Beta GEONALY is in open beta. Found a rough edge? Tell us.
Blog · Measurement
Share of voice in AI search is your share of the brand mentions across a fixed set of prompts, assistants and runs. This post gives the formula we use, a worked example with the arithmetic shown, the two ways the number can mislead you, and how it differs from share of search.
26 of 114 brand mentions are yours. 20 prompts, four assistants, one run each. The arithmetic is below.
Key facts
Share of voice in AI search is your brand's share of all the brand mentions in the answers to a fixed set of prompts, run in a fixed set of assistants a fixed number of times. GEONALY counts it across ChatGPT, Claude, Gemini and Perplexity and reports it next to mention rate and citation rate, which answer different questions.
Ask an assistant which managed hosting to pick for a small company in Czechia and it answers with a shortlist. Three brands, sometimes five, occasionally one. Share of voice is your share of those named brands, counted across many such answers.
Three words need pinning down first. A mention is your brand named in the answer text. A citation is a link to one of your pages as a source. A position is where you sit in the shortlist when there is one. Share of voice, as we count it, is built from mentions. Citations and position are tracked next to it, not folded into it.
The phrase has an older meaning in advertising, where share of voice is usually a brand's share of category ad spend. And it is different again from share of search, which is a brand's share of branded search volume in its category. That one gets its own section below.
Fix three things before you count anything: the prompt set, the assistants, and the number of runs. Then:
The denominator of share of voice is every brand named across the answers, yours included. It is an open set: whoever the assistants name is in it. That is deliberate. The competitor you did not list is the one the assistant is recommending instead of you, and a closed list would hide exactly that.
Mention rate uses a different denominator, the number of answers. It tells you how often you appear at all. The two can move in opposite directions in the same month, which is why we report both.
Take 20 buying questions for one market, run once in each of four assistants. That is 80 answers. Suppose the tallies come out like this. The figures are illustrative.
| Assistant | Answers | Name you | Other brand mentions | All mentions | Share of voice |
|---|---|---|---|---|---|
| ChatGPT | 20 | 9 | 21 | 30 | 30.0% |
| Claude | 20 | 6 | 21 | 27 | 22.2% |
| Gemini | 20 | 7 | 23 | 30 | 23.3% |
| Perplexity | 20 | 4 | 23 | 27 | 14.8% |
| All four | 80 | 26 | 88 | 114 | 22.8% |
Share of voice across the four is 26 divided by 114, which is 22.8%. Mention rate is 26 divided by 80, which is 32.5%. Read them together: you appear in a third of answers, and when brands are named you are about a fifth of the names.
The per-assistant split is not decoration. Perplexity names you in 4 of 20 answers while ChatGPT names you in 9 of 20. Those are different problems with different fixes, and an average of the four would hide both.
If a new competitor starts appearing in answers next month, your share falls even if your own mentions are unchanged. This is the same trap we hit with competitor sets in share of search, where one rival with three times the branded searches changed a client's picture overnight. Always report the raw counts next to the share, so a falling share with steady mentions is read for what it is: someone else rising.
Assistants do not give the same answer twice. Their web results change, and the models themselves sample. Semrush's 230,000-prompt study saw Reddit's share of ChatGPT's top 25 cited domains move from about 60% to about 10% within six weeks1. With 114 mentions in the example above, each mention is 0.9% of the total, so a single extra mention moves share of voice by most of a point: about 0.7 points if it is yours. Below about a hundred mentions, monthly swings of a few points are noise. Run more prompts, run them more than once, and read a rolling average.
| Metric | What it adds |
|---|---|
| Mention rate | A share can be read against how often you appear at all. |
| Citation rate | Being named without a link means the assistant learned about you from someone else's page. |
| Average position | Being named sixth of six counts the same as first in a plain share. The GEO paper weighted visibility by position for this reason3. |
| Split by assistant and by market | The four assistants disagree, and the markets disagree more. |
| Competitors named instead of you | Listed by count. The most useful line in the report, and the one most tools leave out. |
They sound alike and measure opposite ends of the funnel.
| Share of search | Share of voice in AI search | |
|---|---|---|
| What it measures | Your share of branded searches in the category | Your share of brand mentions in assistant answers |
| Where the data comes from | Search volume, monthly | Prompts you run, on your schedule |
| What it tells you | How many people already want you by name | Whether an assistant puts you in front of people who do not |
| Evidence behind it | Tracked market share with a lead of up to a year in cars, in the IPA study by Les Binet2 | New. No published study links it to market share yet. |
We report both because they answer different questions. Share of search says how much demand you own. Share of voice in AI says whether you are in the answer when someone with no brand in mind asks. When the first is high and the second is low, the assistant is handing your category to somebody else.
Published 21 September 2026. If a model or a figure above changes, this line says what changed and when.
Short answers. The detail is in the sections above.
There is no published benchmark. Compare it with two things you can measure: your share of search in the same category, and your own number last month. A share of voice well below your share of search means the assistants under-represent a brand people already look for.
Enough that one mention barely moves the share. In the worked example above, 80 answers produced 114 brand mentions, so each mention is 0.9% of the total and one more mention of the brand moves its share by about 0.7 points. Twenty prompts per market, run in four assistants, is a workable floor. Repeat runs tighten it further.
We count a citation separately. A link to your page is recorded as a citation whether or not your name appears in the text. A mention is your name in the answer. The two rates are reported side by side.
ChatGPT, Claude, Gemini and Perplexity, each queried and reported separately. The four disagree often enough that an average across them would hide the useful part.
The AI Visibility trial runs on your real brand for 7 days, with no card and no call. The first answers are in it when the trial opens.