What Is AI Share of Voice and How Do You Measure It?

AI share of voice is your brand's fraction of all brand mentions in the AI answers to a defined set of questions. If the engines name your company and your competitors a combined forty times across your prompt set, and ten of those mentions are yours, your share of voice is one quarter. Our glossary entry gives the short version.

The important word is relative. Share of voice does not ask whether you appear; it asks how much of the conversation you hold compared with the alternatives. That makes it the closest AI-era equivalent of market share in attention, and it is a number content can move: the authors of the Generative Engine Optimization paper reported that targeted content changes can boost visibility by up to 40% in generative engine responses. It also means the number depends heavily on who you count as a competitor and which questions you ask, a point that comes up again in the pitfalls below.

You can get a quick read on your competitive position with the free AEO Grader scan, which includes a competitor and sentiment snapshot from one Gemini check grounded in Google Search. That snapshot is an estimate. A measured share of voice needs answers from each engine, which is what the rest of this article is about.

Three AI metrics get used interchangeably, and mixing them up leads to bad decisions. Visibility, sometimes called mention rate, is absolute: the fraction of answers that name you at all. Share of voice is relative: your mentions divided by everyone's mentions. Citation share is about links rather than names: how many of the sources an answer cites point at your website.

Microsoft now reports a version of the third metric. Its AI Performance report, introduced in February 2026, counts how often your content is cited in Copilot and Bing's AI summaries, and in June 2026 Bing added Citation Share to Bing Webmaster Tools, defined as the percentage of citations attributed to your site out of all citations shown across all sites. Useful as it is, it measures your pages being used as sources in Microsoft's AI surfaces, not your brand being named or recommended.

The gaps between the three are informative. High visibility with low share of voice usually means every answer lists a crowd of vendors. High share of voice with low citation share means engines recommend you based on what other sites say about you, not on your own pages.

A manual measurement takes an afternoon the first time and an hour a month after that. The method is the same whether you track one engine or four; the only difference is how many rows end up in the spreadsheet. Decide on every rule before you run the first prompt, and write the rules down so next month's count is comparable with this one.

A worked illustration: say you run eight prompts on one engine. You are named in three answers. Across the same eight answers, competitors collect nine mentions in total. Your share of voice on that engine is three divided by twelve, one quarter, while your visibility is three of eight answers. Calculate each engine separately, then pool the counts across engines for an overall figure rather than averaging the four percentages, so an engine that answered fewer prompts does not get the same weight as one that answered all of them.

The first pitfall is sample size. With twenty answers, a single extra mention moves your visibility by five percentage points, so a month-on-month swing of that size can be pure chance. Statisticians handle this with a confidence interval around a proportion; the NIST/SEMATECH Engineering Statistics Handbook walks through the Wilson method for calculating confidence intervals for proportions. You do not need the formula to act on the lesson: small prompt sets give wide error bars, so look for movements that persist over several measurements.

The second is run-to-run variance. Researchers at Thinking Machines Lab note that asking ChatGPT the same question multiple times provides different results, and that LLM APIs are still not deterministic in practice even at temperature zero. A single run per prompt is a snapshot of a moving target.

The third is prompt bias. A question that contains your brand name will almost always produce an answer that mentions you, which inflates share of voice if those prompts dominate the set. Unbranded category questions are where a mention has to be earned, so report them separately. Finally, location and wording change answers: OpenAI's web search guide lets developers pass an approximate user location to refine results, so a local business should test the way its local buyers ask.

The $10 AEO Grader report follows the manual method above, automated and applied to four engines. It starts by building a profile of who your company is from your website, so that the questions it sends identify the right business when your name is shared with others. It then sends the same panel of buyer questions to ChatGPT, Claude, Perplexity and Gemini, each with its live web search switched on. The panel asks what you offer, whether you are trustworthy, how you compare with competitors, what the alternatives to you are, and which options are best in your category.

For every answer, the report looks for your brand name, tolerating a one-letter typo in longer names, and then a separate check confirms that the answer is describing your company rather than a namesake. Only confirmed answers count as a mention, and answers about a different business are labelled as such. Competitors come from two places: a starting list and the companies each engine itself named in its answer, with your own brand and its variants removed. Each competitor counts once per answer.

Share of voice is then answers naming you, divided by answers naming you plus competitor mentions. It is calculated per engine and, for the overall figure, pooled across every answered question rather than averaged across engines. Calls that fail are left out of the counts, and a report is only delivered once all four engines have answered at least four questions each. When no competitor is named at all, you see a note to that effect instead of a percentage.

A share-of-voice figure on its own invites the question "why?", so the report puts the evidence beside it. You see visibility and share of voice for each engine, how many times each engine cited your own site, and visibility broken down by the type of question asked. That breakdown is the answer to the prompt-bias pitfall: it shows whether you are named only when asked about by name, or also when a buyer asks for the best options in your category.

You also get the list of competitors the engines named with their mention counts, the domains cited most often across all answers, and every engine's actual answer to each question with its sources, so any number can be traced back to the text that produced it. A site audit with a prioritized fix list comes with it. The report does not repeat each question several times, so the honest way to read the number is as one careful measurement; the included re-scan within 14 days shows how visibility and share of voice changed since the first run.

For a trend line, monthly tracking at $9 a month runs a fresh four-engine report every month, emails you what changed per engine, and alerts you if a competitor is being named more often than you. The full scoring approach is on the methodology page.

Share of voice earns its keep when it changes what you do next. Treat it as one of a small set of metrics rather than a headline on its own; our guide to measuring AEO performance shows how it fits alongside citation accuracy, AI referral traffic and sentiment in a monthly report. Then use the per-engine and per-question breakdowns to choose the work.

If your share is low on one engine only, the cause is usually access or evidence specific to that engine; our explainer on why AI engines give different answers covers the usual suspects. If it is low everywhere on category questions, you need third-party coverage and pages that answer those questions directly. If you want to sanity-check a single engine yourself first, the manual routine in does ChatGPT recommend my business? takes twenty minutes. And when you are ready for a measured baseline, start with the free AEO Grader scan and move to the four-engine report when you need the real answers behind the number.