AI share of voice is the percentage of a brand's eligible AI answers — across ChatGPT, Claude and Perplexity — in which that brand gets named or recommended, out of every answer where it had a real chance to be. It's the same competitive idea marketers have tracked for decades — how much of the conversation one brand owns versus its rivals — adapted to a channel with no fixed ad inventory and no single leaderboard: a distribution of generated answers you have to sample repeatedly to read correctly. Here's what changed, and how to calculate your own number.
Classic share of voice is the percentage of total market conversation — media mentions, ad impressions, search visibility — that one brand owns versus its competitors, in a defined channel and time window. The formula is simple: your brand's measure divided by the total market's measure for that channel, whether the channel is social mentions, PR coverage, paid-search impression share, or organic keyword visibility. Marketers have used it for decades to answer one question — are we getting louder or quieter, relative to everyone else fighting for the same attention.
Proof: HubSpot — The beginner's guide to share of voice.
Bottom line: the concept isn't new — what's new is the channel it now has to cover.
AI share of voice tracks the same competitive question, but the conversation is a set of AI-generated answers to buyer questions — not a fixed pool of ad slots, headlines or search results pages. There's no single ranking to screenshot: ChatGPT, Claude and Perplexity each decide independently who to name, and the same question can get a different answer the next time you ask it. That's exactly why it's treated as its own dimension of AI brand intelligence rather than something a social-listening tool can bolt on — the signal lives inside model outputs, not public posts.
Why it matters: a share-of-voice number pulled from one chat session tells you almost nothing — the channel itself is probabilistic, so the measurement has to be too.
A trustworthy AI share-of-voice figure holds three things fixed: a defined set of buyer questions, every engine you care about, and more than one run per question — because the same prompt can return a different answer on repeat. Reflexa's own AI Visibility Index publishes exactly this kind of measurement in the open: a fixed set of 100 real buyer questions, asked three times each to ChatGPT, Claude and Perplexity, for 900 observed answers per category edition. In the September 2026 project-management edition, Asana's share of voice worked out to appearing in the top 3 recommendations in 359 of its 882 eligible answers — about 41%.
Proof: Reflexa AI Visibility Index — methodology and worked example.
Do this: never report a single-engine, single-run number as "your" AI share of voice — treat it as one sample from a distribution until you've repeated it.
Your AI share of voice for a category is your brand's top-3-or-named count divided by the number of eligible answers you were asked about — not the total number of questions you ran. A question that already names your brand (like "what are the alternatives to [you]?") creates an artificial opportunity to be mentioned, so it shouldn't count in your own denominator — though it stays fair game for every competitor who wasn't named in the question. Run the same fixed question set across ChatGPT, Claude and Perplexity, log who gets named in each answer, and divide.
Rule of thumb: the denominator matters as much as the numerator — inflate it with brand-anchored questions and every share-of-voice number looks artificially low.
A high share of voice built on wrong facts is a liability, not a win — being named often doesn't mean being named correctly. A brand that appears in most answers but with a wrong price, a discontinued feature, or a competitor's claim attached to its name is arguably worse off than one that's simply invisible, because the wrong version is now the one reaching buyers at scale. Read share of voice next to whether the mentions themselves are accurate before treating a high number as good news on its own.
Good news: the fix for both is the same discipline — clear, consistent, corroborated facts published where the engines can find them.
There's no version of this that gives you one fixed number and leaves it there. Research on how consistently language models answer the exact same prompt on repeat finds answer-consistency rates that vary widely by model and are routinely well short of 100%, even with randomness settings turned down — which is precisely why a single question, asked once, isn't a measurement. Treat AI share of voice as a tracked trend, re-run on a fixed schedule, not a score you check once and file away.
Proof: arXiv — The Non-Determinism of LLMs: Evidence of Low Answer Consistency in Repetition Trials.
Sources: HubSpot — The beginner's guide to share of voice · arXiv — Non-determinism of LLMs, repetition-trial consistency · Reflexa AI Visibility Index
The free check reads it across ChatGPT, Claude and Perplexity — 3 minutes, evidence included.