AI share of voice: what it is, how to calculate it, and why every tool reports a different number

What AI share of voice is

AI share of voice is the percentage of all brand mentions in a set of AI answers that name your brand. It measures how much of the conversation your brand holds when an AI assistant answers questions in your category, compared with the other brands it names.

Vendors define it in nearly the same words. The Peec AI documentation defines share of voice as your brand's mentions in AI responses compared with all brands mentioned, yours included. Semrush's guide, published July 17, 2026, gives the formula as your AI mentions divided by the total AI mentions across all brands in your category, times 100. HubSpot's glossary sets it apart from traditional share of voice, which it describes as visibility in paid ads, organic search or social media.

AI share of voice has no platform-reported equivalent. In paid search, Google reports impression share as impressions divided by the impressions an ad was eligible for, and estimates that denominator from its own ad auctions. Google and Microsoft, the two companies with site-owner reports for their AI answers, do not report how often those answers name each brand. Every AI share-of-voice figure is therefore computed by whoever chose the prompts and read the answers, and the method is part of the number.

How to calculate AI share of voice

AI share of voice is calculated by counting brand mentions in a fixed set of AI answers and dividing your count by the count for every brand named:

AI share of voice = (your brand mentions ÷ mentions of every brand named) × 100

1. Write the prompts the way buyers ask them, without your brand name. A prompt that names you, such as a question about your pricing, puts you in almost every answer and inflates your share.
2. Fix the conditions: the engines, signed in or not, the country, the language and the number of runs per prompt. Write them down, because changing any of them changes the number.
3. Fix the brand list: you and the rivals you track. Decide whether brands an engine names that are not on your list join the denominator. Leaving them out raises every listed brand's share.
4. Count mentions in the answer text itself. Count each mention of a brand once, so "Acme.io" is one mention, not one for "Acme" and another for "Acme.io".
5. Divide your mentions by the mentions of every brand in the denominator and multiply by 100. Work out each engine separately before combining them.
6. Report the number with its sample: the prompts, runs and answers counted, the dates, and the formula used.

Steps 2 and 6 matter because the answers themselves vary. Anthropic's documentation for its Claude models says its models are not fully deterministic even at a temperature of 0, so the same input can return a different output on another call. One run of a prompt is one draw, not a reading.

Five ways to compute it from the same answers

The same AI answers can produce an AI share of voice anywhere from 4% to 47%, depending on which variant is computed. The worked example below uses five answers, one for each of five category questions, with three brands: you, Rival A and Rival B. The mention counts and search volumes are invented for the example.

Each variant answers a different question:

• Pooled mention share rewards long answers about you. Answer 1 names you six times, and that one answer carries most of the 46.7%.
• A per-answer average gives every answer one vote, so a single answer that repeats your name cannot dominate it.
• Share of appearances ignores repetition inside an answer. HubSpot's glossary uses this basis in one example (30 of 100 responses) and a mention basis in another (25 of 100 mentions), without saying they differ.
• Search-volume weighting rewards appearing on the questions people search for most. The Ahrefs Brand Radar academy describes impressions as mentions weighed against Google's search volume, and share of voice as your percentage of those impressions. Ahrefs does not publish its exact arithmetic, so the last row follows that description, not Ahrefs' code.

Answer 4 needs its own rule. It names no brand at all, so it has no denominator, and your share in it is undefined rather than zero. Counting it as 0% moves the average from 29.8% to 23.8%, although nothing about your brand was measured. Leave such answers out of share of voice and report them separately, as the share of answers that name any brand.

One related measure looks like a share but is not one. The Peec AI visibility score is the percentage of responses that mention your brand: 2 of 5 answers, or 40%, in the example. Rival A's visibility is 80% and Rival B's is 60%, so visibility scores across brands add up to more than 100%.

Share of voice vs share of citations

Share of voice counts brand names in the answer text, while share of citations counts the links an answer gives as sources, and the two can move apart. A brand can be named with no link to its site. A page can be cited in an answer that recommends a rival.

Bing's Citation Share is the closest thing to a first-party share metric. Microsoft's June 16, 2026 post defines it as your site's citations as a percentage of every citation shown, from any site, for one grounding query. The same post says it does not expose competitor domains or traffic share, and that it is not meant as a competitive scoreboard.

Google's generative AI performance report counts impressions of links to your site in AI Overviews and AI Mode. Its help page describes one metric: impressions of links to your own site. Neither the Google nor the Microsoft report counts brand names in the answer text, so share of voice is always measured by sampling answers.

How many prompts and runs you need

An AI share-of-voice reading needs hundreds of answers to separate a real change from run-to-run noise. The simplest way to size a sample is through visibility, the share of answers that name you, because each answer either names you or does not.

The NIST Engineering Statistics Handbook gives the Wilson interval for a proportion. The table applies it to a brand that a long series of answers would name 30% of the time.

The number of answers is prompts times engines times runs. Fifty prompts on two engines with three runs each is 300 answers. Twenty prompts asked once is 20 answers, and a 30% result from it is consistent with anything from about 15% to 52%.

A change between two readings needs the same care. Using the two-proportion z-test the NIST handbook describes, a move from 30% to 40% between two months of 100 answers each is not significant at 95% (z = 1.48). With 200 answers a month, the same move is significant (z = 2.10).

These ranges treat every answer as independent. Repeated runs of one prompt are related, so the true range is wider than the table shows, and adding distinct prompts narrows it more than repeating the same ones. Pooled share of voice counts mentions, which come in clumps inside answers, so it has no formula this simple. Size its sample by answers, not by mentions.

Why AI visibility tools report different numbers

AI visibility tools report different share-of-voice numbers for the same brand because they count different things, weight them differently and ask different questions. The table sets out what each vendor says about its own metric.

Beyond the formula, five things differ between tools:

• The questions. A tool that runs your own list measures the market you chose. Ahrefs' Brand Radar methodology draws its questions from Google's People Also Ask corpus and its own keyword database.
• The surface. Ahrefs runs its prompts through the free, public web versions of ChatGPT, Perplexity, Gemini and Copilot. An API call and a consumer app need not answer alike, as the guide to ranking in Gemini sets out for Google's API.
• The window. Ahrefs refreshes ChatGPT, Perplexity, Gemini and Copilot monthly and reports over a 90-day window.
• The matching. Ahrefs describes mentions as string matches for a term. A string match counts a common word that is also a brand name, and misses a product named differently from its brand.
• The audience. Ahrefs says its metrics show potential visibility, not actual audience reach. Google's guidance on third-party SEO tools says such tools have no access to its internal ranking data and cannot guarantee performance.

Comparing share of voice across tools is therefore comparing methods. Compare within one tool, with the same prompts, over time. The same applies to a hand-kept log, such as the ones described for ChatGPT and Perplexity.

Where overads fits

overads is a marketing workspace where AI agents draft posts, publish them once approved and report on ads. Its Signals product is built to measure AI share of voice the way this guide describes, and its AI-answer scan is switched off on the live product.

As built, overads Signals puts a fixed list of tracked questions to Gemini and Perplexity when a scan runs, and works out share of voice like this:

• A model reads each answer to judge which brands it talks about and in what tone. Code then counts the mentions and does the arithmetic, so the same answer always gives the same share.
• A brand's names are matched longest first, so "overads.io" counts as one mention, not one for "overads" and another for "overads.io".
• An answer that names no brand gives no reading, not 0%.
• The headline share pools mentions over a chosen window, 30 days by default, with every brand the answers name in the denominator. Names the engines raised that are not on the tracked list are shown apart as candidates.

What Signals does not do:

• Signals is built for Gemini and Perplexity only. It does not cover ChatGPT, Microsoft Copilot, Claude, AI Overviews or AI Mode.
• Signals does not weight share of voice by search volume or by position in the answer.
• A team that needs share of voice running now, across ChatGPT or Google's AI features, or weighted by search demand, is better served by a dedicated tool. The comparison of overads and Peec AI sets out what one such tool covers.

For a team that wants the reading beside its other marketing work, how B2B growth teams use overads shows Signals next to ad reporting in one workspace. The method in this guide needs no tool at all: a fixed prompt list, a spreadsheet and the same formula each month give a share of voice you can defend.