AI Share of Voice: how to actually calculate it

Share of Voice is the only AI visibility metric that tells you whether you're winning. Mention rate tells you that you exist. Share of Voice tells you how much of the answer space you own relative to everyone else competing for the same buyer. And most of the Share of Voice numbers people quote in board decks are calculated wrong, usually in the direction that flatters them.

What AI Share of Voice actually measures

AI Share of Voice is your brand's share of all brand mentions produced by AI engines across a defined prompt set, in a defined period.

The definition has three load-bearing parts, and dropping any one makes the number meaningless:

• Defined prompt set. SoV is always relative to a set of questions. "Our SoV is 31%" is an incomplete sentence. It's 31% on those 25 prompts. Change the prompts and the number changes, which is exactly why you version the set and don't quietly edit it.
• Defined engines. ChatGPT, Perplexity, Gemini and Google AI Overviews produce different answers with different numbers of brands. A blended number without stated weights hides more than it reveals.
• Defined period. Engines change. Answers drift. A number without a date on it is a fossil.

Say all three out loud every time you report it: "31% Share of Voice, on our 25-prompt buyer set, across four engines, week of March 23."

The formula

$$\text{SoV}{\text{brand}} = \frac{A{\text{brand}}}{\sum_{i} A_{i}}$$

Where A is appearances: the number of prompt-runs in which a brand appeared at least once. Not name counts. Appearances.

Unpacked into something you can put in a spreadsheet:

1. For each prompt, for each run, for each engine, record which brands appeared. One appearance per brand per run, no matter how many times the name shows up in the text.
2. Sum your brand's appearances: that's your numerator.
3. Sum appearances for every brand, including yours and including brands you don't consider competitors: that's your denominator.
4. Divide. Multiply by 100.

The denominator is where most people cheat, usually without meaning to. If you only count appearances for you and your three named competitors, you've computed Share of Voice within a competitive set you invented. That's a legitimate metric, and it has a name: relative Share of Voice. It just isn't the same number, and it will be substantially higher. Label it accordingly.

A worked example, with real arithmetic

Ten prompts, one engine, three runs each. That's 30 prompt-runs. Here's what the scan returned:

Your mention rate is 12 / 30 = 40%. You appear in 40% of prompt-runs.

Your Share of Voice is 12 / 80 = 15%. You hold 15% of the answer space.

Those two numbers feel very different, and both are correct. The gap between them is the story: engines are naming an average of 80 / 30 = 2.67 brands per answer, so simply being present isn't enough. Bolt is at 21 / 80 = 26.3% SoV, and Cinder at 18 / 80 = 22.5%. You're third, and not close.

Now the trap. Compute relative SoV against only your three battlecard competitors (Bolt, Cinder, Drift). Denominator becomes 12 + 21 + 18 + 9 = 60. Your SoV: 12 / 60 = 20%. Same data, five points higher, purely from redrawing the boundary. Neither number is a lie. Presenting the 20% without saying which brands are in the denominator is.

Next month you improve. Your appearances go from 12 to 16. Mention rate: 16 / 30 = 53%, up 13 points, great. But the engines also got chattier: total appearances rose from 80 to 96. Your SoV: 16 / 96 = 16.7%, up only 1.7 points. You got more visible and barely gained ground, because everyone else did too. Mention rate alone would have had you announcing a win. This is the exact scenario Share of Voice exists to catch.

Four ways people get Share of Voice wrong

1. Counting raw name occurrences

The most common error, and the most inflationary. If an answer mentions "Acme" four times in a paragraph and each competitor once, raw counting gives Acme 4/8 = 50% SoV from a single answer where all four brands were, in fact, recommended equally. You've measured how wordy the model was about you, not how visible you are. Count presence per run. Always.

2. Cherry-picking the prompt set

You add three prompts about your strongest niche use case, SoV jumps six points, and everyone congratulates the content team. The instrument moved, not the world. Lock the set, version it, and if you must expand it, report the old set and the new set side by side for at least one cycle so the discontinuity is visible.

3. Blending engines without weights

Perplexity answers tend to list more brands than ChatGPT answers, because Perplexity is a retrieval-first product that surfaces multiple cited sources. That mechanically depresses per-brand SoV on Perplexity. Averaging your four engines into one figure silently down-weights whichever engine you happen to dominate. Report per engine. If you must blend, weight by where your buyers actually are; for most B2B SaaS, ChatGPT dominates that mix, given 900M weekly active users versus Perplexity's 780M queries per month as of its last official figure.

4. Ignoring position

Share of Voice treats "first brand named" and "eighth brand named, with a caveat" as identical appearances. They aren't. A brand named first in a list of three is doing something very different from a brand named last in a list of ten. SoV is the headline; position is the qualifier that keeps it honest. Track average position alongside it, and if a competitor is at a similar SoV but consistently named first, you're losing to them despite the tie.

Weighted Share of Voice, for when the basic version isn't enough

Once the basic number is stable, two refinements earn their complexity.

Position weighting. Give each appearance a weight based on rank in the answer: first mention 1.0, second 0.7, third 0.5, fourth-and-beyond 0.3. Sum weighted appearances instead of raw ones. This distinguishes the brand the model actually recommends from the brand it lists as an also-ran.

Sentiment weighting. An appearance in "Acme is the best choice for teams under 50" is not an appearance in "Acme is often cited but users report reliability issues." Score each mention positive, neutral, or negative and either report negative-mention rate separately or subtract it. Negative mentions in AI answers are corrosive precisely because they're delivered as neutral-sounding fact.

Don't start with either of these. Get an honest unweighted number first, run it for two months, and only add weighting when the basic metric stops answering your questions.

Why raw mention count misleads

Here's the failure in one line: mention count grows when engines get more verbose, and Share of Voice doesn't. That's the entire argument for the metric.

Over the last two years the answer surfaces have expanded. AI Overviews now appear on roughly 48% of tracked queries, up 58% year over year, and on 82% of B2B technology queries. More answers, longer answers, more brands per answer. A brand tracking raw mention counts through that period would show a beautiful upward line and could still have been losing share the entire time. The rising tide lifted the count and left the share flat.

The same distortion hits in the other direction. When an engine tightens its answers and starts naming three brands instead of six, everyone's mention count drops and the marketing team panics about a penalty that doesn't exist. Share of Voice would have stayed flat and told them nothing was wrong.

Absolute counts measure the engine. Share measures you.

Where the number comes from

Whatever you use to produce it, the pipeline needs four properties: a fixed prompt set, clean sessions (no memory, no personalization), repeated runs per prompt, and stored raw responses so the denominator can be recomputed when you change your competitive set. Without stored responses you can never recalculate history, and the first time someone asks "what would our SoV be if we excluded the marketplaces?" you'll have nothing.

Spottlo computes SoV this way by default: presence per prompt-run, four engines reported separately, competitor set derived from what actually appeared in the answers rather than from a list you guessed at up front. That last part matters more than it sounds, because the competitors AI names are frequently not the ones on your battlecards.

What to do next

1. Recompute your current SoV using presence, not name counts. If your number drops, it was wrong before.
2. State your denominator explicitly in every report: all brands, or a named competitive set. Pick one, label it, don't switch.
3. Report SoV and mention rate together. They diverge, and the divergence is the insight.
4. Split by engine before you blend, and check whether one engine is carrying your average.
5. Run a free scan to get a baseline SoV across four engines, then re-measure in four weeks against the same prompt set. The first delta is the only one that tells you whether anything you did worked.