Named, not just quoted
What is share of voice in AI answers - and how do you measure it honestly?
Every AI visibility vendor shows you a share of voice chart. Most of them won't show you the denominator - which means the chart is decoration, not evidence.
28 Labs · September 2026
Share of voice in AI answers is your brand's naming rate relative to competitors, measured across the identical set of real buyer questions, asked of the same engines, in the same wave. Ask "best X for Y" and "who should I use for Z" across ChatGPT, Claude, Gemini, and Perplexity; count how often each brand gets named, or named first; your share of voice is your count divided by the category total. It's the closest thing this category has to a scoreboard - but only if the question set is frozen and every brand's count re-derives back to a specific answer.
What does share of voice actually count?
Share of voice is a relative metric, not an absolute one. You can't calculate it from your own brand's data alone - you need the same fixed question set asked once, in the same wave, and you need to count every brand the engines name, not just yours. Your share is your naming count divided by the total naming events across the whole category, including you.
Which naming rate you use changes what the number means. Being named anywhere in an answer is a different win than being named in the top three, which is a different win again from being the brand the engine recommends first. Share of voice can be built from any of the three - Visibility Rate, Top 3 Rate, or #1 Pick Rate - as long as you say which one you're reporting. We define all three in the AI visibility metrics glossary. A brand can post a strong share of voice on Visibility Rate and a weak one on #1 Pick Rate at the same time - both numbers are true, and they tell different stories about whether the brand is present or actually chosen.
Share of voice isn't citation share
In June 2026, Microsoft added a metric called Citation Share to Bing Webmaster Tools: your share of all citations for a given grounding query, inside Copilot and Bing's AI answer summaries. It's the first first-party citation-share number any platform has published, and it's easy to mistake for the same thing as share of voice. It isn't.
Citation Share
- Measures how often an engine reads and quotes your page while building an answer
- Counted per grounding query - the engine's internal search, not the user's actual question
- Covers Microsoft surfaces only: Copilot and Bing's AI summaries
- Says nothing about whether the finished answer named or recommended you
Share of voice
- Measures how often the finished answer names your brand, or names it first
- Counted against the real questions a buyer would actually ask
- Requires you to ask the engines yourself - ChatGPT, Claude, Gemini, Perplexity
- Tracks whether AI is actually recommending you, not just reading you
A brand can dominate citations for a query and still lose the recommendation - the engine reads the comparison page for facts, then names a competitor as the pick. The reverse happens too: a brand can get named with almost no citations, because the recommendation came from the model's training data or another source entirely. Neither number substitutes for the other.
We cover the citation side in full in what the Bing AI Performance report can and can't tell you. Read both pieces together if you're trying to understand your full AI footprint - being cited and being recommended are two separate jobs.
What honest measurement requires
Share of voice is easy to fake and easy to trust by accident. To make it mean something, four things have to hold at once:
- A frozen question set. Change the questions between reads and you've changed the metric, not the market. A question set that shifts every time someone runs the report isn't measuring movement - it's measuring the new questions.
- A schedule. AI answers are volatile. The same question asked twice in one week can return different names. One read is a snapshot; a share of voice trend needs waves run on a fixed cadence, so movement between waves is real and not noise.
- Per-engine breakdowns. Share of voice differs radically by engine, because ChatGPT, Gemini, Claude, and Perplexity retrieve and rank differently. A brand can lead its category on one engine and barely register on another. Blend them into one number and you erase the difference that actually matters for where you invest.
- Raw, re-derivable counts. "Named in 14 of 60 answers" is a measurement. "34% share of voice" with no stated question set, engine list, or wave date is a claim wearing a percentage sign.
Where the metric gets faked
Three failure modes show up constantly in this category, and all three produce a number that looks like share of voice while measuring almost nothing:
- Cherry-picked prompts. Run twenty questions, keep the eight that flatter the brand paying for the report, call the average "share of voice." The number is real math on a fake sample.
- Hand-tuned brand-name questions. Asking "tell me about BrandX" and counting the answer as a win mistakes presence for recommendation. The engine will discuss any brand you name directly - that's not the same signal as a buyer asking "who should I use for X" and the engine naming you first, unprompted.
- Rolling composite scores. A single number blended from unknown weights, refreshed on no fixed schedule, that moves without anyone being able to say which question or which answer caused the move.
The composite score warningIf a share of voice number can't be traced back to a specific frozen question set, a specific wave date, and per-engine counts you could re-ask yourself, don't trust the trend line. A composite score with a hidden denominator isn't a softer version of measurement - it's a different thing wearing the same label, and it will tell you the story whoever built it wanted you to hear.
The real use of share of voice: the gap list
The number itself is a scoreboard. The useful part is what's underneath it - the specific questions where a competitor gets named and you don't. That gap, question by question, is the actual work list. Not "improve AI visibility" as a goal, but "we lose the 'best X for Y in [market]' question to two named competitors, every wave, on every engine except Perplexity" as a fact you can act on.
Read that way, share of voice stops being a vanity metric and becomes a diagnostic. It tells you exactly where to build the content, the structured data, or the third-party citations that close a specific, named gap - not a general instruction to "do more AEO."
What we do about it
We run a frozen question set of real buyer questions across ChatGPT, Claude, Gemini, and Perplexity, on a fixed wave schedule, and report share of voice per engine, broken out by Visibility Rate, Top 3 Rate, and #1 Pick Rate separately. Every number re-derives to the specific question and the specific dated answer behind it. No blended score, no hand-picked prompts, no composite nobody can check. That's what makes the gap list real - and the gap list is the part that turns into work.