelled with what produced it. "ChatGPT app", "Gemini app" and "Google AI Mode" mean the real consumer app was asked, through a data provider that operates the app for us, in a fresh session with no history; the method line under each answer says what we know about that session. "Claude Haiku 5.5 (API with web search)" or "Perplexity (API with web search)" mean the maker's own programming interface with web search switched on; these can differ from what the maker's app would say. "OpenAI gpt-6-luna (training data)" means a model answering from training only. We never call an API answer "ChatGPT".

Normal variation

Answers move from check to check even when nothing changed. We work out the usual movement for each order, and call a change real only when it is bigger than that; otherwise we say "within normal variation".

Who counts as "named instead"

Only real businesses of the same kind as yours, which we could confirm exist. Directories, review sites and names we could not confirm are left out; we say how many were left out.

What we store

We keep the answers we collect, permanently: to improve ScanFree and for anonymous combined statistics. When an account is deleted, its business's name, website and questions are removed from them. For answers that come from Google's apps, we keep the names, counts, sources and dates, not the text.

What the evidence tags mean

Each recommendation carries one. Official: the AI company documents it, and we link the page and the date we read it. Seen in our data: our own answers show the link, which is not proof of cause. Likely: the reason is clear, but we have not measured the effect. Info only: for your information.

What we do not know

Nobody outside the AI companies knows exactly how they choose what to say. Studies, ours included, show what tends to go together, not what causes what. We cannot promise that any change will make an AI name your business.

Questions

How to track AI visibility?

Ask the same questions, in fresh sessions, on a schedule, and keep every answer with its date. We ask each phrase twice on each AI at every check, and call a change real only when it is bigger than the usual movement for that order.

How to check AI visibility score?

We don't make one. Each figure is a count with its total, such as named in 15 of 40 live answers, so you can see how many answers it rests on. Live search and training data are counted apart.

What is a good AI visibility score?

There is no standard score, so there is no number to aim for. More useful: who is named instead of you, on which questions, and whether your own count moves from one check to the next by more than normal variation.

How to measure AI exposure?

Count how often AI answers name your business, and how often they cite your website as a source, for the questions your customers ask. Named and cited are different things, and we report them apart.

Does AI still hallucinate in 2026?

Yes. Answers can get facts wrong or name businesses that may not exist. That is one reason we count a business as named instead only when we could confirm it is real and of your kind, and why we ask every question twice.

See it on a real report.

Our own site, checked the same way, zeros included.