Definition · Measurement
The formula is simple. The four inputs feeding it are judgment calls, which is why two tools measuring the same brand routinely report different numbers.
By Izzy Gregorio · Updated August 2026 · 8 min read
In short
AI share of voice is the percentage of AI-generated answers in a category that cite your brand, calculated as your brand citations divided by total category citations, multiplied by 100. A share of voice figure with no named competitor set and no published prompt list is a vanity metric wearing a percentage sign.
The formula
The calculation in circulation is straightforward.
(Your brand citations ÷ total category citations) × 100
Run 25 prompts across four engines. Count how many responses cite your brand. Count total brand citations across all responses. Divide, multiply, done.
Now the part nobody puts in the marketing material. Four decisions sit upstream of that division, and each one moves the result by a wide margin.
| Decision | The options | How much it moves the number |
|---|---|---|
| Which prompts | Broad category terms, buying-intent terms, or a mix | Enormously. Broad definitional prompts flatter incumbents. Buying prompts flatter specialists |
| Which engines | One, or four, weighted equally or by audience | Substantially. SOCi's 2026 index found ChatGPT surfaces a local business 1.2 percent of the time against Gemini at 11 percent |
| Mention or citation | Named only, or named with a resolving link | Roughly doubles or halves most scores |
| Who is in the denominator | Everyone named, or a fixed competitor set | Changes the meaning entirely. Open denominators drift month to month |
Pick your four answers, write them down, and never change them silently. A prompt set you quietly edit produces a trend line measuring your editing.
The distinction
These get used interchangeably and they measure different things.
| Term | What it counts | What it tells you |
|---|---|---|
| Share of model | Responses that mention your brand, out of total responses | Your presence. Are you in the conversation |
| Citation rate | Mentions that include a link, out of total mentions | Your attribution. Do you get credit |
| Share of voice | Your citations out of all brand citations in the category | Your position relative to everyone else |
Semrush named the gap between the first two in 2025 and called it the mention-source divide, finding fewer than one in five brands achieve both frequent mentions and consistent citations.
That gap is diagnostic. High mentions and low citations means the model knows you and your pages are not extractable. Low mentions and high citations means your pages work and the entity is weak.
The blended score hides which one you have, which is precisely why a single number is the least useful format for this data.
Before you trust the number
Both are fixable, and neither is obvious until the score swings and nobody can explain why.
Answer volatility. AirOps found only 30 percent of brands remain visible across consecutive AI responses. Ask the same question twice and roughly seven times out of ten a different set of brands comes back. A score built on one run per prompt is measuring that volatility as much as it is measuring you.
Denominator drift. If your calculation counts every brand the engines happen to name, the comparison set changes month to month and your percentage moves without anything about you changing.
The fixes are different and you need both. Volatility is fixed by running each prompt at least three times per engine and recording the rate at which you appear rather than whether you appeared. Drift is fixed by naming a competitor set. Fixing only the second one and expecting a stable number is how teams conclude the whole method is unreliable.
The judgment call that matters most
An open denominator counts everyone the engines happen to name. That sounds neutral and it is actually unstable.
A fixed competitor set fixes that. Choose three to five brands you actually lose deals to, and calculate against that set only.
Two things happen when you do. The number becomes stable enough to trend. And it becomes usable in a leadership meeting.
Compare the two sentences.
Our share against these four named competitors moved from 8 percent to 14 percent.
Our AI share of voice is 11 percent.
The first is a sentence a board can act on. The second is a number somebody will ask about once and never again.
Translate it
| If you are the | The version of this number you need | The decision it drives |
|---|---|---|
| CMO | Share against a named competitor set, trended over three months | Whether to fund citation supply or structural fixes |
| Marketing director | Citation rate by page | Which twenty pages get restructured first |
| Founder still holding marketing | One verbatim answer where a competitor was named instead of you | Whether this becomes a priority at all |
| Ecommerce lead | Share of voice on buying-intent prompts only | Whether product data is legible enough to be selected |
| Agency being evaluated | The frozen prompt set itself | Whether they are measuring or reselling a dashboard |
Notice that only one row wants a single blended percentage, and it is nobody's row.
Do it without buying anything
An afternoon produces a defensible baseline. Every month after that takes a fraction of the time, because the instrument already exists.
Write 25 prompts. Five prompt types across five real customer situations: definition, comparison, alternatives, use case, buying.
Name three to five competitors. Fixed set, written down, brands you actually lose deals to.
Run every prompt three times per engine, in a logged-out browser, across ChatGPT, Claude, Gemini, and Perplexity. Three runs is the minimum that survives answer volatility.
Log four columns per response. Your brand mentioned, your brand cited with a link, which competitors named, and the date.
Calculate. Your citations divided by total citations across your fixed set, times 100.
Save the prompt list verbatim. This is the instrument. Do not edit it.
One question worth sitting with. If your agency reported an AI share of voice number last quarter, could they produce the prompt list that generated it? Not a description of the methodology. The actual list, and how many times each prompt was run.
That single request separates measurement from a dashboard subscription, and it costs nothing to ask.
Common questions
Your brand citations divided by total category citations, multiplied by 100. The arithmetic is simple. The result depends entirely on four upstream choices: which prompts you run, which engines you include, whether mentions without links count, and which competitors are in the denominator.
No credible benchmark exists, because no independent cross-platform standard has been published and category dynamics vary widely. Semrush's 2026 study of 1,094 US categories found 53.7 percent had no consistent leader at all, so absolute scores are low in most categories. Trend your own number instead.
Share of model measures how often you appear across total responses, which is presence. Share of voice measures your citations against all brand citations in the category, which is position relative to competitors. One tells you whether you are in the conversation, the other tells you how much of it you hold.
Yes. Write 25 prompts, name three to five competitors, and run everything three times per engine in a logged-out browser across four engines, logging mentions, citations, and competitor names. That produces a defensible baseline in an afternoon, and the instrument is reusable every month afterward at a fraction of the effort.
Two causes, and most teams only fix one. Answer volatility: AirOps found only 30 percent of brands remain visible across consecutive AI responses, so a score built on single runs moves on its own. Denominator drift: if your calculation counts every brand the engines happen to name, the comparison set shifts month to month. Fix both by running each prompt at least three times per engine and using a fixed competitor set.
Get the data
The prompt set structure, the competitor-set method, and the citation data behind the numbers on this page. Including the four decisions written out, so you can copy the method rather than rebuild it. Sources you can trace and a calculation you can defend in a board meeting.