AI share of voice: how to measure it, with the formula and a worked example

AI share of voice is the portion of brand mentions in AI-generated answers that go to your brand, compared with the competitors you track. In plain words: ask a fixed set of buyer questions, count how many times each tracked brand is named in the answers, and divide your count by the total, separately for each AI engine and each market, over repeated runs.

The AI share of voice formula

These are the two category-convention measures, and they answer different questions. Visibility rate (often called mention rate) asks how often you are named at all. Mention-share SOV asks how much of the naming goes to you rather than to your rivals.

The two move separately, which is why you want both. In a crowded category where answers name four tracked brands on average, you can be named in 60% of answers and still hold only 15% share of voice (60% ÷ 4). Adding a competitor to your tracked list leaves your visibility rate untouched and lowers your share of voice, because the denominator grew. So a share-of-voice figure always comes with a list of the competitors behind it, and that list has to stay fixed for two periods to be comparable.

Two neighboring measures are easy to confuse with these. Citation share counts the sources an answer links to, not the brands it names. Share of answer asks whether you are the brand the engine recommends, rather than one name in a list. More on that below.

A worked example, with repeat runs

Say you sell project management software in the US. You write 20 buyer prompts, run each one in ChatGPT once a day for seven days, in English, logged out, and track three competitors: A, B and C. That gives 20 answers a day and 140 answers for the week. For each answer you note which of the four brands it names, counting each brand once per answer.

The headline figures. Pool the week rather than averaging the daily percentages: your mention-share SOV is 44 ÷ 200 = 22%, and your visibility rate is 44 ÷ 140 = 31.4%. The same pooled arithmetic for every tracked brand gives the competitive picture, and the shares add up to 100% because the denominator is the same 200 mentions:

The range. Sort your seven daily share-of-voice readings: 16, 20, 20, 20, 25, 25 and 26.7%. The 25th percentile is 20% and the 75th is 25%, so the middle half of your readings sits between 20% and 25%. That span is the P25-P75 range. The honest way to report the week is "22% share of voice in ChatGPT, US English (P25-P75 20-25%)". Do the same for visibility rate: the daily readings 20, 25, 30, 30, 35, 40 and 40% give a range of 27.5-37.5% around the pooled 31.4%.

Reading next week. If next week comes in at 25% with a range of 22-28%, the two ranges overlap heavily, so the three-point rise may be noise rather than progress. A change you can stand behind is one that moves clear of the range you started from.

Why a single run misleads

Look at the example again. Day 3 on its own says your share of voice is 16% and you are named in one answer in five. Day 7 says 26.7% and two answers in five. Same prompts, same engine, same market, nothing changed on your side, and an eleven-point swing. A screenshot of either day would have told a confident story, and the two stories would contradict each other.

That swing is not an artifact of made-up numbers. It is how AI engines behave, which is called answer variability. In our native-language study (October 2026), we asked the same coffee buying questions twice, in the same language, on the same day. In Germany, two runs of the same question on Gemini, and on Microsoft Copilot, had a median of only 38% of named brands in common (brands named in both runs divided by brands named in either). Across the markets and engines we tested, that figure ranged from 20% (Google AI Mode in Japan) to 78% (Perplexity in Japan), and ChatGPT in Germany sat at 28%.

So repeat runs are not a refinement. They are the measurement. One run is a single draw from a range of possible answers. Seven daily runs give you a pooled figure and a range that shows how settled it is. Thirty give you a figure that one odd answer can barely move.

"A share of voice without a range is a screenshot. With a range, it is a measurement."

Want to see how often AI engines name your brand right now? Get a free visibility audit and see where you stand across ChatGPT, Gemini and Perplexity.

Per engine and per market, never blended

Compute share of voice separately for every engine and every market, and keep the rows separate. A rise on ChatGPT and a fall in Google's AI Overviews can average to "no change" while your buyers, who mostly use one of them, see a real shift. If you need a single number for a slide, build it as an explicit roll-up of the rows, weighted by where your buyers actually are, and show the rows beneath it. For a B2B brand that might lean on ChatGPT and Perplexity; for a consumer brand, Google's AI Overviews may deserve more of the weight.

Language matters even more than engine. In the same study we asked each question in the market's language and in English, from the same country, so the only thing that changed was the language. Across Germany, France, Japan and Brazil, the local-language and English answers had between 0% and 21% of brands in common (median per question), and on every market and engine we tested that was lower than between two runs of the same question in one language. In Germany on Perplexity, for example, 11% of brands were shared between the German and English answers against 74% between two German runs. In France on Perplexity it was 4% against 64%.

The sources shifted too. In Germany, Perplexity's German answers drew a median 85% of their citations from .de sites; its English answers, 5%. An English-language panel run for the German market is therefore measuring a different shortlist from the one German buyers see. Write each market's prompts in the language its buyers use, and treat each market as its own measurement.

How major tools define share of voice

Two numbers both labelled "share of voice" can be computed differently, so a figure from one tool is not automatically comparable with a figure from another, or with your spreadsheet. Our comparisons of the field (the best AI visibility tools, Profound alternatives, Peec AI alternatives and tool pricing, checked 26 September 2026, with several vendors re-checked 17 October 2026) record each vendor's prices, engines and markets. They do not record how each vendor calculates share of voice, so we will not paraphrase anyone's formula here. Ask the vendor instead:

• What is the denominator? All mentions of tracked brands gives mention-share SOV. All answers gives a visibility rate. They are different numbers with different meanings.
• Which competitors are in it, and who chose them? Change the list and the share changes.
• Does a brand named three times in one answer count once or three times?
• Is it per engine and per market, or blended? If blended, ask how the engines and markets are weighted.
• How many runs sit behind it, and is there a range? A clean number with no spread is hiding the noise, not removing it.
• Are the prompts written in each market's language, or translated from English?

How to measure AI share of voice in a spreadsheet

You can do all of this by hand, and doing it once is the best way to understand the data before you automate anything. If you would rather see a first reading for your own brand without the spreadsheet, the free AI visibility checker runs a one-off version of this panel in your market and emails you the result.

1. Write the prompts. 20 to 50 questions a buyer would type, from broad ("best project management tool for a 10-person agency") to specific ("project management tool that syncs with Xero"). Leave your brand name out: a question that names you only measures whether the engine has heard of you.
2. Fix the competitor list. Pick the three to five competitors you want to be measured against and do not change the list mid-series.
3. Fix the setting. One engine, one country, one language, logged out or in a clean profile. A German answer and an English one are different measurements.
4. Run every prompt once a day for at least a week. Paste each answer into an Answers tab, one row per answer: Date, Engine, Market, Prompt, Answer ID. In Google, log only the searches where an AI Overview actually appeared, and note how often it appeared as a separate figure.
5. Log the mentions. In a Mentions tab, add one row for each tracked brand an answer names, once per answer: Date, Engine, Market, Answer ID, Brand. An answer that names none of your tracked brands still stays in the Answers tab, because it still counts toward visibility rate.
6. Compute visibility rate. With Engine in column B and Brand in column E of Mentions, and Engine in column B of Answers: =COUNTIFS(Mentions!E:E,"YourBrand",Mentions!B:B,"ChatGPT") / COUNTIFS(Answers!B:B,"ChatGPT")
7. Compute mention-share SOV. Same numerator, with all tracked mentions as the denominator: =COUNTIFS(Mentions!E:E,"YourBrand",Mentions!B:B,"ChatGPT") / COUNTIFS(Mentions!B:B,"ChatGPT") Add a condition on the Market column to split by market.
8. Compute the daily readings and the range. A pivot table on Mentions with Date as rows and Brand as columns gives one count per brand per day. Divide your column by each row's total for your daily share of voice, then take the middle half: =PERCENTILE.INC(DailySOV,0.25) and =PERCENTILE.INC(DailySOV,0.75)
9. Report it whole. The pooled figure, its P25-P75 range, the engine, the market, the date span and the competitor list, and the same row for each competitor. Repeat per engine and per market rather than averaging them.

The spreadsheet range is the spread of your daily readings. It is simple and honest, and it tells you how much the days swing. It does not narrow as you add days, which is the one thing it cannot show you; the method below does.

How Stellarcast measures it

Stellarcast runs this method every day, so you do not have to keep the spreadsheet. What it does, specifically:

• Mention-share SOV per engine and per market. Your brand's mentions divided by the mentions of your brand plus the competitors you track, over a trailing 7, 14, 30 or 90-day window.
• One mention per brand per answer. Each tracked brand is counted once, at its first appearance in the answer. A brand name that appears only because the engine repeated your question back is not counted.
• A P25-P75 range that narrows as days accrue. The range is computed by redrawing the days behind the figure many times and taking the middle half of the results, so it shows how much the figure could still move and always contains it. With a single day of readings there is nothing to redraw, so no range is shown. These are percentiles, never standard errors.
• Visibility rate beside it. The share of answers that name you, with its own P25-P75 range. It needs no competitor list, so it is there from the first day.
• Six engines, daily, on every paid plan: Google AI Overviews, ChatGPT, Gemini, Perplexity, Google AI Mode and Microsoft Copilot. Claude and Grok are available on Enterprise.
• Prompts drafted for each market. Prompts are drafted as a local buyer would ask them in that market's language, not translated from English, and you confirm them before they are measured.
• Append-only, date-stamped readings, so you can see what the number was on any past date.

The full method, including how answers are collected and how brands are matched, is in how Stellarcast measures AI visibility.

One thing share of voice is not: a citation count for your own site. Bing Webmaster Tools reports how often Copilot cites your pages, which is useful and different. Our own 13 weeks of that data came to 5,608 Copilot citations, 79% of them to two pages (the full breakdown). It says how often you are a source, not how much of the naming in your category goes to you.

Share of voice versus share of answer

Share of voice is about presence: how much of the naming goes to you. Share of answer is about ownership: are you the brand the engine recommends, rather than one name in a list? You can have both a high visibility rate and a healthy share of voice while appearing fourth of five in every answer, with a competitor as the one the engine actually puts forward. Presence tells you the engine knows you exist; ownership tells you it trusts you enough to recommend you. Track both, and do not let a strong presence number hide a weak recommendation one.

What share of voice can and cannot tell you

What it can tell you: your presence in a defined set of prompts, your standing against named competitors, the trend over time, and where you are strong or absent by engine and by market. Those are real, actionable readings.

What it cannot tell you: your true share across every prompt real users type, the volume of people who saw any answer, or how logged-in and personalized answers differ at scale. AI share of voice is a sampled, directional index over the prompts you chose. Read as literal market share, it will mislead you; read as a consistent thermometer, it is one of the most useful numbers you can track.

The takeaway

Measure AI share of voice as mention share: your mentions divided by all tracked brands' mentions in the same answers, counted once per answer, per engine and per market. Keep visibility rate beside it, because the two answer different questions. Run the prompts daily, pool the days, and report the figure with its P25-P75 range, never a single run. Write each market's prompts in its own language. Do that, and you have a number that means the same thing next month as it does today.

Choosing a tool to do this? We compare the field in the best AI visibility tools in 2026 and Profound alternatives, with every price on one table in AI visibility tool pricing, compared.

Winning a citation is only half the job: citations fade as engines refresh and competitors publish. See AI citation decay for what causes it, how to measure it and how to keep your place.