How-to

AI Share of Voice: How to Measure It in AI Search Answers

AI share of voice is the percentage of all brand mentions, across a fixed set of AI answers, that belong to your brand. To measure it, run the same buyer questions many times on each AI engine, count mentions for you and named competitors, divide yours by the total, and report the number of answers alongside.

Key points

  • AI share of voice is your mentions divided by all tracked brands' mentions across a fixed set of AI answers.
  • Visibility is the share of answers naming you, so report it next to share of voice because the two can move apart.
  • A single AI answer proves nothing, since the SparkToro and Gumshoe study of January 2026 found brand lists repeated in under 1 in 100 runs.
  • By our Wilson calculation, an observed 30% spans 10.8% to 60.3% on 10 answers and 25.1% to 35.4% on 300.
  • Freeze your prompts, engines and competitor list, and only call a change real when it exceeds the interval.

What is AI share of voice?

AI share of voice is your brand's share of all brand mentions across a fixed set of AI answers. If assistants named brands 120 times in your sample and 30 of those mentions were yours, your share is 25%.

The metric is relative by design. It only means something next to a named list of competitors, a named set of prompts and a named set of engines. Change any of the three and you have a different number.

There is no agreed standard yet. Vendors define the term differently, and some pages are not consistent with themselves. AirOps' guide to share of voice divides brand appearances by total answers in one example, then by all brand citations in its FAQ. Those are two different metrics, so always ask a tool or agency which denominator it uses.

How is it different from SEO share of voice and from visibility?

SEO share of voice estimates your share of clicks from a set of keywords. AI share of voice counts brand mentions inside generated answers. Visibility is simpler still: the share of answers that name you at all, with no competitor in the sum.

Ahrefs describes share of voice as a measure that began in advertising and now covers several channels. Its general formula is brand visibility over total market visibility. For organic search, its rank tracker reports the percentage of clicks from tracked keywords that land on each site. That article does not cover AI answers at all.

MetricWhat it countsDivided byQuestion it answers
SEO share of voiceEstimated clicks from tracked keywordsAll clicks on those keywordsHow much search traffic do we win?
AI visibility (mention rate)Answers that name your brandAll answers collectedHow often are we in the answer?
AI share of voiceMentions of your brandMentions of you plus tracked competitorsHow much of the answer space is ours?
Citation rateAnswers that link to your siteAll answers collectedHow often are we used as a source?

The two AI numbers can move apart. Your visibility can stay flat while your share of voice falls, because answers have started naming more rivals. LLM Pulse's guide (July 2026) draws the same line: share of voice is relative, mention rate is absolute. Report both.

How do you calculate AI share of voice?

Divide your brand's mentions by the total mentions of every tracked brand in the same answers, then multiply by 100.

AI share of voice (%) = your brand mentions / mentions of all tracked brands x 100

Set one counting rule before you start. The simplest is to count a brand once per answer, even if the answer names it three times. Then the mention count for a brand equals the number of answers naming it.

The table below is an invented example, not real data. It assumes 25 prompts, each run four times on one engine, giving 100 answers.

Brand (example only)Answers naming itVisibilityShare of voice
Your brand3030%25.0%
Competitor B4545%37.5%
Competitor C2525%20.8%
Competitor D2020%16.7%
Total120 mentionsn/a100%

Your brand appears in 30% of answers but holds 25% of the mentions. Add a fifth competitor to the tracked list and your share drops again, although nothing in the answers changed. That is why the competitor list has to stay fixed.

Why does a single AI answer mislead?

AI assistants can give a different answer every time you ask. One answer is a single draw from a range of possible answers, not a ranking you can screenshot.

This is documented behaviour. OpenAI's API guide says chat completions are non-deterministic by default. A seed setting makes output mostly consistent, and backend changes can still alter it.

The best public evidence for brand lists comes from a SparkToro and Gumshoe study published in January 2026. About 600 volunteers ran 12 prompts through ChatGPT, Claude and Google's AI, for 2,961 runs. The chance of getting the same list of brands twice was under 1 in 100. The chance of the same list in the same order was about 1 in 1,000.

Rand Fishkin's conclusion was that ranking position in AI answers is unreliable. He judged visibility percentage, measured over many prompts and many runs, to be a reasonable metric, though not a precise one.

Treat the study with care. It was not peer reviewed, and its co-author works at an AI tracking company, which the post discloses. Participants used default settings. Search Engine Land's write-up lists the open questions, including how many runs you need and whether API answers match what real users see. Nobody outside the AI companies can settle that second point today.

Google's surfaces differ from each other too. Google's documentation says AI Overviews and AI Mode may use different models and techniques, so responses and links vary. Our comparison of AI Overviews and AI Mode for brands covers this in more detail.

How many answers do you need for a number you can trust?

You need more than most teams collect. With 10 answers, an observed 30% could plausibly sit anywhere from about 11% to 60%. With 300 answers, that range narrows to about 25% to 35%.

A confidence interval is the range of true rates that fits your sample. A 95% interval comes from a method that captures the true rate in about 95 of every 100 samples. A wide interval means you do not know much yet.

For a yes or no measure such as "did this answer name us", use the Wilson interval. E.B. Wilson introduced it in 1927, and the NIST statistics handbook describes it. Its lower limit cannot fall below zero. The simpler textbook formula can shrink to zero width near 0% or 100%, which falsely suggests certainty.

We computed the table below ourselves in Python, using the Wilson formula at 95% confidence for an observed rate of 30%.

Answers collectedAnswers naming you95% Wilson intervalWidth
10310.8% to 60.3%49.5 points
30916.7% to 47.9%31.2 points
1003021.9% to 39.6%17.7 points
3009025.1% to 35.4%10.3 points

Two things follow. On 30 answers each, a brand at 30% and a rival at 38% cannot be told apart. And you need roughly four times the answers to halve the width.

There are two honest limits. The interval fits visibility cleanly, because each answer is one yes or no. For share of voice it is an approximation, since mentions inside one answer are not independent. It also covers run-to-run noise only. It says nothing about whether you picked the right prompts.

Which prompts and engines should you track, and what goes wrong?

Track the questions your buyers ask, on the assistants they use, and then freeze the list. A stable set matters more than a perfect one.

  1. Collect real questions. Use sales calls, support tickets, site search and Search Console queries. Write them as a customer would type them.
  2. Settle on a set size. LLM Pulse suggests 30 to 100 prompts, mixing category, comparison and branded questions. That is a vendor rule of thumb, not a tested threshold.
  3. Separate branded prompts. Prompts that contain your name are not a fair test of whether assistants recommend you. Report that group on its own.
  4. Fix the competitor list. Name the rivals you will count and write down the counting rule.
  5. Choose engines by audience. Pick the assistants your customers use. Report each engine separately before you blend them.
  6. Run every prompt several times. Keep the same schedule, and store the date, engine and full answer text.

Do not agonise over exact wording. In the SparkToro study, real people's prompts for the same need barely resembled each other. Search Engine Land's summary notes the tools still returned similar brand sets for the same intent. Cover intents, not phrasings.

If your share is low, the fixes sit elsewhere. See our guides on getting your brand mentioned in ChatGPT and showing up in Google AI Mode and Gemini.

Common mistakes

  • Using one screenshot as proof of anything.
  • Reporting rank position inside an answer as if it were stable.
  • Changing prompts or competitors halfway through a series.
  • Mixing branded prompts into the headline figure.
  • Blending engines into one number and hiding the differences.
  • Quoting a percentage without the number of answers behind it.
  • Counting mentions and linked citations as the same thing.
  • Reading share of voice as market share. This guide makes no claim about its effect on sales.

How to measure whether it worked

Compare like with like: the same prompts, engines and competitors, pooled over enough answers that the change is larger than the interval. Anything less is noise.

Start with a baseline of two to four weeks before you change anything. Then pool answers into weekly or monthly windows. LLM Pulse argues that daily readings are noisy and monthly ones lose signal. Daily collection is still useful, because it builds the sample. Twenty-five prompts run once a day give 175 answers per engine each week.

Judge changes against the table above. A move from 30% to 33% on 100 answers per period sits well inside a 17.7 point interval. A move from 30% to 45% on 300 answers per period does not. Keep a dated log of what you published or changed. Models and their settings change, so a shift may come from the engine and not from you.

Traffic data will not confirm it neatly. Google says clicks from AI Overviews and AI Mode are counted within the Web search type in Search Console, with no separate breakdown described. Watch branded search and direct enquiries alongside.

A spreadsheet can do all of this if you are disciplined. Our walkthrough on tracking brand mentions in ChatGPT shows the manual route. CiteSieve, the tracker Braivex makes, asks six AI engines your customers' questions once a day and stores every answer. It shows visibility as a share of answers with the answer count and a 95% interval. For other options, see our comparison of AI visibility tools and prices.

Questions and answers

What is a good AI share of voice?

There is no universal benchmark, because the figure depends on how many competitors you track. With four tracked brands, an even split would be 25% each. Compare your share with your own baseline and with your nearest rival on the same prompts and engines, not with numbers published for other industries.

Is AI share of voice the same as citation share?

No. Share of voice counts answers that name your brand in the text. Citation share counts answers that link to your website as a source. An assistant can recommend you without linking to you, and can cite your page while recommending someone else. Track the two separately and say which one you are reporting.

Can I measure AI share of voice by hand?

Yes, for a small set. Run each prompt several times in each assistant, paste the answers into a spreadsheet and mark which brands appear. It is slow, so keep the prompt set small and the schedule fixed. Record the date and engine, and keep the full answer text so you can recount later.

How often should I report AI share of voice?

Report on a window long enough to hold a useful sample, which for most small brands means weekly or monthly. Always show the number of answers and the interval beside the percentage. If a period has too few answers to separate you from a competitor, say so and wait for more data.

Sources

  1. OpenAI API docs: Advanced usage (reproducible outputs)
  2. Google Search Central: AI features and your website
  3. NIST/SEMATECH e-Handbook of Statistical Methods: Confidence intervals for a proportion
  4. SparkToro: AIs are highly inconsistent when recommending brands or products (January 2026)
  5. Search Engine Land: AI recommendation lists repeat less than 1% of the time
  6. Wikipedia: Binomial proportion confidence interval (Wilson score interval)
  7. Ahrefs: What Is Share of Voice? How to Measure It Across Channels
  8. AirOps: Share of voice
  9. LLM Pulse: Share of Voice in AI Search: How to Calculate It

Sources were read on 11 November 2026. Engines and their documentation change; check the linked page for the current wording.

More guides

Find out what AI says about your brand.

A site audit on the page in seconds, then live answers by email. No account, no card.