# How to Measure AI Share of Voice Across ChatGPT, Gemini, and Perplexity

**Author:** John Morabito (Founder, /winston)
**Published:** September 15, 2026
**Updated:** October 1, 2026
**Reading time:** 10 minutes
**Canonical:** https://www.winstondigitalmarketing.com/playbooks/how-to-measure-ai-share-of-voice/

For twenty years the question was "where do we rank." In an AI answer there is nothing to rank. The assistant reads the web, decides which brands to name, and hands the user a short list. So the question changes to something more useful and more brutal: when someone asks about what you sell, how often does the engine name you, and how do you stack up against the competitors it names instead? That is AI share of voice, and it is becoming the metric that matters. Here is how to actually measure it, across the engines that count, in a way you can track and improve rather than guess at.

## What AI share of voice actually is

AI share of voice is the percentage of relevant AI answers that name your brand, measured against the other brands named in those same answers. It reframes visibility for a world where the ten blue links are collapsing into a single synthesized response.

Think about what happens when a buyer asks an assistant "what are the best options for X." The engine returns a handful of named brands. You are either in that set or you are not, and if you are, you are one of several. Share of voice captures both facts at once: your presence, and your presence relative to everyone else's. A keyword ranking tells you that you are position four on a page most people no longer scroll. Share of voice tells you that, across the questions your buyers actually ask, the AI names you in a certain share of answers and names your top competitor in more. That second number is the one that predicts whether you are winning the AI-mediated market or losing it quietly.

This is the same shift we argue in why citation share is replacing rankings (https://www.winstondigitalmarketing.com/playbooks/citation-share-replaces-rankings/): the unit of measurement moved from the ranked page to the named entity. Share of voice is how you put a number on it.

## Step one: build a stable prompt set

The entire metric rests on the prompts you choose, so this is where the work starts and where most attempts go wrong. You are not tracking keywords; you are tracking the questions a real buyer would type into an assistant. Those are longer, more conversational, and more intent-loaded than keywords: "what is the best CRM for a small law firm," "which dispensaries deliver in Brooklyn after 9pm," "is [brand] worth it compared to [competitor]."

A useful prompt set spans the buyer journey: broad category questions where you want to be discovered, comparison questions where you are weighed against rivals, and brand questions where people are checking you out specifically. Cover the intents that actually lead to revenue, not just the ones where you already look good.

The one rule you cannot break: once you have chosen the set, keep it stable. Share of voice is only meaningful as a trend, and a trend requires measuring the same prompts every time. If you change the prompts between measurements, you are comparing two different things and the movement is noise. Build the set deliberately, then hold it steady and only evolve it slowly and on purpose. If you need help generating a strong set, that is the subject of GEO prompt research (https://www.winstondigitalmarketing.com/playbooks/geo-prompt-research/).

## Step two: run the prompts across the engines that matter

Here is the part single-tool checkers get wrong: the engines do not agree with each other. Ask the same question of ChatGPT, Google's AI Overviews, Google's AI Mode, Perplexity, and Gemini, and you will get different brands named and different sources cited. A brand can dominate share of voice in Perplexity and be nearly invisible in Google's AI Overviews. If you measure one engine and call it your AI visibility, you are looking at a fraction of the picture and drawing conclusions from it.

So the measurement has to run across the full set of engines your buyers use. Today that is:

- ChatGPT
- Google AI Overviews
- Google AI Mode
- Perplexity
- Gemini

These are the five surfaces where cited, real-time answers are being read and acted on, and they are exactly the five the Winston GEO Tracker (https://www.winstondigitalmarketing.com/geo-tracker/) measures. It does not track Claude, because Claude does not surface the same kind of cited, live-web recommendations these five do, so tracking it would not tell you anything about the answers your buyers are getting. For each prompt on each engine, you record two things: were you named, and which competitors were named alongside you.

## Step three: do the math

With that data, the calculation is simple. Across your prompt set, on a given engine or across all of them:

    AI share of voice = your brand mentions / all brand mentions

If, across your tracked prompts, the engines produced one hundred brand mentions total and eighteen of them were you, your share of voice is eighteen percent. Track it per engine, because your position differs by engine and the per-engine number tells you where to focus, and track it in aggregate for the headline trend. The absolute number matters less than two things: how it moves over time, and how it compares to the specific competitors showing up in your answers.

That competitive framing is the whole point. Share of voice is inherently relative. Knowing you are named in eighteen percent of answers means little until you know the market leader is named in forty and a rival you did not worry about is named in twenty-five. The metric turns a vague sense of "are we visible in AI" into a leaderboard you can read at a glance.

## How many prompt runs you need for a number you can trust

AI answers are not fixed. Ask ChatGPT the same question twice and you can get a different list of brands. So one run of one prompt is a coin flip, not a measurement. How much you can trust your share of voice number depends on the size of the sample underneath it: prompts, times runs, times engines.

The math is the ordinary math of a proportion. The margin of error shrinks with the square root of the number of answers you collect. With 100 answers on one engine, a share near 50 percent carries a 95 percent margin of roughly plus or minus 10 points. With 400 answers it is roughly plus or minus 5. Halving the noise takes four times the answers.

In practice that comes down to three rules:

- Run each prompt more than once per measurement window on every engine, and count mentions across all the runs instead of trusting a single answer.
- Size the sample to the decision. If you need to see a 5 point shift on one engine, you need a few hundred answers on that engine per window. A 25-prompt set run once will not show it. You get there by adding prompts, adding runs, or pooling a week of daily runs.
- Do not read a move smaller than the noise. Say your share slips from 22 to 19 percent on 80 answers. That is inside the margin, and nothing happened. Put the sample size next to the number so nobody reacts to it in a meeting.

## Per topic, per engine, per competitor

One blended number hides most of what you need. Share of voice can be measured per topic, and it should be. Tag every prompt with the topic or product line it belongs to and compute the share for each tag on its own. A brand that leads on pricing questions and barely appears on "alternatives to" questions has two different problems, and the blended number averages them into one that describes neither.

Engines work the same way. Report each one separately as well as in total, because Perplexity and Google's AI Overviews pull from different sources and will disagree about you. Then name the competitors. Share of voice against "everyone else" is less useful than share of voice against the three rivals your sales team actually loses deals to.

Those are also the questions to put to any tool you are evaluating. Which engines does it run? Does it repeat prompts or take one answer per prompt? Does it report per engine and per topic? Do you choose the competitor set? Can you export the raw answers and check its counting? The vendor comparison is in [the best AI citation tracking tools for share of voice](https://www.winstondigitalmarketing.com/playbooks/best-ai-citation-tracking-tools/).

## Why you cannot do this by hand for long

You can measure share of voice manually. Once. Sit down, run your prompts across five engines, and tally the mentions, and you will have a real snapshot and probably a few surprises. The problem is that a snapshot is nearly useless, because AI answers change constantly as the engines recrawl the web and reweight sources. The answer you got last Tuesday can be different today. A number you measured once and never again tells you where you were, not where you are.

What you actually need is the trend, and the trend requires running the same prompt set across the same engines on a schedule, which is a lot of repetitive work to do by hand every week. This is exactly why the measurement gets automated. The GEO Tracker runs your prompt set across all five engines on a schedule, tallies your mentions and your competitors' automatically, and shows share of voice as a trend and a competitive leaderboard rather than a one-time count. It runs daily if you want it to, at $0.75 per prompt, which is deliberately far below what most tracking tools charge, because the whole point is that measuring often should be affordable enough to actually do. The comparison of the tooling landscape is in the best AI citation tracking tools (https://www.winstondigitalmarketing.com/playbooks/best-ai-citation-tracking-tools/).

## From measurement to movement

Measuring share of voice is only worth it if you use it to grow the number, and the measurement itself tells you how. The tracked data shows you exactly which prompts name your competitors but not you, and the sources cited in those answers show you where the engine is getting the information behind the answer. That is your roadmap. You are absent from a prompt because the engine is pulling from sources that do not mention you, so the work is to become one of the sources it pulls from. Turning those absences into a prioritized, action-tagged list is the whole subject of AI content gap analysis (https://www.winstondigitalmarketing.com/playbooks/ai-content-gap-analysis/).

That work is generative engine optimization: publishing genuinely citable content that answers the prompts where you are missing, earning third-party mentions and reviews on the sources the engines already trust, and keeping your brand a clear, consistent entity so an engine can confidently name you. Then you re-measure and confirm the number moved. Share of voice is the scoreboard; the GEO work is how you score. Run without the scoreboard and you are optimizing blind; run without the GEO work and you are just watching a number you never change.

The place to start is a baseline. You cannot manage a number you have never measured, so the first step is simply seeing where you stand: which engines name you, which name your competitors, and where the gaps are. That is what the free AI visibility audit behind the GEO Tracker gives you, no call required. Establish the baseline, then decide what to move. If you would rather stand the whole tracking program up yourself step by step, the operational checklist is in [how to set up AI visibility tracking](https://www.winstondigitalmarketing.com/playbooks/how-to-set-up-ai-visibility-tracking/).

If you run marketing for a company with several product lines or brands, the prompt set and the topic tags are most of the work, and they are the part in-house teams most often get wrong. That is the piece we build alongside your team: the prompt set, the baseline, and a read on which gaps are worth closing first. A 20-minute call is enough to tell whether it is worth doing. Book a 20-minute call (https://www.winstondigitalmarketing.com/contact/).

We run AI share of voice measurement and the GEO work that improves it, tracked with the GEO Tracker, as part of our generative engine optimization (https://www.winstondigitalmarketing.com/services/generative-engine-optimization/) practice.

## Frequently asked questions

### What is AI share of voice?

AI share of voice is the percentage of relevant AI answers that name your brand, measured against the competitors named in those same answers. Instead of asking whether you rank for a keyword, it asks: across a fixed set of prompts a real buyer would type, how often does the engine mention you, and how does that compare to the rivals it mentions? If an assistant names five brands across your tracked prompts and you are one of them, your share of voice for that set is your slice of those mentions. It is the AI-era version of the visibility metric, because in an AI answer there is no ranking to hold; there is only whether you are named or not, and how you stack up against everyone else who is.

### How do you measure AI share of voice?

You build a fixed set of prompts that represent how your buyers actually ask, run them across the AI engines on a schedule, and record for each answer whether you are named and which competitors are named alongside you. Share of voice is then your mentions as a proportion of all brand mentions across that prompt set, tracked over time. The two things that make it valid are a stable prompt set, so you are comparing like with like from one measurement to the next, and consistent scoring of what counts as a mention. Do it once and you have a snapshot; do it on a schedule and you have a trend you can actually manage. Doing this by hand across engines is tedious, which is why most teams use a tracker to run the prompts and tally the mentions automatically.

### Which engines should you measure AI share of voice across?

The ones your buyers actually use, which today means ChatGPT, Google's AI Overviews, Google's AI Mode, Perplexity, and Gemini. These are the surfaces where AI answers are being read and acted on, and they do not agree with each other, so measuring only one gives you a distorted picture. A brand can have strong share of voice in Perplexity and be nearly absent in Google's AI Overviews, and you would never know if you only checked one. Measuring across the full set is the only way to see your true position, which is why the Winston GEO Tracker measures all five. It does not track Claude, because Claude does not surface the same kind of cited, real-time web answers those five do.

### How is AI share of voice different from citation share?

They are closely related and often used together, with a subtle difference in emphasis. Share of voice is about being named in the answer relative to competitors: what proportion of the brand mentions across your prompt set are yours. Citation share, in the strict sense, is about your domain being cited as a source the engine links or attributes. In practice both measure the same underlying thing, your presence in AI answers versus everyone else's, and a good tracking program watches both: whether the engine mentions your brand, and whether it cites your pages. If you want the strategic argument for why this class of metric is replacing keyword rankings altogether, that is the subject of our piece on why citation share is replacing rankings.

### How do you improve AI share of voice once you are measuring it?

You use the measurement to find where you are absent and why, then close those gaps. The tracked data tells you which prompts name competitors but not you, and the cited sources in those answers tell you where the engine is getting its information, which is where the work has to happen. Improving share of voice is then the core generative-engine-optimization work: publishing genuinely citable content that answers those prompts, earning third-party mentions and reviews on the sources the engines trust, and keeping your entity clear and consistent so engines can confidently name you. You re-measure on a schedule to confirm the work is moving the number. Without the measurement you are guessing; with it, share of voice becomes a metric you deliberately grow.

### How many prompt runs do you need for a reliable AI share of voice number?

Enough answers that the number stops swinging between measurements. AI answers vary from run to run, so count mentions across repeated runs of each prompt rather than single answers. The margin of error on a proportion shrinks with the square root of the sample: about 100 answers on one engine gives roughly plus or minus 10 points near a 50 percent share, and about 400 gives roughly plus or minus 5. Pick the sample size from the smallest change you need to detect, and report the sample size next to the number.

### Can AI share of voice be measured per topic?

Yes, and it should be. Tag each prompt with the topic or product line it belongs to, then compute share of voice for each tag separately. A blended number across every topic averages strong and weak areas together and hides where you are actually losing. Per-topic share tells you which product line or question type to work on first.

### Can a tracker show competitor share of voice in Google AI Mode?

Yes, if it runs prompts in AI Mode specifically and records every brand named in each answer, not just yours. Check that the tool treats AI Mode as its own engine rather than folding it into AI Overviews, because the two surfaces can name different brands. The Winston GEO Tracker runs AI Mode alongside ChatGPT, Gemini, Google AI Overviews and Perplexity and shows your share against competitors as a leaderboard.