AI share of voice: how to measure it across ChatGPT, Gemini and AI search

AI share of voice is how often AI assistants name your brand when people ask the questions your buyers ask, compared with how often they name your competitors. It is the AI-answer version of the old advertising and search metric: of all the attention in a category, how much goes to you. Some teams call it “share of model”.

The idea is simple, but the measurement is not. AI answers change from run to run. In a January 2026 study of 2,961 runs across ChatGPT, Claude and Google's AI, SparkToro and Gumshoe found less than a 1 in 100 chance of getting the same list of brands twice, and the order was even less stable.^[1] The same study found that how often a brand appears across many runs is a reasonable metric, even though any single answer is unreliable.^[1] That is why share of voice is always a rate over many answers, never a reading of one.

Four numbers cover almost every question a team asks. Measure them separately, because they move for different reasons and point to different fixes.

Mention rate

This is your visibility on its own terms. It does not depend on which competitors you track, so it is the most stable headline number. Mention rates across brands do not add up to 100%, because one answer can name five brands.

Share among named brands

This is the classic share of voice. It adds up to 100% across the brands you track, which makes it easy to chart, but add or remove a competitor and everyone's share changes even if no answer did. Fix the competitor set up front and change it only on purpose.

Citation share

Engines with web search (Perplexity, ChatGPT search, Google AI Overviews and AI Mode) list the pages they used. Citation share tells you whether your own site is one of them. A review site can be cited in an answer that names you, and your blog in one that recommends a rival. Google says AI Overviews and AI Mode may run several related searches for one question (“query fan-out”), so one answer can cite many pages.^[4]

Average position

Being named first matters more than being named fifth. Research on generative engines weights a source by where it appears, giving earlier mentions more credit.^[2] Average position is useful, but only next to mention rate: a brand named once, first, has a perfect average position and almost no visibility. Because the order of lists is so unstable,^[1] treat position as a secondary signal.

Here is a made-up example for one engine in one market. A project management software company tracks three rivals on 20 unbranded questions, each asked in three wordings, for 60 answers in total. The answers name the four tracked brands 99 times between them and cite 240 pages.

How to read it: Rival B leads clearly. Its range does not overlap yours, so the gap is real. Rival C looks ahead of you, but your ranges overlap, so with 60 answers you cannot say C is truly more visible. Your citation share beats C's although C is mentioned more: engines read your site but lean on other sources when choosing whom to recommend. Next, check which sites they cite on the questions where B appears and you do not.

The ranges here are simple Wilson intervals, a standard method for proportions that behaves better than the textbook formula at small samples.^[3] Glossy Data's Honest Score uses a pooled model, so its ranges are a little different, but the logic is the same.

Your share of voice is only as good as the questions behind it. A good set reads like real buyers, covers the category and stays fixed long enough to show a trend.

1. Start from buyer language. Use wording from sales calls, support tickets and site search. Write questions, not keywords: “best accounting software for a small agency with multi-currency invoicing”.
2. Cover the journey. Mix early research (“how do I choose a CRM”), shortlist questions (“best CRM for real estate teams”), comparisons (“HubSpot vs Pipedrive”) and late-stage checks (“CRM with a Gmail integration and monthly billing”). Questions close to a decision deserve more weight.
3. Keep branded questions apart. If you are Acme, “Is Acme any good?” will name you nearly every time. Track those answers for accuracy, but leave them out of the score.
4. Paraphrase. Ask each question in three to five natural wordings.
5. Localize by market. A buyer in Germany asks in German and gets different sources. Write native questions per market instead of translating word for word.
6. Size it. Twenty to fifty core questions per segment is a practical start. Many questions across a topic beat many repeats of one.
7. Version it. Date your set. When you add or retire questions, note it so a jump in the chart is not mistaken for progress.

More than most people expect. The table shows the 90% range around a 50% mention rate for different numbers of answers, using the Wilson method.^[3] The range is widest at 50% and narrower near 0% or 100%.

Practical rules: do not report a weekly change smaller than the range; pool a few weeks of answers when volumes are low; and always show the number of answers next to the percentage. A 6-point weekly move on 60 answers is usually noise.

• Mixing branded and unbranded questions. The quickest route to a flattering, useless number.
• Reading too much into small samples. A screenshot of one ChatGPT answer proves nothing about the next one.^[1]
• Blending engines. Engines use different models and sources, so a blended number can rise while your most important engine falls. Show each engine, then a weighted total with visible weights.
• Blending markets and languages. Visibility in the US says little about Brazil or Japan. Keep markets separate and label any market asked in English.
• Moving the competitor set. Adding a big rival lowers everyone's share among named brands overnight.
• Treating position as rank. AI answers are not ten blue links. Lists are reordered constantly,^[1] so a single “rank 1” claim is not meaningful.
• Counting citations as recommendations. Being cited means the engine read you. Being named means it recommended you. Track both.

Report a few stable numbers with their uncertainty, per engine and market, over a rolling window of a few weeks: mention rate with a range for each engine, share among named brands, the top cited domains, and the questions where you gained or lost. Label changes inside the range as within noise.

The most useful view is the gap list: questions where a rival is named and you are not, with the pages the engine cited. It turns a chart into a to-do list. See competitor share of voice and sources and fixes for how Glossy Data lays this out, and the methodology for the statistics.

Glossy Data asks paraphrased, unbranded questions on ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity and Claude every week, in 49 countries, with native question packs in 7 languages. Every number carries its sample size and a 90% range, and changes inside the noise are labeled. Branded questions are shown but not scored. Paid plans add competitor share of voice and a Source Map of the sites each engine cites. Microsoft Copilot is coming soon.