AI share of voice is the percentage of AI answers, when buyers ask ChatGPT, Claude, Gemini or Perplexity about your category, where the assistant names or recommends your brand. The AI share of voice formula is simple: your brand’s mentions divided by all brand mentions in your category, times 100. The formula is the easy part.
What makes an AI share of voice number worth reporting is the sampling around it: a fixed set of buyer prompts, run several times on each assistant, with recommendations counted separately from passing mentions, because one answer is basically a coin flip.
It’s also the number an AI search visibility report leads with, because it’s relative: your brand against the brands the assistant names, week after week. A rank tells you where you sit in Google, and AI share of voice tells you who’s winning the category.
The share of voice formula
The share of voice formula is your brand’s mentions divided by every brand’s mentions in the category, times 100. It’s an old advertising ratio: share of voice originally meant your share of the category’s paid ad impressions, and Palo’s glossary still works the example of 500 impressions in a market of 1,000 as 50%. Social listening later carried the ratio across by swapping impressions for mentions, and Sprout Social states it as your brand metrics divided by total market metrics.
Let’s say the answers to your prompt set name brands 150 times between them, and your brand accounts for 60 of those mentions. Your share of voice is 60 ÷ 150, so 40%. Whether that’s good depends on how many brands are sharing the rest: against five competitors it’s a strong hold on the conversation, against one dominant incumbent it’s a fight.
Where usage varies is the denominator: the inputs change by field, and the ratio never does. Some teams divide by all the mentions in their category, some by all mentions in the whole market, and some count conversations instead of single mentions. Pick one, write it down and hold it fixed, because a share of voice trend is only a trend if the denominator didn’t move underneath it.
The AI share of voice formula
AI share of voice is the same ratio with AI answers in it: your share of the brand mentions that ChatGPT, Claude, Gemini and Perplexity give back when a buyer asks about your category. The numerator is the answers where the assistant names you. The denominator is all the brand mentions across your prompt set.
Two denominators are both defensible, and you should pick one before you start. Share of answers divides by the prompts you ran: ExposureNinja’s method computes AI responses mentioning your brand over the total prompts scanned, so 10 mentions across 20 prompts reads as 50%. Share of mentions divides by everything the assistant named, and switching between the two partway through a trend is the one move that breaks the number, because they aren’t comparable.
One distinction the formula can’t hold on its own is between being named and being put forward as a pick. ExposureNinja’s method tells you to log direct brand mentions and product recommendations separately, and then its formula has one numerator, AI responses mentioning your brand, so the split it told you to log doesn’t make it into the number. A brand named in passing doesn’t get the buyer, a brand put forward does, so count them separately and report two shares if you have the patience: one for being named, one for being recommended.
How to measure AI share of voice
You measure share of voice in AI search by fixing a set of buyer prompts, running each one several times on each assistant, and counting what comes back. It’s a routine on a schedule, and the number is a trend before it’s a level. Each step has a place where it goes wrong in practice.
-
Build a fixed prompt set from how buyers actually ask: comparisons, “best of” questions and category questions, in the words a buyer types rather than your keywords. Twenty to fifty prompts is enough to start. When the market’s questions change, add the new prompt and keep the old one running a few weeks before you retire it, so the trend line doesn’t break.
-
Run every prompt several times per assistant. The same prompt can name you on one run and skip you on the next, so a single-run number is mostly noise. Five runs per prompt per assistant is a workable floor, which comes to 600 answers if you track 30 prompts across the four assistants.
-
Record each brand the answer named, and whether the assistant put it forward or named it in passing. Log the position too: Semrush’s method for share of voice in AI chatbots weighs position as well as frequency, on the logic that the first name in an answer carries more than the last.
-
Compute the share, and treat small samples as noise. On 600 answers, a move from 18% to 21% is 18 answers, which is signal. A move from 18% to 18.5% is three answers, which is noise wearing a trend line.
-
Track the whole category alongside your own share. When your number drops and every named competitor drops with you, a model update moved the category and your brand didn’t get worse. Keep a few buyer questions untouched as a comparison set, so when you start working the others you can show the lift was yours.
That’s the whole method, and it doubles as the honest way to measure search visibility: a share of the answers buyers actually receive, taken over weeks, rather than a single position snapshot.
What a good share of voice percentage is
There’s no universal good share of voice percentage, because the number only means something against your own category. If the assistant names two brands for your category, 50% is a dead heat. If it names twelve, 20% is a commanding lead.
A 50% share of voice means your brand took half of whatever you fixed as the denominator: half the answers that named any brand, or half of all mentions. It usually means the category’s answers are concentrated on you, which is a nice problem. Check it against your recommended share before you celebrate, because half of all mentions and few of the recommendations is a losing 50%.
What the number can’t tell you is everything around it: whether the recommendation came with a caveat, whether the buyer acted, what it did to revenue. It’s a share of being named and put forward. Anyone quoting an industry benchmark for share of voice is inventing it, because the denominator is your category rather than an industry.
Visibility score vs. share of voice
A good SEO visibility score tells you where you stand in Google, and it’s built from rank positions. Sistrix’s visibility index, for example, combines the keywords a site ranks for with their search volume and the click rate of each position. So a good visibility score is one that’s rising against your own history, and it still says nothing about which brand an assistant picks.
| SEO visibility score | AI share of voice | |
|---|---|---|
| Built from | Google rank positions, search volumes, click rates | Prompts, repeated runs, the brands each answer names |
| Measures | Your standing in Google’s results | Your share of the answers where buyers ask about your category |
| Moves when | Your Google rankings move | The assistants’ answers change |
| Can’t tell you | Whether any assistant cites or recommends you | Where you rank in Google |
The two numbers are separate, and the separation is measurable. In our 10x AI Visibility Refresh study, refreshing 98 posts moved AI citations up 10x in 91 days with nothing new published, while the relationship between Google rank and AI citation was basically zero (Spearman rho 0.017, p = 0.914), and 41 cited pages had no Google impressions at all. The citation lift came while rank stayed flat.
Neither number predicts the other. If your visibility score climbs while your share of voice sits still, the work is landing somewhere the assistants aren’t reading.
Measuring this in practice
Tracking share of voice in AI search by hand gets old within a week, and this is the job Omnipresence does. It tracks whether ChatGPT, Claude, Gemini and Perplexity recommend you or a named competitor for the questions your buyers ask, with every question asked many times so the share you report doesn’t swing on a single answer. It keeps a few questions untouched, so a lift in the worked ones is provably your work, and it shows share of voice against your named competitors by day.
The limits are the honest ones. It tracks those four assistants and nothing else, so no AI Overviews or AI Mode, and you bring your own model key and pay the providers at list price. Its data on which brand an assistant put first is partial, so read the recommendation share as the headline and the first-place detail as rough.
Fix the prompts, repeat the runs, keep recommendations separate from mentions and hold the denominator still. Do that for a quarter and you’ll have a share of voice number that holds up when someone challenges it, and someone will.