AI Share of Voice: How to Measure It Without Fooling Yourself

Three formulas hiding behind one label

Ask three vendors for your AI share of voice and you may get three numbers, all correct, because they measure different things. Before you set this month's figure against last month's, or against an agency's pitch, find out which one you are looking at. For a quick first read before any of this, the free AI visibility check runs one customer question through several engines live and shows who they named.

Bing uses the third definition for its own metric. Citation Share in Bing Webmaster Tools, added on June 16, 2026, is your site's percentage of all citations shown for the same grounding query in Copilot and Bing's AI summaries. It is first-party, free and, as Bing says plainly, observational: it does not reveal who took the rest. For most owners we lead with mention rate, because a business the engines do not know has a bigger problem than its citation mix.

Working the numbers on a sample prompt set

Here is a hypothetical with round numbers so the arithmetic shows. An HVAC company in San Antonio tracks 20 questions: emergency repair, new system quotes, brand comparisons, maintenance plans. It asks each question three times on four engines (ChatGPT, Google AI Mode, Gemini and Perplexity). That makes 240 answers a month.

1. 1

   Freeze the question list

   Write the 20 questions once and stop editing them. A changed word is a new question, and the trend line breaks. The prompt tracking guide covers how to choose them.

2. 2

   Save every answer in full

   Keep the answer text, the engine, the date, the location used and any cited links. Screenshots alone cannot be counted later.

3. 3

   Tag each answer

   Mark every business named, and note whether your company was recommended, merely listed or mentioned with a warning. Flag wrong facts as you go.

4. 4

   Compute the mention rate

   The company appears in 54 of 240 answers: 22.5%. Its largest competitor appears in 96: 40%.

5. 5

   Compute share of mentions

   The 240 answers name businesses 380 times in total. The company's 54 mentions are 14.2% of that, which tells you how crowded the answers are.

6. 6

   Split by engine and question type

   A blended 22.5% might be 40% in Perplexity and 5% in AI Mode. That split, not the total, tells you where to work next.

Nothing here needs special software. A spreadsheet handles 240 rows comfortably. Tools earn their fee when the set grows to hundreds of prompts, several cities or daily runs, and the GEO tools comparison lists what they charge as of October 2026.

Why the same question gets a different answer on Tuesday

A results page is a list that changes slowly. An AI answer is written fresh each time. Ask the identical question twice and the businesses named, their order and the sources cited can all shift. Several mechanisms push that way, and none of them means your work failed.

• Generation, not retrieval. The model composes new text on each run, so a borderline business appears in some answers and not in others.
• Fan-out searches vary. Google says AI Overviews and AI Mode may issue several related searches across subtopics to build one response, and that mix is not fixed.
• Personal context. Google says AI Mode personalizes some results using previous conversations and places the user searched for in Search and Maps. Your logged-in test is not your customer's.
• Location. Ask from San Antonio, then from Houston, and a local question gets different local answers. Pick a location for each prompt and keep it.
• Different surfaces. API answers, which most tracking tools collect, can differ a little from the consumer apps people actually open.
• Sources come and go. Gemini does not attach sources to every response, and each engine refreshes its index on its own schedule.

This is also why ranking language does not fit. There is no number one in an AI answer; there is a probability of being named, and share of voice is your estimate of that probability. The wider mechanics of how engines pick names are in how AI chooses businesses.

How many answers before a change means something

Treat each month's mention rate as an estimate with a margin of error. The standard back-of-the-envelope formula for a proportion gives a rough 95% margin of 1.96 times the square root of p times (1 minus p), divided by n, where p is your rate and n the number of answers. At a 25% mention rate the margins work out like this, and the change between two months has to be larger still, because both months carry their own error.

Not every question deserves the same vote

A plain mention rate treats "how does a heat pump work" the same as "who can replace my AC this week in Stone Oak". Only one of those questions arrives holding a credit card. Weighting fixes that. Group the prompt set by intent and report each group on its own line, or give buying questions double weight in the headline figure and say so in the report.

Free share figures from Google and Microsoft

Third-party tools estimate share of voice from the outside, and Google states that no third-party tool has access to its internal ranking or AI systems. Two platforms publish versions of their own, each limited to its own engine, and a third gives the raw exposure behind them.

• Bing Webmaster Tools, AI Performance. Citations in Copilot and Bing's AI summaries, the grounding queries behind them, Citation Share per query and, since June 2026, Intent and Topic views plus period comparison. No clicks and no competitor names.
• Google Merchant Center, AI performance report. For retailers: a brand's share of voice for shopping queries in AI Mode and AI Overviews, for English queries in Australia, Canada, India, New Zealand and the US.
• Search Console, Generative AI performance report. Not a share figure, but impressions by page in AI Overviews and AI Mode, which shows where Google's AI features show you at all.

None of these covers ChatGPT, Perplexity, Gemini or Claude. For those engines someone has to ask the questions and count, whether that is a person with a spreadsheet, a subscription tool or a managed program. The rest of our generative engine optimization guide explains where each engine gets its sources.

What a share-of-voice score leaves out

So a share figure belongs inside a report that shows the answers, not alone on a slide. DataCram's own platform, AICiteBeam, tracks competitor share of voice and the sources each engine cites across ChatGPT, Claude, Gemini, Perplexity and Google's AI Overviews, and alerts us when visibility moves. In client programs that data sits next to the fixes made that month. Programs start at $625 a month after a free GEO audit; GEO pricing explains the range, and AI visibility monitoring covers the tracking side in detail.