Free Growth Assessment
SEO & AEO

AI Share of Voice: How to Measure Your Company’s Presence in AI Answers

A percentage means nothing until you know what it’s a percentage of. Here are the formulas, a worked example and a quarterly sheet one person can run.

By Rhett Collver Last updated: November 22, 2026 11 min read

AI share of voice is the share of AI answers that name your company, out of all answers that name any company you track, across a fixed set of prompts. Fix the prompts, rivals, and formula before you count.

A hand sorting small blank white cards into four piles on a dark slate desk
Count the answers first. Choose the denominator second.

A vendor shows you a dashboard. It says you hold 18% AI share of voice.

Ask one question. Eighteen percent of what?

The answer changes by vendor, and so does the number, because some divide by every answer, some divide by every company name, and some weight each answer by estimated search volume before they divide anything. Same answers. Different percentages. Nobody has to be lying for that to happen.

That makes this a ratio problem before it’s a content problem. Get the ratio right and you have a gauge you can track. Get it wrong and you’ll cheer or panic over an artifact. Our AI visibility audit starts at the same place: agree on what’s being counted, then count it.

Key Takeaways

The Short Version

Four decisions come before any counting: write down the prompts, the rivals, the platforms, and the formula, then run each prompt several times, because one run proves nothing. Report how often you appear and never where you rank inside the answer. Read the number beside your leads, not instead of them.

C4 sells AI visibility work, so that’s the interest on the table. The steps below need a spreadsheet and one person’s time. They don’t need a tracker or a retainer.

What AI Share of Voice Means

Share of voice is an old advertising idea. It’s your slice of the attention in a market. The AI version moves that idea to a new place: the answer an assistant writes when a buyer asks it a question.

AI share of voice is the percentage of AI-generated answers, in a defined set of prompts, that mention your company compared with the competitors you track. It’s a ratio of presence. It doesn’t measure rank, tone, traffic, or sales. Whoever runs the count decides the prompts, the rivals, and the math.

There is no standard. Here are four definitions in the wild.

Same Name, Different Math

SourceWhat it countsWhat it says about its limits
Ahrefs, Brand RadarThe share of AI visibility your brand captures against the competitors you track, across the prompts you care about. Its formula divides responses that mention you by responses that mention any tracked brandTwo tools can report different numbers for the same brand without either being wrong
Semrush, AI Visibility ToolkitThe percentage of mentions your brand receives in AI-generated answers compared to competitors in your marketIts metrics are directional signals, not exact numbers
Similarweb, AI SearchBrand mentions with no linked source. Citations, where your page is linked, are counted as a separate numberThere is no universal benchmark
A spreadsheet you ownWhichever of the three versions below you pick and write downOnly what you decide to record

Every company in that table sells AI visibility tracking, so each gains when you buy a tracker. Each also warns you, in its own words, about what its number can’t do. Believe that part.

How to Calculate AI Share of Voice From the Same Answers

Here’s an invented example, and it’s arithmetic, not a finding. You test 25 prompts on one platform and run each one four times, which gives you 100 answers to count, and you track your own firm and three rivals that sell into the same market.

Your firm is named in 30 answers. Rival A is named in 45, Rival B in 25, and Rival C in 20. That’s 120 mentions in all. Seventy of the 100 answers name at least one of the four firms, and the other 30 name nobody you track.

Now divide three ways.

Same 100 Answers, Three Denominators

VersionDivide your 30 byYou getThe question it answers
Presence rateAll 100 answers30%How often do I show up at all?
Share of answers naming a tracked firmThe 70 answers that name any of the four43%When a tracked firm gets named, how often is it me?
Share of mentionsAll 120 mentions25%Of every name handed out, how many were mine?

Same data. Thirty, forty-three, twenty-five.

Which one is right? All three, for the question each one asks. Ahrefs’ published formula is the middle one, and Semrush’s wording, a percentage of mentions against competitors, reads like the last. A tool that doesn’t say which it uses hands you a number with no meaning attached.

The rivals look different under each version too. On share of mentions they land at about 38%, 21%, and 17%, and with yours at 25% the four add up to 100 apart from rounding. On the middle version they add up to more than 100, because one answer can name several firms. That’s fine. Just don’t draw them as slices of one pie.

Pick one version. Write it down. Never switch halfway through a year, because a change of formula moves the number as surely as a real change does.

Two more versions exist. Citation share counts how often your pages are linked as a source, not merely named. Similarweb says a brand can hold a strong AI share of voice and a near-zero citation share at the same time, and it calls the first a recognition signal and the second a trust signal. Its own page states the citation-share formula two ways, one counting citation events and one counting prompts, which proves the point about as well as I could.

Ahrefs adds a fourth: weight each answer by modeled search volume, so a prompt behind 20,000 searches counts for more than one behind 1,000. It’s a reasonable idea, and it’s an estimate sitting on an estimate. Ahrefs says as much: impressions are “a modeled estimate of exposure, not a count of people who saw an answer.”

Build the Prompt Set Before You Count Anything

Your prompt list is your market. Change the list and the number moves, though no buyer has changed their mind about you. That’s why it comes first.

  • Buyer questions at three points. Working out the problem, building a shortlist, and comparing two firms. A prompt such as “which MSPs handle compliance for a 40-person law firm” sits in the second group.
  • Three to five rivals. Name the ones your buyers actually shortlist, not the ones you envy. Ask your last ten lost deals who won.
  • One list of platforms. ChatGPT, Gemini, Perplexity, and Google’s AI features answer differently, so count each on its own before you blend them.
  • A frozen version. Save the list with a version number. Next quarter’s additions become version two, and you rerun the old set beside it.

How many prompts? My rule of thumb is 20 to 30, run three to five times each on each platform, which gives you 60 to 150 answers per platform and is enough to see a pattern without hiring a researcher. It’s a judgment call, not a study. SparkToro’s authors say their own research needed far larger samples, so treat my range as a floor.

Need starting questions? Our AI search prompt library lists the ones buyers type into an assistant. Borrow what fits and cut the rest.

Why One Run of One Prompt Tells You Nothing

Ask an assistant the same question twice and you can get two different lists. That isn’t a bug. It’s how these tools work.

SparkToro and Gumshoe measured it. In surveys run in November and December 2025, 600 volunteers ran 12 prompts through ChatGPT, Claude, and Google’s AI Overviews, 2,961 runs in all, and SparkToro published the results on January 28, 2026. For ChatGPT and Google’s AI, the chance of getting the same list of brands twice was under one in a hundred, and Claude was only slightly more likely to repeat. The odds of the same brands in the same order were nearer one in a thousand.

Presence held up better than order. The author writes that visibility “across dozens to hundreds of prompts run multiple times is a reasonable metric,” and he treats ranking position as close to noise. In one follow-up, the cancer center City of Hope appeared in 69 of 71 ChatGPT answers, a 97% visibility rate, yet was the top mention in only 25.

Read the limits. Volunteers used default settings, the work wasn’t peer reviewed, and the authors say they need far larger samples, while Gumshoe sells AI tracking and SparkToro sells software. The author says he started as a skeptic. Weigh all of that, and keep the direction of the finding.

A hand laying a brass ruler across four stacks of blank white cards on a slate desk
Measure the same stacks twice and expect a small difference.

Google says something similar in plainer words. Its page on AI features and your website says AI Mode and AI Overviews “may use different models and techniques, so the set of responses and links they show will vary.”

So what does a single 30% mean? Take our invented 30 in 100. Textbook math for a proportion puts the 95% range at roughly plus or minus 9 points, so 21% to 39%, but that assumes every answer is an independent draw, and repeat runs of one prompt aren’t. The real range is wider, so treat 9 points as a floor.

Slide from 30% to 34% and you have learned nothing.

The action is plain. Report presence across many runs, report it with a range, and delete any column called “position in the answer.”

What the Tools Count and What They Leave Out

If you buy a tracker, ask three questions before the demo ends. What’s the denominator? How many times is each prompt run? Which platforms and which country? Then read the vendor’s own caveats, which are usually more honest than its marketing.

  • Ahrefs. Weights answers by modeled impressions, and says two tools can show different numbers for one brand with neither being wrong.
  • Semrush’s data page says its prompt database is captured from real requests, not through the LLMs’ APIs. Then comes the warning: “AI search and LLM responses are fast-changing and highly personalized, which means no platform can provide exact numbers on visibility.” It calls its metrics “reliable directional signals.”
  • Similarweb. Counts mentions and citations apart. “There is no universal benchmark.”
  • A number with no stated denominator is a slogan.

Directional is the honest word. Use a tool for the direction, and keep a sheet whose rule you can explain to your own board. If you can’t explain the number to a buyer, don’t hand it to a board.

Where AI Share of Voice Stops

It counts presence. Ahrefs says its own metric doesn’t tell you whether mentions are positive, whether AI actually recommends you, or whether you get traffic. Named isn’t the same as recommended. A buyer can read your name in a list of firms to think twice about.

It also comes with no target. Ahrefs says there isn’t a universally good AI share-of-voice score, so this post gives you no “good” range, and the ranges to compare against sit in our post on MSP AI visibility benchmarks. Read them with the same three questions: what was counted, how many runs, which platform. Compare yourself with your own last quarter first.

Tie the number to money. Add a “how did you find us” line to your form and your first call, and log every answer that names an AI assistant. Check AI-referred visits in GA4 as well. Our post on tracking AI visibility for an MSP shows the setup.

If share of voice climbs and leads don’t, you have a different problem. A rising number you can’t connect to a lead is a vanity metric with better branding.

A Quarterly AI Share of Voice Sheet One Person Can Own

One person. One sheet. One date a quarter. Six steps, in this order.

  1. Write the prompt list and call it version one.
  2. Pick three to five rivals and the platforms, then write the formula as one sentence.
  3. Run each prompt three to five times on the same day, signed out, in a private window. Record the platform, the mode and the date.
  4. Log every firm named in each answer, and the domains linked if you want citation share.
  5. Calculate the presence rate, the share of answers naming a tracked firm, and the share of mentions, each with its range.
  6. Write down what changed since last quarter, including any edit to the prompts, rivals, or formula.

What the Sheet Holds

ColumnWhat goes in itWhy it earns its place
Prompt and stageThe exact wording, and where it sits in the buying pathLets you add or drop a prompt without losing history
Platform, mode and dateChatGPT, Gemini, Perplexity or Google, which mode, and the dayAnswers change, and the date says which version you saw
Run number1 to 5 for each promptShows how much one prompt wobbles
Firms namedEvery tracked firm in the answer, yours includedThe raw count behind all three percentages
Domains linkedEach cited source, when the tool shows themNeeded only for citation share
NotesLocation, signed in or out, anything oddNext quarter’s you won’t remember

What do you do with a low number? Fix the inputs, not the dashboard. In my view the work sits in two places, the sites assistants draw from, which we cover in what the ChatGPT source studies show, and the mentions of you on other sites, which our post on off-site brand mentions explains.

No vendor page I read promises a timeline for moving the number. Don’t accept one from anybody else.

Want a second opinion on your prompts and rivals before you count? Book a growth strategy call. We sell this work, and we’ll still tell you when a spreadsheet is enough.

Before You Buy a Tracker

So which number counts as a good one?

Your own last quarter, measured the same way. Ahrefs says there isn’t a universally good AI share-of-voice score, and Similarweb says there’s no universal benchmark.

Compare like with like. Read any published range only after you’ve checked what it counted.

Is this the same as citation share?

No. Share of voice counts mentions. Citation share counts how often your pages are linked as a source, and Similarweb says a brand can score high on one and near zero on the other.

Track mentions first. Add citations once the mention count holds steady.

How many prompts do I need before I trust it?

More than feels necessary. My rule of thumb is 20 to 30 prompts, run three to five times each, per platform. That’s a judgment call, not a study.

SparkToro’s authors say their own sample needed to be far larger, and the plus or minus 9 points above is a floor, so keep the range next to the number.

Does the platform I test change the answer?

Google says AI Mode and AI Overviews use different models and techniques, so the responses and links vary from one to the other.

Count each platform on its own. Blend them only if you can say how, in one sentence.

Can a spreadsheet do this, or do I need a tracker?

A spreadsheet can handle a small prompt set. It’s slow, and it’s yours.

A tracker earns its fee when you need many prompts, frequent runs, or several markets. Ask the vendor for its denominator before you ask for the price.

How often should I run it?

Quarterly for the full sheet. Monthly if you’ve just changed something you want to see move.

Weekly runs chase noise, given how much one prompt wobbles from run to run.

AI Share of Voice AI Visibility Answer Engine Optimization SEO & AEO
Share
READY TO ACT?

Find out how often AI answers name you

Bring us your buyers’ questions and the three to five firms they shortlist, and we will talk through the prompt set, the formula and what a first count should look like.

Book a Growth Strategy Call