AI Search Share of Voice: How to Measure Your Brand

AI search share of voice is the share of a fixed set of AI answers in which your brand appears, measured against a stated denominator: a frozen question set, named engines, one market and one time window. It is four separate shares, not one, because being named, being cited, being recommended and being named beside competitors are different events.

AI search share of voice: the short answer

1. Write the denominator before you count anything. Question set, engines, market, time window, and how many answers came back. Change any of them and you have a new metric, not a new month.
2. Keep four shares apart. Mention share, owned citation share, recommendation share and competitor co-mention share answer four different questions. One blended number hides the one you need.
3. Pick one of two calculation methods and label it. Answer share divides by answers. Brand-pool share divides by all brand mentions. The same data gives two different percentages, and both are correct for their own question.

This page is the umbrella. If you only want the single metric Prime AI Visibility calls share of citation, our explainer on share of citation and how to read a movement covers that one number in depth. Here we define the family it belongs to, show the arithmetic, and lay out a dashboard you can build in a spreadsheet.

What share of voice meant before AI answers

Share of voice is an old advertising measure. In a 1990 Harvard Business Review article, John Philip Jones compared a brand's share of category advertising spend with its share of market, and the comparison only worked because both sides were countable: dollars spent, units sold. Paid search kept the idea and changed the unit. Google Ads defines impression share as impressions received divided by total eligible impressions. Again, a countable event and a defined pool.

That is the pattern to keep. A share needs a unit you can count and a pool you can define. Generated answers break the old units. There is no impression log you can download, no ordered list with a position number, and the same question can produce a different answer an hour later. Google's own documentation for AI features says the set of responses and links shown in AI Overviews and AI Mode will vary. So the pool has to be built by you: a fixed set of questions, asked of named engines, on recorded dates. The answers that come back are the pool. Everything below is a share of that pool.

Why the AI version needs four shares, not one

Inside a single answer, four different things can happen to your brand, and each one deserves its own count. Prime AI Visibility records the first three events under the names Brand named, Website cited and Explicit recommendation, and keeps them separate on purpose, because naming or citing a brand does not establish a recommendation. The shares built from those counts are your calculation, not a labelled metric in the product.

Share	Formula	Question it answers
Mention share	Answers naming the brand (Brand named) ÷ relevant answers	How often are we part of the answer at all?
Owned citation share	Answers with at least one link to our website (Website cited) ÷ relevant answers	How often does our own site appear as a source?
Recommendation share	Answers that clearly recommend us (Explicit recommendation) ÷ relevant answers	How often are we the pick, not just a name?
Competitor co-mention share	Answers naming us and at least one competitor ÷ answers naming us	When we appear, how often is it in a crowd?

Mention share is the number most vendors mean when they say AI share of voice. In our editorial articles it is called share of citation, the percentage of relevant answers that name the brand at least once; that is an editorial term, not a metric label inside Prime AI Visibility's reports. Relevant means answers where naming a brand is on-topic for the question. An answer that asks you to clarify, refuses, or wanders off the category is not in the denominator.

Owned citation share counts a different event: at least one qualifying link to your website inside the answer. An answer can name you without linking to you, and it can link to your pricing page without spelling out your name. Prime AI Visibility counts each answer once under Website cited, and counts the individual links separately as Source links, so ten links in one answer do not inflate the share.

Recommendation share is the strictest. The answer has to clearly recommend you. "Harbor Payroll, Ledgerline and PayNorth are common choices" is a mention. "For a 20-person firm, Harbor Payroll is the simplest option" is a recommendation. Most brands find this number is a fraction of their mention share, which is exactly why it must not be blended in.

Competitor co-mention share uses a different denominator on purpose. Its base is the answers that already name you, and it asks how often a rival stands next to you in those answers. A high co-mention share with a low recommendation share is a specific, actionable pattern: you are in the consideration set and losing the comparison.

The denominator: write the scope line first

Most disagreements about AI search share of voice are denominator disagreements in disguise. The fix is boring and effective: write a scope line at the top of the report before any percentage, and never publish a share without it.

A scope line has five parts:

1. Question set. A fixed, versioned list. "40 buyer questions, set v3, unchanged since 1 July 2026."
2. Engine set. Named tools, and which mode. "ChatGPT, Perplexity, Claude, Gemini; consumer defaults; API-based collection." Say if a Google surface is a live capture or a grounded preview, because they are not the same thing.
3. Market and persona. Language, country, and any persona instruction used in the prompt.
4. Time window. Dates of the runs, and how many repeat runs per question.
5. Answer accounting. Requested answers, returned answers, relevant answers, missing answers. Percentages then state which of these they divide by.

Prime AI Visibility's published metric definitions follow the same rule: reports identify the questions, tools, market, time period, requested answers, returned answers, usable answers and missing answers, and every percentage uses its stated denominator. Missing answers matter because a share that divides by returned answers goes up when answers go missing, without anything improving. If 40 questions across four tools request 160 answers and 6 never returned, the report shows 160, 154 and the relevant count side by side instead of quietly dividing by 154. Prime AI Visibility applies the same discipline to its Score: missing requested answers cannot improve it.

A scope line also decides what you may compare. Two reports are comparable when the question set, tools, persona, market and measurement method line up. If they do not, the honest label is "different scope", not "down 5 points".

Two ways to calculate it, and why they disagree

Read five vendor blogs on AI share of voice and you will find two formulas presented as if they were one. They are not, and the difference is the most useful thing on this page.

Method A: answer share

Count answers. For each relevant answer, record a 1 if the event happened (brand named, website cited, or explicitly recommended) and a 0 if it did not. Divide by the number of relevant answers.

Mention share
  = answers naming the brand
  ÷ relevant answers × 100

Owned citation share
  = answers linking to our website
  ÷ relevant answers × 100

Recommendation share
  = answers clearly recommending us
  ÷ relevant answers × 100

Co-mention share
  = answers naming us and a competitor
  ÷ answers naming us × 100

Each answer counts once. A brand named three times in one answer is still one answer. These are the per-answer events Prime AI Visibility records as Brand named, Website cited and Explicit recommendation, and answer share is the method to use when the question is "how often does a buyer who asks this see us?"

Method B: brand-pool share

Count brand appearances. Across all relevant answers, tally every brand that is named, counting each brand once per answer. Your share is your appearances divided by all brand appearances in the pool.

Brand-pool share
  = answers naming our brand
  ÷ sum over all tracked brands of
    (answers naming that brand) × 100

This is the closest cousin of the advertising original, because it is a share of a pool that adds up to 100 percent across all brands. It is the method to use when the question is "of all the brand attention in these answers, how much is ours?" It moves when a competitor gains even if your own count is flat, which is either the point or a trap, depending on what you were asked.

Why they rarely match

Method A divides by relevant answers. Method B divides by brand appearances. The two are equal only when the pool of brand appearances happens to equal the number of relevant answers, one brand per answer on average. Whenever answers name more than one brand on average, which is what happens when buyers ask for options, Method B's denominator is bigger and its share is smaller; when most answers name no brand at all, the reverse can happen. Method B can also fall while Method A rises, because a competitor's gain enlarges the pool. Neither is wrong. A dashboard that shows one of them without saying which, or that switches between them from one quarter to the next, is producing noise with a percent sign on it.

There is a third variant worth naming so you can recognise it: weighted share, where each question is weighted by an estimate of how often buyers ask it. It can be useful for prioritisation, but the weights are guesses, and a weighted share is only defensible if the weights are published beside it and frozen for the same window as the question set. If a vendor cannot show you the weights, treat the number as a rank of their opinion, not a measurement.

A worked example with real arithmetic

The brand below is fictional. The arithmetic is not.

Harbor Payroll sells payroll software to small accounting firms. Its scope line: 40 buyer questions (set v3), four tools (ChatGPT, Perplexity, Claude, Gemini), United States, English, no persona, one run per question in the week of 1 to 7 September 2026. That requests 160 answers. 154 returned. 149 were relevant; 5 asked for clarification or answered a different question.

Counts from the 149 relevant answers:

Event	Count
Answers naming Harbor Payroll (Brand named)	57
Answers linking to harborpayroll.example (Website cited)	31
Answers clearly recommending Harbor Payroll (Explicit recommendation)	12
Answers naming Harbor Payroll and at least one competitor	41
Brand appearances, all five tracked brands, one per brand per answer	335

Method A shares:

• Mention share: 57 ÷ 149 = 38.3%
• Owned citation share: 31 ÷ 149 = 20.8%
• Recommendation share: 12 ÷ 149 = 8.1%
• Competitor co-mention share: 41 ÷ 57 = 71.9%

Method B share:

• Brand-pool share: 57 ÷ 335 = 17.0% (the 335 is Harbor Payroll 57, Ledgerline 96, PayNorth 74, Quillpay 58, Tallyhaus 50)

Both 38.3% and 17.0% describe the same week. The first says Harbor Payroll is in roughly four of ten answers. The second says it holds about one sixth of all brand attention in those answers, because Ledgerline and PayNorth appear more often. A board slide that shows "AI share of voice: 38%" and a competitor deck that shows "17%" are both right and are answering different questions. Label them.

The per-engine split is where the work starts:

Tool	Relevant answers	Answers naming the brand	Mention share
ChatGPT	38	18	47.4%
Perplexity	37	14	37.8%
Claude	37	13	35.1%
Gemini	37	12	32.4%

Fifteen point gap between the strongest and weakest tool. An average would have hidden it. The next question is which sources each tool displayed in the answers that named a competitor and not you, which is the reason to keep the answers themselves, not just the counts.

How many questions before a change means anything?

A share is a proportion estimated from a sample, and small samples move on their own. The width of that uncertainty can be estimated. Using the Wilson score interval, the method for proportions presented in the NIST/SEMATECH Engineering Statistics Handbook, and treating each answer as an independent draw, which a fixed question set with repeated runs only approximates, the 95% interval around Harbor Payroll's mention share is:

• 57 of 149 answers: 38.3%, interval roughly 30.8% to 46.3%
• If the same brand had run only 40 answers and been named in 15: 37.5%, interval roughly 24.2% to 53.0%

Treat these as illustrations of sampling noise under that assumption, not a guarantee about the market. Even so, read that second line twice. With 40 answers, a reported move from 37% to 45% is inside the range the sample alone could produce. Three habits keep you honest:

1. Report the denominator next to every share, so a reader can see 15 of 40 for what it is.
2. Repeat runs. Two or three runs per question in the same window shrink the noise and show which answers are stable.
3. Compare like with like. Only put two shares on the same chart when their scope lines match. Prime AI Visibility labels such a comparison a Comparable change and reports observed gains, losses and recoveries without claiming that a content edit caused them.

None of this requires statistics software. A spreadsheet with the counts and the interval formula is enough, and the discipline matters more than the decimals.

A sample dashboard

One page, five blocks, no blended score. Every number carries its numerator and denominator.

Block	What it shows	Example
Scope line	Question set version, tools, market, window, requested / returned / relevant answers	v3, 4 tools, US-EN, 1 to 7 Sep 2026, 160 / 154 / 149
Four shares	Mention, owned citation, recommendation, co-mention, each as count and percent, with the previous comparable window beside it	57/149 = 38.3% (prior 52/147 = 35.4%)
Per-tool split	Mention share by tool, same layout	ChatGPT 18/38, Perplexity 14/37, Claude 13/37, Gemini 12/37
Competitor pool	Each tracked brand's answer count and brand-pool share	Ledgerline 96 (28.7%), PayNorth 74 (22.1%), Quillpay 58 (17.3%), Harbor 57 (17.0%), Tallyhaus 50 (14.9%)
Evidence and gaps	Count of missing answers, link to the saved answers, the list of questions where competitors were named and you were not	6 missing; 23 gap questions listed

Two design rules. First, the comparison column only appears when the prior window is comparable; otherwise it shows "new scope". Second, nothing on the page is a rank, a traffic estimate or a forecast, because the underlying data cannot support any of the three. If you report to executives, the one-page executive reporting method shows how to present movement without claiming your edits caused it, and the same rule applies here.

Mistakes that make the number meaningless

• Changing the question set mid-series. Adding five questions where you already do well raises the share without anything changing in the world. Version the set, and start a new series when it changes.
• Mixing methods. Method A last quarter, Method B this quarter, one line chart. This happens more often than it should when a team switches tools.
• Counting links as mentions, or mentions as links. They are different events with different denominators. Keep Brand named and Website cited apart, exactly as the published definitions do.
• Forgetting brand aliases. "Harbor Payroll", "Harbor", "HarborPay" and a misspelling are all the same brand to a buyer. Approve the variants once and normalise before counting.
• Averaging across tools. A 15 point gap between engines is the finding. The average is where the finding goes to disappear.
• Borrowing a benchmark. A published "average AI share of voice for SaaS" was measured on someone else's question set, engines and dates. It is not your denominator. Our guide to benchmarking your brand's AI citations against competitors explains how to build a fair one instead.
• Reading a share as a rank. No engine publishes an ordered list of brands inside an answer. A share tells you how often you appear, not where.
• Claiming cause. A share that rose after a page rewrite is an observation. The engines do not document why one answer names a brand and the next does not, and the same prompt can vary between runs.

Doing it by hand versus with a tool

You can run this whole method manually for one engine. Our walkthrough of monitoring brand mentions in ChatGPT shows the record you keep per answer, and the Gemini equivalent covers the same discipline on a second engine. Forty questions on four tools is 160 answers a month, which is many hours of careful work, and the hard part is not the asking but the consistent classification of each answer into named, cited, recommended and co-mentioned.

Prime AI Visibility is built around the same record, and we should be plain about what it does and does not do. It runs the questions you choose across ChatGPT, Perplexity, Claude, Gemini and an AI Overview-style preview using Google-grounded answers, which is not a capture of the live Google Search AI Overview panel. Monthly checks are the default; weekly and temporary daily schedules are optional, and on-demand checks use credits. It saves the returned answers and reports Brand named, Website cited, Source links and Explicit recommendation separately, each against its stated denominator, and its reports identify requested, returned, usable and missing answers. It does not report rank, AI traffic estimates, forecasts or causal attribution. API-based observations can differ from what a person sees in a consumer app, and the step-by-step description of how Prime AI Visibility works says so. What it gives you for this method is the counts and the denominator. Share of voice, co-mention share and the brand pool are not labelled metrics in the product; you calculate them from the saved evidence, and deciding which answers are relevant and which competitors appear in each one still takes a human read of the saved answers.

Methodology and disclosure

Method note, 24 September 2026. This article was written by the Prime AI Visibility editorial team, and Prime AI Visibility sells the measurement software described in the previous section, so read the product paragraphs as a vendor's description of its own tool. The metric names and their definitions were checked against Prime AI Visibility's published metrics, how-it-works and pricing pages on 24 September 2026, and the outside sources below on the same date. The worked example uses a fictional brand and invented counts so the arithmetic can be followed; it is not a client dataset. We ran no vendor's software for this piece and make no claim about any engine's internal workings.

References

1. John Philip Jones, Ad Spending: Maintaining Market Share, Harvard Business Review (January to February 1990). https://hbr.org/1990/01/ad-spending-maintaining-market-share
2. Google Ads Help, About impression share: impression share = impressions ÷ total eligible impressions (checked 24 September 2026). https://support.google.com/google-ads/answer/2497703
3. Google Search Central, AI features and your website, noting that the responses and links shown in AI Overviews and AI Mode will vary (checked 24 September 2026). https://developers.google.com/search/docs/appearance/ai-features
4. NIST/SEMATECH, e-Handbook of Statistical Methods, section 7.2.4.1, confidence intervals for a proportion and the Wilson method (checked 24 September 2026). https://www.itl.nist.gov/div898/handbook/prc/section2/prc241.htm
5. OpenAI Help Center, Searching the web with ChatGPT (checked 24 September 2026). https://help.openai.com/en/articles/9237897-chatgpt-search
6. Anthropic Help Center, Using web search on Claude (checked 24 September 2026). https://support.anthropic.com/en/articles/10684626-using-web-search-on-claude
7. Prime AI Visibility, Metrics and methodology, reporting vocabulary and what Prime does not report (checked 24 September 2026). https://primeaivisibility.com/metrics
8. Prime AI Visibility, How it works and Pricing (checked 24 September 2026). https://primeaivisibility.com/how-it-works and https://primeaivisibility.com/pricing

Next steps

1. Read the share of citation explainer if your team already reports one number and you want to be sure which denominator it uses.
2. Build a fair competitor benchmark before you put any share next to a rival's.
3. Write your scope line, then create a Prime AI Visibility workspace and run the first check on the questions you just fixed.