How to measure AI search visibility — and what the numbers can't tell you

"Are we showing up in ChatGPT?" is now a question marketing teams get asked in board meetings. Answering it with a screenshot is tempting and almost always misleading: ask the same question again and the answer may change. This guide covers the metrics that make AI visibility measurable, how to read them without fooling yourself, and what no AI visibility number can tell you.

In short

• Measure across many prompts, several engines and repeated runs — never a single answer.
• The core metrics are mention rate, position, share of voice, sentiment and cited sources.
• Every score needs a confidence range; changes smaller than the range are usually noise.
• Visibility is a leading indicator. It doesn't show what one person sees, prove causation, or measure revenue on its own.

Why AI visibility needs a different kind of measurement

Classic rank tracking assumes one query, one results page, one position. AI answers break that assumption in three ways:

• Answers are generated. The same prompt can produce a different list on the next ask.
• Engines differ. ChatGPT, Perplexity, Claude and Gemini retrieve different sources and reason differently.
• Context differs. Language, location, the exact wording and the model version all change the answer.

So instead of a position, you're estimating a probability: across the questions your buyers ask, how likely is an assistant to recommend you, and how prominently? That calls for survey-style measurement — a fixed prompt set, several engines, repeated on a schedule — and for statistics that admit uncertainty.

The core metrics

Mention rate

The share of answers that mention your brand (or one of its aliases) at all. It is the simplest and most robust number. If you appear in 30 of 100 answers, your mention rate is 30%.

Position

When an answer contains a list, where do you appear? Being named first carries more weight than being named fifth, both for the reader and, often, for how the answer is summarised elsewhere. Average position is only meaningful across answers where you appear in a list.

A weighted visibility score

Mention rate and position are easier to track as one number. In Mentioned, each answer earns a weight — 1.0 if you're ranked first, 0.15 less for each place below (with a floor of 0.3), 0.6 if you're mentioned outside a ranked list, and 0 if you're not mentioned. The visibility score is 100 × the average weight across the latest answer for every prompt and engine in a rolling seven-day window. The methodology page documents it in full.

You don't have to use this exact formula. What matters is that the definition is written down, stable and applied consistently.

Share of voice

Your mentions as a share of all mentions of the brands you track. Share of voice puts your number in context: a 40% mention rate is excellent in a crowded category and worrying in a niche where one competitor appears in every answer.

Sentiment

Being mentioned as "expensive and hard to set up" is not the same as being recommended. Track whether answers describe you positively, neutrally or negatively, and read the answers behind any change.

Cited sources

Which domains do assistants cite when they answer your prompts? The most actionable view is the gap: sources that are cited in answers recommending your competitors but never in answers that mention you. Those are the pages most likely to be shaping the shortlist without you.

Per-engine breakdown

Report every metric per engine as well as overall. A drop that only happens on one engine usually has a different cause — and a different fix — than a drop everywhere.

Sampling, variance and confidence ranges

Because answers vary, any score is an estimate from a sample of answers. Two practical consequences:

1. More answers mean a more reliable estimate. More prompts, more engines and repeated samples of the same prompt all narrow the uncertainty.
2. You need a range, not just a point. Mentioned shows each score with an approximate 95% confidence range — "62 ± 6" — based on how much the individual answer weights vary. If last week was 60 ± 6 and this week is 62 ± 6, nothing has demonstrably changed.

A rolling window helps too. Mentioned uses the latest answer for each prompt and engine within the last seven days, so a single unusual answer is smoothed out by the rest of the set.

Reading the trend without fooling yourself

Some patterns are worth recognising:

• A step change on one engine, across most prompts usually means the engine changed — a new model, a new retrieval setup — rather than anything you did.
• A slow climb across engines after you shipped new pages or earned listings on cited sources is the pattern GEO work should produce.
• A jump right after you edited the prompt set is a measurement change, not a market change. Annotate it.
• One prompt flipping back and forth between "mentioned" and "not mentioned" is normal variance. Look at the set, not the single prompt.

Tie movements back to work you shipped. Mentioned's weekly recommendations record when each action was completed and measure the change in visibility since then, which makes this easier — but remember that correlation in a noisy system is a hint, not proof.

What the numbers can't tell you

Every honest AI visibility report should say what it doesn't measure.

What one particular person sees

Consumer apps may use different models, system prompts, personalisation, memory and location than the APIs used for measurement. Visibility tracking is a consistent, repeatable measurement of how the models answer — not a recording of what your prospect saw on Tuesday.

Coverage beyond your prompt set

The score describes the prompts you track. If your buyers ask questions you don't track, the score is silent about them. Review the set regularly — our guide to choosing buyer prompts covers how.

Causation

When the score moves, many things change at once: your content, competitors' content, the sources, the models. Visibility tells you what changed in the answers; working out why takes judgement, and the answers themselves.

Revenue

Being recommended is a leading indicator. To connect it to business outcomes, pair it with AI referral traffic, self-reported attribution ("how did you hear about us?") and pipeline data.

Perfect extraction

Deciding whether an answer mentions your brand is automated. Similarly named products, abbreviations and misspellings can trip any system up. Aliases help, and you should always be able to open the full answer behind a data point and check it yourself.

A reporting cadence that works

• Weekly: glance at the score, share of voice and any alerts. Read the answers behind any change larger than the range.
• Monthly: review per-engine trends, the top cited sources and gap domains, and progress on recommended actions.
• Quarterly: review the prompt set, competitors and segments; reset targets.

For a leadership slide, three numbers are usually enough: the visibility score with its range, share of voice against your top competitors, and the trend over the quarter — with one sentence on what you shipped and one on what you'll do next.

Checklist

• A written, stable definition of your score
• Metrics reported per engine and overall
• Confidence ranges shown next to every score
• Changes annotated when the prompt set changes
• Every data point traceable to the full answer
• Visibility paired with traffic and pipeline data

Want a first measurement? The free audit asks five buyer questions on two engines and scores the answers in about a minute.