AI share of voice

What AI share of voice measures

Every definition of AI share of voice is a fraction: your brand mentions on top, all brand mentions underneath. The interesting half is underneath, and the published definitions do not agree about what is down there.

The denominator is a chosen prompt set, not the category.

Five vendor definitions, read 4 September 2026What your own tool divides by

How the main tools define the metric, in their own words.

The clearest case is the first row, because it needs no comparison across sources. Both answers sit on one page.

One page, the formula and the caveat that undoes it

1. 1The formula is printed: your brand mentions over total category mentions across tracked prompts, times one hundred.
2. 2The next bullet says the score varies by platform, and gives the example of appearing often in ChatGPT and barely registering in Perplexity.
3. 3Both lines are true and they do not sit together comfortably. A single percentage cannot carry a number that changes with the platform it was measured on.

Mentions are counted across a defined set of prompts, and two of the definitions put that in writing. Semrush describes the denominator as all brands in your category and names no prompt selection or sample size.

A second page calls it the percentage of AI-generated responses in your category that mention you, then names a prompt set in the instructions underneath. Its definition and its method are counting different things.

A third page defines the metric and never says what the denominator is at all. Google’s own AI answer puts it relative to your total category competitors, a denominator none of the tools here measures.

One platform sells two numbers under the one name, a mention-based share and a citation-based share, and they count different events in the same answer. Why those two events are not interchangeable sits in what AI visibility is.

Why the denominator cannot be audited

The reason this matters more here than in the metric it borrowed its name from is arithmetic, not ethics. The set of prompts a person could ask an answer engine has no edge, so nobody can compute the true total.

The name came from somewhere, and so did an expectation.

Nielsen (August 2009) and the definitions above, read 4 September 2026Which properties survived the borrowing and which did not

The left column is the metric with a measured relationship behind it.

The advertising version has a number attached. What matters here is the shape, not the size.

The universe of possible prompts is effectively infinite, so a vendor selects a small subset of static prompts and aggregates those outputs into a percentage that reads as global. Dan Taylor calls the result a hidden denominator.

The definitions do not hide it. One writes the denominator into its own as total brand mentions across tracked prompts.

And the answers the set produces do not hold still while you measure them.

What to ask before you accept the number

Settle one shorter thing first, because it decides whether either of those questions is worth asking. What does your own vendor divide by? None of this makes the metric useless; it makes it a measurement with a stated sample, and the difference is four questions long.

An undisclosed sample makes the percentage unreadable from outside.

Taylor’s constructive half is worth taking as well, because it separates one number into three that fail differently. Share of mentions is how often you are named. Share of recommendations is how often you are named when the question asks the engine to advise. Share of narrative is what gets said about you when you are named. His example should worry anyone reporting a single figure. A brand can hold a high share of voice while being consistently described as a complex legacy system, in which case the number is going up while the pipeline goes down.

What counts as a good percentage

The most common question about this metric is what number to aim for. It has an answer, and the answer is that the question is malformed unless the category is specified.

A number is a verdict only once the field is counted.

One definition states it plainly: a good percentage depends entirely on your market’s competitive context. In a category with two serious players, 50 percent suggests parity. In a fragmented category with ten alternatives, 15 percent may be category leadership. A benchmark quoted without the number of competitors behind it is not a benchmark.

The same page offers the closest thing to a distribution anyone publishes. Most B2B brands appear in under 30 percent of relevant category queries, regardless of their conventional search rankings. That is one vendor’s observation, not a study, and it is useful in one direction only. If your number is far above it, the first thing to check is the size of the prompt set, not the strength of the brand.

Read the practitioner case against that. A practitioner published a piece on LinkedIn explaining why AI share of voice became his primary metric for the year. He tracks what he describes as thousands of queries in real time. His stated reason is that a buyer asking an engine for the best platform in his category either sees his product or does not. The disclosure stops in the same place as everyone else’s: the piece does not say how many prompts, or how the set was chosen. The percentages it shows are illustrative, not data.

One published method does state its own. The GEO paper by Aggarwal and colleagues appeared at KDD 2024. It sets out a framework for defining visibility metrics, then publishes the benchmark it was tested on.

That is the disclosure every page above skips, and the one to ask your own vendor for. You can disagree with their metric and still see what it counted, over which queries. That is the difference between a measurement and a claim.

Measuring your own share, with the swing between runs reported, is our AI visibility service.

Four ways AI share of voice misleads

If a benchmark cannot be borrowed from outside, the remaining risk is inside your own reporting, and it is four specific readings. Three come from the prompt set behind the denominator being invisible, and the fourth from two different numbers sharing a name.

The number can rise while nothing about you changed.

The fifth one you cannot fix from your side. Almost every number in this field is published by a company selling visibility tracking. That does not make the numbers wrong, but none of them has been checked by anyone independent. What you can check is the wording: every definition above is quoted or linked, so you can read it at the source.

What to record so the number stays readable

If you keep this metric, the fix is not a better formula. It is recording the sample beside the percentage, every time, so that next quarter’s number can be compared to this one.

Write the sample next to the share.

Four values make the figure comparable to your next one. How many prompts. How they were selected. Whether a mention or a linked citation increments it. And how many times each ran. They do not make it reproducible by anyone else; that would take the prompts themselves. What they tell you is whether the sample behind this month is the sample behind last month. A share of voice reported without them is a number that cannot be checked, including by the person who reported it three months later.

What your own analytics will and will not separate sits in how to measure AI visibility. What the underlying score is made of is in the AI visibility guide.

What this metric is good for

It is a usable internal trend line and a poor external benchmark, for the same reason in both cases. Whoever chose the prompt set decided what the denominator counts.

Keep it inside one instrument, and record the sample.

Held still, meaning the same prompts, the same engines and the same counting rule, the percentage becomes readable as your own series. It does not become stable: the answers still move between runs, and platform changes set a floor on precision that no amount of repetition beats. What holding the configuration still removes is the movement you caused. Set beside a competitor’s published figure, or beside a benchmark from a different category, it is comparing two fractions built over different samples. Inspeccia, which sells a visibility product, sizes that instability from another angle with its own production data. Across 167 domains measured twice, the visible ones moved by an average of 30.8 points. That is what a percentage inherits when the counts underneath it move. That is not a small caveat about a good metric; it is the whole reading.