Ask an assistant the same buying question a hundred times and you will not get the same answer twice. In a SparkToro and Gumshoe study published on 30 January 2026, covering 2,961 prompts across ChatGPT, Claude and Google's AI search with 60 to 100 runs each, there was less than a 1 in 100 chance that two runs returned the same list of brands, and less than 1 in 1,000 that they returned it in the same order. That one finding shapes everything about measuring share of voice. A single answer is an anecdote. Share of voice is a distribution, built from many runs of a fixed prompt set, and the discipline lies in sampling it honestly.
What is AI share of voice?
AI share of voice measures how much of the competitive conversation in AI answers belongs to your brand. The IAB defines it in its August 2026 guidelines, Measuring Visibility in the AI Era, as a brand's mentions as a proportion of all competitive category mentions. It sits in the "presence" tier of the framework, alongside mention rate (responses containing your brand divided by total responses), citation rate (responses citing your content divided by total responses) and visibility momentum (the percentage change in those figures between periods).
The distinction matters because the three are often confused. Mention rate is absolute: how often you appear. Citation rate is about attribution: whether the assistant points to your content. Share of voice is relative: your slice of a finite pool of brand mentions, so it can fall even when your mention rate rises, if competitors rise faster.
How do you calculate AI share of voice?
The standard calculation is your brand mentions divided by total category brand mentions, multiplied by 100. If your brand is named 30 times in a set of answers that name brands 120 times in total, your share of voice is 25 per cent. Define "mention" in advance, including how you treat misspellings, product names and parent brands, and apply the same rule to every competitor.
Two refinements are worth deciding on before you start:
- Presence versus prominence. A mention in the first sentence of a recommendation is not the same as the fifth item in a list. The IAB separates presence from prominence, which covers position, ranking order and whether a brand gets depth or only a passing reference. Some teams weight mentions by position or by the share of an answer's words devoted to each brand.
- Mention-based versus citation-based. Counting named brands tells you who is part of the conversation. Counting cited domains tells you whose content the assistant leaned on. For a publisher these can diverge sharply, because an assistant may read your page and name someone else.
How many prompts and runs do you need?
Enough to make the number stable, which is more than most teams assume. The IAB draws a line between directional measurement, suitable for trend monitoring, and decision-grade measurement, suitable for allocating budget. Directional measurement needs at least 50 queries per programme, multiple responses per query, coverage of at least two intent types and a monthly or quarterly cadence. Decision-grade measurement needs a large, diverse query set covering categories and subcategories, enough responses per query for a stable distribution, all four intent types (informational, comparison, recommendation and transactional), weekly or more frequent testing, and documented acceptable variation ranges within a seven-day window.
The practical reading is blunt. If you are testing 15 prompts once a month, you have a spot check, not a share of voice. Report it as directional, say so on the slide, and do not move budget on it.
Why you must measure each assistant separately
Assistants do not share a source pool. Brand24 cites a Writesonic study of 161,286 prompts in which ChatGPT, Gemini, Perplexity and Google AI Overviews cited the same domains in only 3.8 per cent of cases. Whatever the precise figure, the direction is consistent across the research: a brand can lead on one assistant and barely register on another for the identical question. A blended number averages away the one thing you need to act on, which is where you are weak.
Retrieval settings matter as well. The IAB asks providers to disclose whether live web retrieval was enabled when responses were collected, because a model answering from memory and a model answering from pages it has just read are measuring different things. Pick a setting and keep it.
A practical measurement process
- Build the prompt set from real buyer language. Cover awareness, comparison, recommendation and transactional questions for your category, and keep a written record of where each prompt came from. Prompts derived from search keyword data are a proxy for how people phrase questions to an assistant, not a record of it.
- Fix the competitor set. Share of voice is only as meaningful as the denominator. Agree the list of brands before the first run and version it.
- Run each prompt repeatedly on each assistant. Because identical runs diverge, treat each prompt as a small sample. Record the share of runs in which each brand appears rather than a single yes or no.
- Log the context. For each answer, capture which brands appear, in what order, which sources are cited, how the brand is framed and whether any claim about it is wrong. The IAB's portrayal tier covers sentiment, framing, hallucination rate and factual inaccuracy rate.
- Aggregate per assistant, then overall. Report each assistant's share separately before showing any blended figure.
- Track the trend, not the reading. Compare like with like, using the same prompts, settings and cadence, and look at visibility momentum over several periods.
Resist the temptation to publish a "rank in AI". Rand Fishkin's verdict on any tool that reports a ranking position in AI was that it is, in his words, "full of baloney", and his advice was that providers should show their maths before you spend anything.
What does a good AI share of voice look like?
There is no universal benchmark, because it depends on how crowded the category is. One reference point: AthenaHQ's State of AI Search 2026 report, as summarised by Brand24, puts leading brands at an average of roughly 33.6 per cent share of voice, second-placed brands at around 19.5 per cent and third-placed at around 13.3 per cent. Treat that as a shape, not a target. Concentration at the top is normal, and your useful comparison is your own share against the named competitors you actually lose deals to, in the decision-stage prompts closest to a purchase.
The measurement gap is also still wide. The Scribewise and Scrunch survey reported by Digiday found 71 per cent of marketers not tracking competitive share of voice, and the IAB's own framing, via PPC Land, is that only 16 per cent of brands systematically track AI visibility. Anyone running a clean, repeatable programme is ahead of most of the market.
Which tools measure AI share of voice?
Several platforms run prompt sets across the main assistants and report mention rate, share of voice and sentiment, including Profound, the Semrush and Adobe toolkit, Ahrefs Brand Radar, Otterly, Peec and, for Microsoft's own grounding data, Clarity. Their numbers will not agree with one another, because they sample different prompts at different times with different retrieval settings. Our comparison of AI visibility tracking tools sets out how each collects data. The rule that follows is simple: choose one tool, hold its configuration constant and compare it only with itself.
Turning share of voice into action
A number only earns its place if it changes a decision. Look first at low share in high-intent prompts, which is the most commercially urgent gap. Then look at which sources the assistants cite when a competitor wins, because those pages are the ones you need to be present on.
There are two routes to closing a gap. Organic generative engine optimisation earns citations over time through clear, authoritative, retrievable content. Paid content-layer placement puts a brand fact into the publisher content an assistant reads at the moment a relevant question is asked, and blankspace operates on that second route at the CDN edge, with advertisers able to track presence in AI responses before and during a campaign using an independent measurement tool. For publishers, the same logic runs in reverse. Their request logs show which of their pages assistants read, which is the audience an advertiser is trying to reach, and it is data the answer-side tools cannot see.
Frequently asked questions
What is a good AI share of voice?
It depends on the category and your competitors, so there is no universal target. AthenaHQ's 2026 report, as summarised by Brand24, puts category leaders at an average of roughly a third of mentions. Judge yourself against your named competitors in high-intent prompts and watch the trend across periods rather than a single reading.
How often should I measure AI share of voice?
The IAB's guidance is monthly or quarterly for directional trend monitoring and weekly or more frequent for decision-grade measurement. Because AI answers shift quickly and vary between runs, a one-off snapshot is unreliable whichever cadence you pick.
How many prompts do I need to track?
The IAB sets a floor of at least 50 queries per programme for directional measurement, with multiple responses per query and at least two intent types covered. Decision-grade measurement needs a larger, more diverse set covering all four intent types, run weekly or more often, with documented variation ranges.
Why does my share of voice differ between ChatGPT and Perplexity?
Each assistant draws on a different pool of sources and cites differently. Research cited by Brand24 found four major assistants cited the same domains in only 3.8 per cent of cases. Measure each assistant separately, and show a blended figure only after the individual ones.
How do I improve my share of voice in AI answers?
Through organic generative engine optimisation, which builds durable citations, and through paid content-layer placement, which puts your brand into the content assistants read at the moment of a relevant query. Prioritise the decision-stage prompts where a mention is closest to a purchase, and re-measure with the same method.