AI share of voice: what it is, how to calculate it, and how to benchmark against competitors

Share of voice is a measure of participation: how much of a category’s conversation your brand holds compared with the others. What changes from one channel to the next is what counts as “the conversation.”

• In advertising: your spend or impressions over the category total.
• In SEO: your visibility across a set of keywords, weighted by position.
• In media and social: your mentions over the category’s total mentions.
• In AI: your mentions over all brand mentions in the answers assistants give to questions about your category.

The difference from SEO is one of format. A classic search engine shows ten results, and you can be the seventh. An AI assistant answers with a paragraph that names three, four or five brands, sometimes just one, and the person asking rarely keeps searching. In each answer your brand is either there or it isn’t. That is why AI share of voice works more like media share of voice than like SEO: you count who gets named, and where in the list.

Almost every measurement of a brand in AI ends in two different numbers, and half the misunderstandings come from confusing them.

• Incidence, or visibility rate: how many answers your brand appears in out of all valid answers. It is expressed as “k of n,” for example 45 of 90. It answers the question “how often does AI name me?”
• Share of voice: how much of all brand mentions goes to you. It answers “how much of what AI says about my category is me?”

The two can move in opposite directions. If engines start naming fewer brands per answer, your share of voice can rise even though you appear in fewer answers. And an “AI visibility index” that sums everything up in one score usually blends these two numbers with position and sentiment. It works as a headline, but to decide what to do you have to look at each component on its own.

The arithmetic is simple. Whether the number is useful depends on four decisions you make before doing it:

1. What counts as a mention. The brand name and its variants, plus its products if the question is about products. The cleanest approach is one mention per brand per answer, not one for every time the name repeats within the same text.
2. Which brands go into the denominator. Every brand that appears in the answers, not just the competitors you declared. If you leave out the brands AI names that weren’t on your radar, your percentage comes out inflated.
3. Which questions count. Only the ones that don’t name your brand. “Is Yerbal Norte yerba mate any good?” will always name you; mixing it with “what’s the best mild yerba mate in Uruguay?” pushes the number up without saying anything about your real position.
4. Which answers count. Only the ones the engine actually gave. This one deserves its own section.

Not every query produces an answer. Google doesn’t generate an AI Overview for every search, only when it decides one adds something. An API can return an error or empty text, and an engine can decline to answer. If those queries go into the denominator as “answers that didn’t name you,” your number drops for a reason that has nothing to do with your brand.

In the example further down, Google answered 20 of 40 queries and the brand appears in 14 of those 20: 70%. Divide by 40 instead and you get 35%, and the report would say the brand does badly on Google when Google actually names it in almost three of every four answers it gives. The rule is not measured is not zero: a query with no answer leaves the calculation and is reported separately, as the engine’s coverage.

For the same reason, every percentage has to travel with its sample. A bare “70%” doesn’t distinguish 14 of 20 from 700 of 1,000, and those are claims of very different sizes.

ChatGPT, Claude, Gemini, Meta AI and Google’s AI Overviews don’t answer the same question the same way. Each has its own model and its own way of using the web, so share of voice is calculated engine by engine. An average of all five hides exactly what matters: where you are strong and where you don’t show up.

The most useful split is between two families of engines:

• Chat engines (ChatGPT, Claude, Gemini, and Meta AI, the assistant inside WhatsApp, Instagram and Facebook) answer largely from what the model learned. If they name you, it’s because your brand became associated with the category in the material they were trained on.
• AI search (Google’s AI Overviews) builds the answer from the pages Google ranks today for that query and paraphrases them. If it names you, it’s usually because a page of yours, or one that talks about you, ranks well for that question.

That’s why a young brand can have a good share of voice in AI search and almost none in chat: today’s web knows it, the model doesn’t yet. It also works the other way: a brand with many years behind it can dominate in chat and lose in AI search to competitors with better recent content. Averaging the two families gives a number that describes neither, and each one is improved with different actions.

A screenshot of ChatGPT naming your competitor is not a benchmark. It’s one answer, from one engine, on one day, with one way of asking. For the comparison to be fair, it has to meet these conditions:

1. 01

   The same questions for every brand

   One set of category, comparison and purchase-intent questions, and every brand is measured on that set. If each brand gets its own questions, you are comparing questions, not brands.

2. 02

   The same country and the same language

   Answers change from one market to another. The question has to be written in the market’s language and name the country in full (“in Uruguay,” not “UY”), and AI search has to be queried with that country’s location.

3. 03

   The same period

   Every brand in the same run. Comparing your measurement from this month with a competitor’s from three months ago mixes the brand effect with the time effect: models get updated and the web changes.

4. 04

   The same engines and the same number of runs

   If you measured your brand on five engines and the competitor on two, or yourself with three runs per question and them with one, the comparison doesn’t hold.

5. 05

   The full competitive set

   Declare your direct competitors, but also count the brands AI names that weren’t on your list. They tend to be the biggest surprise.

6. 06

   By topic, not just in total

   A category holds several conversations: price, trust, service, speed. You can lead the total and disappear when the question is about price, and the average won’t show it.

Category: yerba mate in Uruguay. The brands and numbers are fictional, built to show the calculation. We asked 40 questions that name no brand (“what’s the best mild yerba mate in Uruguay?”, “which yerba mate is best for drinking every day in Uruguay?”) on all five engines: 200 queries in total. First, Yerbal Norte’s incidence on each engine:

The right-hand column is the most common mistake. With non-answers in the denominator, Google looks like Yerbal Norte’s worst engine (35%); counted correctly, it’s the best (70%). Splitting the two families, the brand appears in 76 of 160 chat answers (47.5%) and in 14 of 20 AI search answers (70%): today’s web ranks it better than the models remember it. Now, share of voice against the competition:

Yerbal Norte leads with 30%. Had only the four declared brands been counted, the number would have been 33.3% (90 of 270): the 30 mentions of brands nobody had on the list change the picture. And with four brands competing head to head, parity is 25%, so sitting at 30% means being somewhat above your proportional share, not dominating the category.

• Measuring with one question or one screenshot. Answers vary from one run to the next; one answer is not a sample.
• Counting non-answers as zeros. It penalizes the engines that answer less often, for reasons unrelated to your brand.
• Mixing in questions that name your brand. They inflate share of voice, because the answer names you by definition.
• Leaving out undeclared brands. It inflates your percentage and hides real competitors.
• Averaging chat and AI search. They are different mechanisms, and they improve with different actions.
• Comparing measurements from different periods or countries. The model or the market may have changed, not your brand.
• Reporting percentages without a sample. A “60%” with no “k of n” can’t be questioned or reproduced.

The Share of Voice module, Refractia’s competitive arena, calculates your share of AI mentions against your competitors, per engine, with rankings for organic citations, discovery rate and visibility over a 30-day window. Every rate is shown with its sample, queries with no answer don’t count as absence, and chat and AI search are reported separately. Questions are generated in the market’s language and name the country in full, and Google’s AI Overviews are fetched with each country’s location.

Performance by Topic splits the category into topics and measures, in each one, incidence, position against each declared rival, and the split between chat and AI search. Evolution shows the share of voice, visibility, sentiment and position lines for your brand and for each competitor from one scan to the next.

The limits, stated plainly: we measure ChatGPT, Claude, Gemini, Google’s AI Overviews and Meta AI; we don’t measure Perplexity, Copilot, Grok, or Google’s AI Mode as a separate surface. And every scan is started by a person, from your team or ours. It is a deep diagnostic, not a daily tracker.