How to measure AI search visibility

A check is one question, asked to one assistant, in one country, once. Every metric below is computed over a set of checks.

Read them together. High visibility with a poor average rank means you’re on the lists but rarely the top pick.

Sample data for a fictional journaling app:

• 15 questions × 3 assistants × 3 runs each × 1 country = 135 checks.
• Named in 41 checks → visibility 41 ÷ 135 = 30%.
• Its 41 positions average to 2.4.
• Across all 135 answers, assistants named 540 products in total (about 4 per answer). 41 of those were this app → share of voice 41 ÷ 540 = 7.6%.

Same app, two different stories: it’s recommended in almost a third of answers, but it holds a small slice of all mentions because answers list several apps each.

Assistants sample their wording, search results shift, and models get updated. The SparkToro and Gumshoe study (2,961 responses to 12 prompts) found ChatGPT and Google’s AI repeated the same brand list less than 1 time in 100, while appearance rates were much steadier. The authors suggest measuring visibility rather than exact position.

To tell signal from noise, start from the standard error of a proportion, then remember that comparing two weeks doubles the variance:

SE (one scan)        = √( p × (1 − p) ÷ n )
SE (week vs week)     = √2 × SE (one scan)

p = 0.30, n = 135  →  SE ≈ 0.039 (±4 pts)  →  difference SE ≈ 0.056 (±6 pts)

In the example, a move from 30% to 34% is well inside one standard error of the difference: probably noise. A move to 45% is about 2.7 standard errors: likely real. Doubling to 270 checks brings the difference error to about ±4 points.

These numbers are optimistic. Runs of the same question aren’t independent (they share a prompt and often the same search results), so the real noise is larger than the formula says. Treat it as a floor, and look for changes that hold across two or more scans.

There’s no standard AI share-of-voice formula. Common versions:

• Mentions-based: your mentions ÷ all product mentions (MentionReach uses this).
• Response-based: responses naming you ÷ all responses in the set. This is effectively visibility.
• Response share among tracked brands: responses naming you ÷ responses naming any tracked brand.
• Volume-weighted: Ahrefs weights each prompt by estimated demand, so popular questions count more.

When comparing tools or reports, check the denominator first. The same data can give 7.6% or 30% depending on it.

Visibility tells you whether you’re named. Citations tell you why. For each cited page, count how many answers cited it and how many of those named you. A page cited 12 times that named you 0 times is exactly where to focus. See how to get cited for what to do next, or see what a scan costs.

• Per assistant: ChatGPT, Gemini and Claude run different search stacks and often cite different pages. A blended number hides that.
• Per country: OpenAI’s web search accepts an approximate user location, and local results differ. Test the markets you sell in.
• Per question: one question at 0% while the rest are fine usually points to one missing page, not a brand problem.