How to Measure AI Share of Voice: Formula, Method and Pitfalls

When SparkToro and Gumshoe.ai had 600 volunteers ask ChatGPT, Claude and Google's AI the same brand-recommendation questions nearly 3,000 times, the odds of getting the same list of brands twice were under 1 in 100. That single finding decides how AI share of voice has to be measured. One prompt typed once into ChatGPT tells you almost nothing. A defined prompt set, run repeatedly on each engine and counted the same way every time, gives you a number stable enough to compare against competitors and to track month after month.

What AI share of voice measures

AI share of voice is the percentage of brand mentions your company earns in AI-generated answers, out of all brand mentions in your category, across a defined set of prompts and engines. It is the GEO counterpart of the classic share of voice that marketers used for advertising and press coverage, with one structural difference: an AI engine does not show ten results and let the buyer choose. It writes one answer that names two, three or five brands, and a brand left out of that answer does not exist for that buyer at that moment.

AI share of voice is the share of all brand mentions in AI answers about your category that name your brand, measured across a fixed prompt set, on each engine separately.

The formula is short: AI share of voice = (your brand mentions ÷ total brand mentions for all tracked brands) × 100. If 20 prompts on ChatGPT produce 48 brand mentions in total and 9 of them are yours, your share of voice on ChatGPT for that prompt set is 18.75%. Everything difficult sits in the inputs: which prompts, which engines, how many runs, and which competitors count. The rest of this guide covers those choices.

Why the number matters more than your rankings now

Rankings no longer tell you how much of the buyer's attention you get. In Pew Research Center's analysis of 68,879 Google searches, users clicked a traditional result in 8% of visits when an AI summary appeared, against 15% when none did, and they clicked a link inside the summary itself in just 1% of visits.

Being named inside the answer is what protects you. Seer Interactive's study of 3,119 informational queries found organic click-through on queries with an AI Overview fell from 1.76% to 0.61% between mid-2024 and September 2025, while brands cited in the AI Overview earned 35% more organic clicks than brands that were not. That is why GEO and SEO now need separate scorecards.

Step 1: build a prompt set that sounds like your buyers

Start with 25 to 50 prompts written the way buyers actually ask questions, not the way you would type keywords. "Which CRM should a 50-person sales team use?" is a useful test prompt. "CRM software tools" is not. Cover four kinds of prompts: category questions ("best payroll software for a small business"), problem questions ("how do I stop losing leads between marketing and sales"), comparison questions ("alternatives to [competitor]") and a few brand questions ("what does [your brand] do").

Keep brand names out of the category and problem prompts. Those prompts measure whether AI engines think of you unprompted, which is the number that matters for new buyers. Brand prompts measure something else, how accurately engines describe you, and mixing the two inflates your score. The gap between both is worth tracking on its own, as covered in our analysis of branded vs non-branded AI share of voice.

Step 2: run every prompt several times on each engine

Because AI answers are generated from probabilities, the same prompt returns a different set of brands from one run to the next. The SparkToro and Gumshoe data put the odds of the same list in the same order closer to 1 in 1,000. Their study also found something more useful: across hundreds of differently worded prompts with the same intent, the engines kept returning a similar group of leading brands. How often a brand appears is stable enough to measure. Its exact rank in any single answer is not.

In practice, run each prompt at least three to five times per engine, in fresh sessions with no prior conversation, and work with the mention rate rather than a yes or no for a single answer. We explain why a single check is an anecdote rather than data in our piece on measuring AI visibility as a distribution.

Step 3: count mentions, and keep position as a secondary signal

For every answer, record each brand mentioned, the order in which it appears, and whether it is recommended, mentioned neutrally or cautioned against. Then calculate share of voice per engine with the formula above.

Here is how it plays out for a mid-market CRM vendor running 40 prompts five times on three engines, which gives 600 answers to analyse. Its brand appears 72 times, its two main competitors 84 and 54 times, for 210 brand mentions in total. Its share of voice is 72 ÷ 210 × 100 = 34.3%. Filtering the same data down to the prompts about affordable CRMs drops its share to 12%, which tells the team exactly where it is losing buyers.

Position still carries information, since a brand named first is read before one named fifth. But given how much order varies between runs, use it as an average over many answers, never as a ranking for one prompt. A share of voice built on mention counts is the number to report; position is context for it.

Step 4: calculate per engine, never as one blended score

ChatGPT, Gemini, Perplexity, Copilot, Google AI Overviews and Google AI Mode retrieve and weight sources differently, so the same brand can score 35% on one engine and 8% on another. Averaging those into one figure hides the most useful signal you have: where you are strong and where you are invisible. Our comparison of how AI engines rank the same brands differently shows how wide those gaps get.

Step 5: record sentiment and sources, not only mentions

A high share of voice with lukewarm wording is weaker than it looks. A brand described as "a cheaper option with limited support" in 35% of answers has a visibility number that hides a positioning problem. Track sentiment next to the count, as we detail in our guide to monitoring AI brand sentiment.

Record the sources each engine cites as well. They explain the score. In Ahrefs' study of 75,000 brands, branded web mentions showed the strongest correlation with visibility in AI Overviews (0.664), brands in the top quartile for web mentions averaged over ten times more AI Overview mentions than the next quartile, and 26% of brands had no AI Overview mentions at all. The cited pages where competitors appear and you do not are your citation gap, and usually the clearest to-do list the measurement produces.

How to read the result

A share of voice has no meaning in isolation. Read it against the competitors measured on the same prompt set and the same engines, because a 15% share can lead a crowded category and trail in a niche one. Our competitive AI share of voice analysis walks through that comparison, and the AI share of voice benchmarks give reference points by category.

Then read the trend. One measurement is a snapshot; the same prompt set run every week or month shows whether content, PR and technical work are moving the number. When it is time to act on the gaps, the methods are covered in how to improve AI share of voice.

The measurement mistakes that distort the score

Testing a single engine is the most common one. ChatGPT is the engine teams know, so it is often the only one tested, while Gemini and the Google AI surfaces can tell a completely different story. Changing the prompt set between two measurements is the second: the comparison becomes meaningless, so freeze the core prompts and add new ones as a separate group.

The third is reading a mention as a recommendation. Appearing eighth in a list of tools "some users also consider" is not the same as being the answer. The last one is stopping at visibility. Share of voice says whether AI engines name you; AI referral traffic in Google Analytics 4 says whether those mentions bring visitors, and the two numbers together tell you whether the work is paying off.