How to calculate AI share of voice (and the denominator mistakes to avoid)

AI share of voice is the client's share of all brand mentions in AI answers, among the brands you track:

Measured over a fixed set of prompts, on the same answers, for a defined period. The formula is simple. What makes the number trustworthy — or meaningless — is how you count.

A worked example

Say you track one client and three competitors on 20 prompts, asked weekly on three engines: 240 answers in a month. You count, for each brand, the number of answers that mention it:

(Fictional data, the same client as our sample report.)

Note that the total, 270, is larger than the 240 answers: one answer can mention several brands. That's expected.

Share of voice vs mention rate

These two get confused constantly, and they answer different questions:

• Mention rate — answers mentioning the client ÷ all answers. In the example, 71 ÷ 240 = 30%. "How often does the client show up?"
• Share of voice — client mentions ÷ all tracked-brand mentions. 26.3%. "When AI recommends someone in this category, how much of that attention goes to the client?"

Mention rate can rise while share of voice falls, if competitors rise faster. Report both: mention rate for presence, share of voice for competitive position. More on the full metric set in how to measure AI visibility.

Rule 1: Count a brand at most once per answer

If an answer names the client three times and a competitor once, is that 75% share of voice? It shouldn't be. Repetition within an answer usually reflects writing style, not preference — a list item with a follow-up sentence names the same brand twice.

Counting once per answer makes the number stable and explainable: "of the answers that mentioned any tracked brand, here's how they split." Some tools count every occurrence. Neither is wrong, but you must know which you're reading, and you can't compare numbers counted differently.

Rule 2: Fix the denominator before you start

The denominator is the brands you chose to track. Change the competitor set and every share changes, even if nothing changed in the AI answers. So:

• Pick 3–5 competitors and keep them. If you add one, mark the month and don't compare across the change.
• Decide on aliases up front. If "Cyclo" and "Cyclo Co." are counted separately, Cyclo's share is split in two.
• Exclude the noise you can't act on. Marketplaces and media brands can be tracked as competitors if clients really lose business to them; otherwise they make shares smaller without making them more useful.

Rule 3: Same prompts, same engines, same period

Share of voice measured on "best X" prompts will look very different from share of voice on "X vs Y" prompts. Compare months only when the prompt set is the same. Report per engine as well as overall — a brand can own Perplexity and be absent from ChatGPT, and the average would hide both facts.

Rule 4: Report the counts, not just the percentage

"26%" invites questions. "Mentioned in 71 of 240 answers; 270 brand mentions in total" answers them. Showing the counts also stops anyone over-reading a small change: on 240 answers, a two-point swing in share of voice is a handful of answers.

Optional: weighted share of voice

Some teams weight mentions by position — first mention counts more than fifth. It's defensible in theory, but position is the least stable thing in an AI answer: the SparkToro/Gumshoe study found identical ordering across runs roughly once in a thousand. If you weight by position, report the unweighted figure alongside it, and explain the weights to the client.

How to get the numbers without a spreadsheet

You can calculate share of voice by hand from a log of answers (one row per answer, one column per brand). For an agency running this every month for many clients, it's what a tracking tool should do. Ranxii counts each tracked brand at most once per answer, calculates share of voice across the client and up to five competitors, shows it per engine, and links each number to the answers it came from — so "why is Veloz at 36%?" has an answer you can open.

Sources

• SparkToro: AIs are highly inconsistent when recommending brands or products (January 2026)