Here is a situation that confuses a lot of marketing teams. One AI visibility tool says your brand has 18% share of voice in your category. Another, run the same week, says 44%. Someone asks which one is right, and the honest answer is that both could be. They are asking different questions, of different AI systems, a different number of times, and counting the answers differently.
That is the central lesson of measuring AI visibility: the formula is trivial, and the prompt set and sample size are everything. Share of voice is simply your mentions divided by all brand mentions. But because the same question rarely produces the same answer twice, the number you get depends heavily on which questions you ask and how many times you ask them. This guide covers how to build a measurement you can actually trust, and how to read it without fooling yourself.
This post covers methodology. For the wider context, see our plain-English guide to AI search optimization; for the brand-accuracy side of measurement, see brand entity optimization.
Why this is harder than rankings
A search ranking is deterministic enough to track: search the same term from the same place and you get broadly the same results. AI answers are not. They are generated fresh each time, so the brands named, their order and the sources cited all shift between runs.
The scale of that variation is striking. In research published by SparkToro in January 2026, around 600 volunteers ran roughly 3,000 queries across a dozen prompts on several AI platforms, repeating each prompt dozens of times. The chance of getting the same list of brands twice was under one in a hundred. That single finding changes how you should approach the whole problem: every AI answer is a sample, not a verdict, and a single screenshot of ChatGPT naming a competitor instead of you tells you almost nothing on its own.
The metric family
"Share of voice" is often used loosely to mean several different measurements. They answer different questions, so it helps to track them separately and name them precisely in your reports.
| Metric | How it's calculated | What it tells you |
|---|---|---|
| Share of voice | Your mentions ÷ all brand mentions | How much of the conversation you own |
| Mention rate | Answers naming you ÷ all answers | How often you appear at all |
| Citation share | Links to your site ÷ all cited links | How often you are the linked source |
| Position-weighted share | Mentions weighted by order named | Whether you are named first or fourth |
| Accuracy and sentiment | Reviewed sample of what is said | Whether being mentioned actually helps |
The distinction between mentions and citations matters in practice. A brand can be named without being linked, and name mentions tend to track awareness while linked citations more often lead to visits. Position matters too, because the first brand an answer names carries far more weight with a buyer than the fourth. And none of these numbers mean much if what the AI says about you is wrong, which is why a small, human-reviewed accuracy check belongs in every report.
ClickUp
All-in-one work platform for tasks, docs, goals, dashboards and automation with AI built in.
Best for: Teams wanting an all-in-one project & work platform
Build the prompt set first
The prompt set is your survey questionnaire, and it determines the answer more than any other choice you make. Build it deliberately:
- Use buyer language, not your own. Draw prompts from sales-call notes, support tickets, community forums and search data. The questions people actually ask rarely match the way your team describes the product.
- Cover the buying journey. Weight toward comparison and recommendation questions ("best X for Y", "X vs Z"), then problem questions ("how do I fix…"), with a smaller share naming competitors directly.
- Leave out your own brand name. "Is [your brand] good?" will mention you every time and inflate the number. Branded prompts belong in the accuracy check, not the share-of-voice set.
- Aim for around 50 prompts. Far fewer and one odd question swings the total; far more and each extra prompt adds little.
- Freeze it. Keep the set fixed between measurements, because changing the questions breaks comparability in the same way that rewriting a survey halfway through a study would.
Finding the questions buyers really ask is a skill of its own, and it overlaps with how you structure pages to answer them, as covered in writing pages AI tools quote directly.
The rule The share-of-voice formula is trivial. The prompt set and the sample size are everything. Every AI answer is a sample, not a verdict, so measure the pattern across many runs and never react to a single screenshot.
How many runs is enough?
Because each answer is a sample, precision comes from repetition, and it comes slowly. This is the part most teams underestimate.
A mention either happens or it doesn't, so each run behaves like a coin flip. The margin of error is roughly 1.96 × √(0.25 ÷ runs).
For a single prompt:
→ About 100 runs → within roughly ±10 points
→ About 400 runs → within roughly ±5 points
For the whole set:
→ 50 prompts × 3 runs = 150 observations → the overall figure lands within roughly ±8 points
→ 50 prompts × 8 runs = 400 observations → within roughly ±5 points
What it means: individual prompt results are too noisy to report on. Make the aggregate across the whole set your headline number, and treat per-prompt results as clues about where to look, not as scores.
Real variance is often lower than this worst case, and tools may sample more heavily, but the logic holds: precision comes from volume.
This is also why you should always ask a tool what its denominator is. Half a percent of a set of fifty thousand citations and half a percent of ten hand-picked prompts are completely different claims presented as the same number.
Calculate, then segment
When you run the set, record every brand named in every answer, not just yours, because the denominator is the whole point. Then calculate share of voice overall and break it down in two directions. By platform, because ChatGPT, Perplexity, Gemini and Google's AI features draw on different sources and can place you very differently; some surface few links by design, so lean on mentions there rather than citations. And by intent, because you might dominate problem questions and barely appear in comparisons, which calls for completely different fixes. Local businesses should add a location cut too, since AI answers for nearby services vary by place, as covered in how AI Overviews are changing local discovery.
Where you have the data, record which sources each answer cited. That tells you not just that a competitor is winning a question, but which page, review site or list is doing the work for them, which is the most actionable output of the whole exercise.
Read it as a trend, not a score
Run the set regularly, weekly or monthly, so you catch sudden shifts such as a competitor publishing a page that starts being cited everywhere. But judge it slowly. Report the overall figure as a quarterly trend, not a weekly score, because small month-to-month movements are usually indistinguishable from normal variation. The discipline is the same as reading a cohort curve: look at the direction across periods rather than any single point.
Then connect visibility to outcomes you can observe elsewhere. AI influence often produces no click at all, so pair share of voice with branded search volume, direct traffic and any referral traffic arriving from AI tools, as we discussed in zero-click content and success metrics. Be honest that the link between being mentioned and being bought is indirect, the same caution we applied to attribution generally, and place it inside an end-to-end funnel audit rather than treating it as a revenue metric on its own.
Turn the gaps into work
Measurement only pays off when it changes what you do. Sort your prompts into three groups: questions where you consistently appear, questions where competitors appear and you don't, and questions where nobody credible appears. The middle group is your priority list. For each, look at what the winning answers cite, then decide whether the fix is a better page on your own site, a stronger presence on the third-party sources being quoted, or clearer brand information overall.
The page-level fixes follow the principles in our GEO playbook and in designing websites for AI search; the wider search context is covered in how AI Overviews are reshaping SEO. Then re-measure with the same frozen set, and give each change enough time and enough runs to show up above the noise.
The short version
AI share of voice is your brand's mentions divided by all brand mentions across a defined set of AI answers. The formula is trivial; the prompt set and the sample size decide the result, which is why two tools can report very different numbers for the same brand. AI answers are non-deterministic, and research in early 2026 found under a one-in-100 chance of the same brand list appearing twice, so treat every answer as a sample. Track related metrics separately: share of voice, mention rate, citation share, position-weighted share, and a reviewed accuracy check. Build a fixed set of around 50 prompts in buyer language, weighted toward comparison and problem questions and excluding your own brand name. Precision comes from volume: a single prompt needs roughly 100 runs to be within ten points, so report the aggregate across the whole set and use per-prompt results as clues. Record every brand and cited source, segment by platform and intent, read the figure as a quarterly trend, pair it with branded search and direct traffic, and turn the prompts competitors win into your work list.
Want an AI visibility number you can actually trust?
We help brands measure and grow their presence across classic and AI search.
Explore SEO →Frequently asked questions
What is AI share of voice?
AI share of voice is the proportion of brand mentions in AI-generated answers that belong to your brand, compared with competitors, across a defined set of questions. The basic formula is your brand's mentions divided by the total mentions of all brands across those answers, multiplied by 100. It is a competitive, relative metric: it tells you how much of the conversation you own in your category, rather than simply whether you appear at all. Closely related measures include mention rate, which is the share of answers that name you at all; citation share, which counts linked sources rather than name mentions; and position-weighted versions that give more credit for being named first. There is still no industry-standard formula, and different tools calculate it differently, so the most important thing is to define your own method clearly and apply it consistently over time.
Why do AI share of voice numbers vary so much between tools?
Mainly because they use different sets of questions, different platforms, different numbers of runs and different formulas. The calculation itself is simple, but the result depends almost entirely on which prompts go into it: a set weighted toward questions your brand is strong on produces a flattering number, while a broad category set produces a lower one. Tools also differ in whether they count name mentions or linked citations, whether they weight by position or by estimated search volume, and whether they query the consumer apps or the underlying models directly. On top of that, AI answers are non-deterministic, so the same prompt returns different brands on different runs. Research published by SparkToro in early 2026, based on thousands of repeated queries, found there was less than a one-in-100 chance of getting the same list of brands twice. Two honest tools can therefore report very different figures for the same brand in the same week.
How many times should you run each prompt?
More than most teams expect, depending on how precise you need to be. Because answers vary from run to run, each result is effectively a sample, and the margin of error shrinks only slowly as you add runs. As a rough guide at 95% confidence, pinning down a single prompt's mention rate to within about ten percentage points takes roughly 100 runs, and to within five points takes nearly 400. That is impractical for every prompt, which is why the sensible approach is to rely on the aggregate: a fixed set of 50 prompts run three times each gives 150 observations, enough for an overall figure within roughly eight points in the worst case. The practical rule is to report the overall share of voice across the whole set as your headline, treat individual prompt results as diagnostic rather than definitive, and never react to a single run.
How often should you measure AI share of voice?
Collect data regularly but judge it slowly. Running your prompt set weekly or monthly is useful for spotting sudden changes, such as a competitor launching a page that starts being cited everywhere, but the number you report to leadership should be a trend over a quarter rather than a weekly score. Month-to-month movements of a few points are usually indistinguishable from normal variation, especially at the level of individual prompts or platforms, and reacting to them wastes effort. Keep the prompt set fixed between measurements, because changing the questions breaks comparability in the same way that changing a survey halfway through a study would. When you do need to update the set, run the old and new versions side by side for one period so you can bridge the two series rather than starting from scratch.
