AI Share of Voice

AI share of voice is the percentage of relevant prompts for which an AI answer engine mentions or cites your brand — brand appearances divided by total prompts sampled, per engine, per period. If ChatGPT surfaces your brand on 6 of 25 tracked prompts, your share of voice is 24%.

It is the metric the AI-visibility tool vendors sell, which is exactly why the definition should come before any dashboard. The metric is a sampling exercise anyone can run with a frozen prompt battery and a spreadsheet — no paid tracker appears anywhere in our own stack [our data].

How is AI share of voice calculated?

Divide the prompts where your brand appears by the total prompts in a fixed battery, per engine, per month. The battery — 20–30 questions your buyers actually ask, frozen so the wording never drifts — is the denominator; appearances in the engine's answers are the numerator.

An illustrative worked example (the arithmetic, not a benchmark):

What goes into the denominator matters as much as its size. A battery built only from brand questions flatters the score — engines mention brands most reliably when asked about them directly — so an honest mix leans on category, comparison, and money questions, where your brand has to earn its way into the answer. Draw the battery from Search Console queries and real buyer questions rather than invented prompts — the same first-party demand sources our own query mining runs on [our data].

Freezing the battery is the method, not a detail. If the prompts change month to month, the trend line measures your prompt-writing instead of your visibility. New prompts go in a separate, dated section; the step-by-step protocol is in how to measure AI share of voice without paid tools.

What counts as cited versus mentioned?

Cited means the engine linked your page — a path exists for the reader to reach you. Mentioned means your brand was named with no link — reputation without a click path. They are 2 separate rates, they move independently, and a single blended percentage hides which one moved.

The mention rate matters more than intuition suggests, because citations rarely convert to clicks: Pew found users clicked a cited source in about 1% of visits that included an AI summary (March 2025 data, Pew Research Center). Most of what an answer engine says about a brand never registers in analytics — which is why a referral report, like our GA4 AI-traffic channel, can only ever see a fraction of the visibility that share of voice samples directly.

Why is AI share of voice a sample, not a census?

Because answer engines are non-deterministic: the same prompt can return different answers, different sources, and different brands across runs on the same day. Any single reading is one draw from a distribution, so the metric is only readable as a trend — the same battery, scored the same way, across months.

Part of what the sample draws from is also frozen. Kevin Indig's 2026 analysis found 24% of ChatGPT answers are generated without fetching any live page (Growth Memo, 2026) — answers reconstructed from training data, which no recent publishing reaches. A share-of-voice loss on those prompts is real, but it is not a loss this month's work caused or can quickly reverse.

These limits apply to every measurement of the metric, paid or manual. A daily-sampling tool draws from the same distribution a monthly spreadsheet does; frequency smooths the trend line and changes nothing about what is underneath it.

Do you need a paid tool to measure AI share of voice?

Not to start, and for a single brand usually not at all. A 20–30 prompt battery run monthly across 4–5 engines costs an afternoon and produces the same wins-and-losses signal a dashboard renders — the full free protocol is documented in our manual measurement method, and what the paid trackers charge for automating it is priced out in our AI visibility tools breakdown.

Against our own interest as a team that sells measurement-heavy builds: the common failure is buying the dashboard before defining the metric. If you cannot yet name the 25 prompts that matter for your business, a tool will happily sample prompts that do not, at $100+ a month, and the percentage it reports will be precise noise. Define the battery first; automate when the afternoon stops being enough.

What does AI share of voice not tell you?

It does not tell you clicks, revenue, or why an answer changed. Share of voice measures presence in answers — the thing analytics cannot see — and nothing downstream of it; connecting presence to business value takes referral tracking and honest attribution, not a bigger percentage. It also cannot distinguish a durable loss from sampling flicker until 2–3 monthly readings agree, which is the price of measuring a non-deterministic system. And it says nothing about accuracy: an engine can mention your brand on every prompt and misstate your prices in half of the answers, which is why the scoring rubric keeps a separate "wrong" column for every run.

What it is good for is the thing nothing else does: telling you whether answer engines know you exist on the questions that matter, and whether the pages you publish are moving that needle. The publishing side of that loop — building passages worth retrieving in the first place — is the subject of the operator's guide to generative engine optimization.