AI visibility is measured with three separate counts per engine, not one score: how often you're named anywhere in an answer, how often you're recommended in the top three, and how often your own site is cited as a source. Track each one on its own, and compute a share by dividing by the number of eligible answers.
This is the practical follow-up to what an AI visibility audit should show you: an audit's Recognition probe is where these three counts actually get produced, and this article is how to read and compute them yourself.
Named, Recommended and Cited each answer a different question about the same AI response, and none of them substitutes for the others. Named asks whether your brand shows up anywhere in the text, even a single passing mention. Recommended is stricter: you have to be one of the top three products the answer actually suggests, not just referenced in passing. Cited asks something different again, whether the engine lists your own domain among its sources, separate from whether it recommends you at all.
| Count | What it checks | What a high count without the others means |
|---|---|---|
| Named | Does your brand appear anywhere in the answer? | You're on the radar, but maybe only as a runner-up or a comparison footnote |
| Recommended | Are you one of the top three actual suggestions? | AI trusts you enough to lead with you, the strongest signal of the three |
| Cited | Is your own domain listed as a source? | The engine used your content directly, even if a rival got recommended |
Bottom line: a brand can be Named often, Recommended rarely and still have strong Citations, three different pictures of the same brand's AI presence.
Because it averages three counts that move for different reasons into a single number that hides which one actually changed. Recommended reflects whether AI trusts you enough to lead with you; Cited reflects whether your content specifically got used; Named just reflects whether you were mentioned at all. Blend them and a brand whose Recommended rate falls while its Cited rate rises can show up as "no change," when in fact two very different things happened at once.
Proof: in the Reflexa AI Visibility Index's September 2026 project-management edition, brands mentioned by AI in 27-51% of eligible answers were cited on their own domain in only 6-18% of those same answers, a 21-33 point gap between the two counts for the same set of brands.
Why it matters: if you only track one number, you can't tell whether a drop is a trust problem (Recommended) or a content problem (Cited), and the fix for each is different.
Pick 10-20 real buyer questions in your category, run each one three times per engine, then divide each count by the number of eligible answers. An eligible answer is one where the question didn't already name your brand outright, questions like "what are the best alternatives to [you]" create an artificial opportunity and should be excluded from your own denominator, the same rule the Index uses.
Worked example: the Index's September 2026 edition asked 100 fixed questions 3 times across 3 engines (900 answers) and, after excluding the 18 answers to two Asana-anchored questions, Asana appeared in the top three recommendations in 359 of its 882 eligible answers, 359 ÷ 882 = 40.7%, reported as 41% Recommended (Reflexa AI Visibility Index, September 2026).
Do this: run the same fixed question set every time you re-measure, changing the questions between runs makes month-over-month movement meaningless.
The same method, run on a fixed set of 100 buyer questions per category, three times each across ChatGPT, Claude and Perplexity, for 900 observed answers per edition. Each answer is scored the same three ways, Recommended for top-3 inclusion, Mentioned for any appearance, Cited when the engine's sources list the brand's own domain, and anchored questions are excluded from that brand's denominator so no brand's number is inflated by a question that already gave away the answer.
Shortcut: the full method, including the six buyer-intent types behind the 100 questions, is published on the AI Visibility Index page if you want to mirror it exactly.
None of this requires buying anything, and each step takes a few minutes:
Bottom line: you can get a first read on all three counts yourself in under 20 minutes, before running anything at Index scale.
With three separate counts per engine: Named (you appear anywhere in the answer), Recommended (you appear in the top three products the answer actually suggests) and Cited (the engine lists your own domain as a source). Each is tracked on its own, never blended into one score.
Named means your brand shows up anywhere in the answer, even a single passing mention. Recommended is stricter: you have to be one of the top three products the answer actually suggests. A brand can be named often but rarely recommended.
Because it hides which count moved. A brand's overall share can hold steady while its Recommended rate falls and its Cited rate rises, two opposite trends that a single blended number averages into "no change."
Pick 10-20 real buyer questions in your category, ask each one 3 times per engine across ChatGPT, Claude and Perplexity, then count how many eligible answers name you, recommend you in the top three, and cite your domain. Divide each count by the number of eligible answers.
100 fixed buyer questions per category, asked 3 times each across ChatGPT, Claude and Perplexity with live web search, for 900 observed answers per edition. Each answer is scored Recommended, Mentioned or Cited, and questions that name a brand outright are excluded from that brand's own denominator.
The free check runs Named, Recommended and Cited across ChatGPT, Claude and Perplexity, 3 minutes, evidence included.