Run your brand through two AI visibility tools this week and you will get two different Share-of-Voice numbers. One says you own 22% of your category in AI answers; the other says 14%. Same brand, same week, same category. Neither tool is broken, and neither number is wrong.
Share of Voice is a ratio, and a ratio is only as meaningful as the two numbers underneath it. Every tool that computes SoV makes three decisions about those numbers, quietly, before it ever shows you a percentage. Once you know what the three decisions are, you stop trusting the number and start reading it, which is the only safe way to use it.
Decision 1: what counts as "presence"
The first fork is whether the tool counts a mention or a citation. A mention is your brand named in an AI answer with no link. A citation is a linked source the engine points to. These are different events, and they happen at different rates.
A brand can be named constantly and cited rarely, or the reverse. A tool that counts mentions will report a higher SoV for a well-known brand the model already knows; a tool that counts only linked citations will report a lower one, and will favour brands whose content the engine actively retrieves. Neither is the "true" presence. They measure two different things, and if two tools disagree here, their SoV numbers cannot be reconciled.
Decision 2: the denominator
The second fork is the size and shape of the sample. SoV over 20 prompts on three engines is a different number from SoV over 50 prompts on five. Add Google's AI Overviews and AI Mode to a sample that previously ran only ChatGPT, Perplexity, and Gemini, and a brand strong on Google surfaces will see its SoV jump, not because anything changed in the world, but because the denominator did.
This is why a SoV figure with no stated prompt set and no stated engine list is close to meaningless. The percentage is real; the thing it is a percentage of is undefined.
Decision 3: how it aggregates (the one that hides)
Want to know how your brand scores on these same metrics?
We'll run 20 prompts across 3 AI platforms and send your report within 24 hours.
The third fork is the most technical and the most consequential, because it is invisible in the output. Once a tool has mentions per prompt per engine, it has to roll them up into one number, and there are two ways to do it.
The wrong way is to compute a rate for each prompt (or each engine) and then average those rates. The right way is to pool: sum your brand's mentions, sum all brands' mentions, and divide once.
They diverge exactly when coverage is uneven, which in AI search it always is. Work a small example. Your brand is one of five in a category, measured across two prompts. Prompt A returns 50 brand mentions in total and your brand gets 10 of them, a 20% share. Prompt B returns just 4 mentions and your brand gets 2, a 50% share. Average those two shares and your Share of Voice reads 35%. Pool them instead, 12 of your mentions out of 54 total, and it reads 22%. The averaging method nearly doubled your share on the strength of a four-mention prompt, while pooling reports what actually happened across all 54 mentions.
Because different AI engines return wildly different volumes for the same brand, that gap is not an edge case; it is the norm. A tool that averages drifts away from the share a pooled tool reports, and neither of them tells you which method it used.
This is the decision that most often explains two tools disagreeing, and it is the one you can never see from the dashboard.
The three decisions at a glance
| The decision | Two common choices | Why it moves your number |
|---|---|---|
| What counts as presence | a mention (named, no link) vs a citation (a linked source) | mentions favour brands the model already knows; citations favour brands whose content gets retrieved |
| The denominator | few prompts on three engines vs many on five | adding an engine a brand is strong on lifts its share with nothing changing in the world |
| How it aggregates | average per-prompt rates vs pool all mentions | averaging over-weights tiny samples; pooling weights by actual volume |
Two tools that answer these three differently cannot produce the same number, and neither is wrong within its own method.
How Cited computes it, in the open
We count mentions, pooled, across a stated prompt set and the five engines in the Cited Index (ChatGPT, Perplexity, Gemini, Google AI Overviews, Google AI Mode). Share of Voice, the second of the Cited 8 metrics, is your mentions divided by every brand's mentions in the category, summing numerators and denominators before dividing, never averaging per-prompt rates. It is one specific, defensible set of the three decisions above, stated plainly so you know exactly what the number is a percentage of. That is the point: not that our number is the only right one, but that you can see how it was made.
How to read any Share-of-Voice number
You do not need to audit a tool's source code. You need to ask it three questions, and a good tool will answer all three:
- Mentions or citations? Know which event you are counting.
- Over what? How many prompts, and which engines. A number without a denominator is decoration.
- Averaged or pooled? If a tool cannot tell you, assume averaged, and treat the number as soft.
A SoV figure that comes with those three answers is a real instrument. One that does not is a vanity metric, and the fastest way to get burned is to compare it against a second tool that answered the three questions differently. And once you can read the number, the next question is whether it is good; for that you need a benchmark of what a strong share actually looks like, which for Indian brands is lower than most teams expect.
Share of Voice is genuinely useful, but only within one consistent method, tracked over time. Pick a tool, learn exactly how it does the arithmetic, and watch the trend. The absolute number is a house style. The direction it moves is the signal. This is the same reason we do not chase prompt volume: the count only means something once you have pinned down what is being counted.
Want a Share-of-Voice number you can actually read, tracked with one consistent method? Start a free trial and track your Share of Voice across the engines.