How Share of AI Voice is calculated, and what the number cannot tell you

Share of AI Voice is a ratio, not a score out of ten. Take a fixed set of prompts, put them to AI answer engines, count how many times your brand is named, count how many times your named competitors are named, and divide.

Your mentions, divided by your mentions plus your named competitors' mentions, inside one week, on one frozen prompt set, on recorded model versions.

That is the whole calculation. Everything that makes the number defensible or useless sits in the four words doing the work: fixed, named, one week, recorded. Change the prompt set and the number moves. Change the competitor list and the number moves. Sample a different week and the number moves. Let a provider ship a new model inside your window and the number can move while nothing in the market has changed at all.

So the arithmetic is simple and the measurement is not. The rest of this guide is about that gap, because a brand manager carrying the figure into a board meeting has to be able to defend both halves: the sum, and the sampling that fed it.

Share of voice arrived in marketing from PR and social listening, where the denominator could plausibly be the whole conversation: all the press mentions, all the posts, all of it. AI answers do not work that way. There is no corpus of every answer ever generated, because each answer is produced on demand, once, for one person asking. You cannot measure the whole conversation. You can only measure the conversation you asked for.

That forces a choice, and the choice is the method. The honest version of the metric compares you against a named set of rivals, so the denominator is a market you defined and can explain in a sentence. Seenby takes your brand name, a description and up to three competitor names at setup, and those names are the denominator for every week that follows.

Two consequences matter more than the decimal places.

• A share with no competitors in it is not a measurement. If the denominator contains only your own mentions, the arithmetic returns one hundred percent every week, forever. Naming rivals is what turns a mention count into a share.
• The competitor set is an editorial call, not a fact. Name the three incumbents and your share looks modest. Name three smaller challengers and it looks strong. Neither list is wrong, but the resulting number is only comparable to itself, and only for as long as the list stays still.

We would rather you argued about the competitor list at setup than about the score in month four.

Seenby reports a single headline share per brand per week. Ranking position and sentiment are extracted and stored per response, and shown in their own views, rather than folded into the headline through a weighting scheme nobody outside the vendor can inspect.

That is a design choice with a reason behind it. A composite score that blends presence, rank and tone into one figure is easy to put on a slide and impossible to argue with, because you cannot tell which component moved. Keeping them apart means a flat share with worsening sentiment reads as what it is, which is a different problem from a falling share, and calls for a different response.

So the dashboard is built around four views that each answer one question. A 30-day trend line for Share of AI Voice, so you can see whether this week is a wobble or a direction. A sentiment breakdown, including the absent count. The top prompts where the brand is named and the top prompts where it is not. And a competitor benchmarking view against the names you chose.

When the number surprises you, the designed next move is the per prompt evidence, not a harder stare at the percentage. Growth and Agency include per-prompt evidence answer by answer and the full scrubbed answer log for as long as the account exists; Starter keeps 30 days of answer log history. The point of the log is that absence has an address: a specific question, on a specific engine, in a specific week.

This is the objection most often raised against the whole category, usually phrased as "the results change every time I ask", and it deserves a straight answer rather than a reassurance.

The objection is partly correct. OpenAI's own API documentation states that chat completions are non-deterministic by default, meaning outputs may differ from request to request, and it exposes a system fingerprint field precisely so that developers can tell variation caused by their own settings from variation caused by changes the provider has made on its side. Providers also retire and replace dated model snapshots on a published schedule, with notice periods and recommended replacements. The ground under any AI visibility measurement moves, and no vendor controls that.

So a measurement method here has two jobs: reduce the noise it can reduce, and label the noise it cannot. A frozen prompt set and a weekly window handle the first. Model-version tagging on every stored response handles the second, and the weekly digest is built to carry drift alerts tied to model-version changes when applicable, alongside the methodology caveat.

The caveat is worth stating in full, because it belongs in your reporting as much as in ours: a week over week move in Share of AI Voice may reflect a model update rather than a real change in how the market sees you. When a model version changed inside the window, treat the move as unexplained until the per prompt evidence says otherwise. A number you can question is worth more than a number you cannot.

Five limits, stated plainly, so you are not surprised by them in a meeting.

It is not a measure of traffic or demand

Share of AI Voice counts how often you are named in a sample of AI answers. It says nothing about how many people asked, what they did next, or what any of it was worth. We make no claim that the score predicts clicks, pipeline or revenue, and you should not present it as if it did.

It is not the whole of AI search

The measurement covers the engines it queries, OpenAI and Perplexity, on your frozen prompt set, in one week. Report it that way. A score with its scope attached survives questions; a score presented as "our AI visibility" invites one you cannot answer.

It is not anyone else's session

Every answer is generated on demand, and identical requests can return different text. The sample tells you what the engines tended to say to the questions you chose, in that window. It is not a transcript of what your buyer saw on Tuesday.

It does not explain itself

The share tells you the proportion. It does not tell you why. Causation lives in the per prompt evidence, the sentiment breakdown and your own knowledge of what you published and when, and even then you are reasoning, not proving.

Rank and sentiment are classifications, not verdicts

Ranking position is recorded for each mention inside a response, and sentiment is classified into positive, neutral, negative or absent. Both are useful signals and neither is a human reading of the answer. Where the stakes are high, read the answer.

A brand manager's real problem is not computing the score, it is defending it to a CMO or a board that has every right to be sceptical of a metric invented eighteen months ago. Five habits make that defence short.

• State the denominator first. "Against these three named competitors" is a six word phrase that prevents most of the argument.
• Quote the sample with the score. The prompt count, the two providers and the week. A percentage with no sample behind it is a decoration.
• Compare like windows only. Week against week, from the same frozen prompt set. The moment the prompt set changes, the series starts again, and you should say so.
• Carry the caveat, every time. If a model version changed in the window, say that before anyone asks. Seenby attaches the methodology caveat to every computed score and surfaces the disclosure on every score display for exactly this reason.
• Pair the share with the absences. The prompts where you were not named are the actionable half of the report. The share is the summary; the absent list is the work.

This guide is the arithmetic and its limits. Three sibling pages go deeper on the parts that do the heavy lifting.

• Share of AI Voice, defined is the shorter definition, for when you need to explain the metric to someone else in a paragraph rather than a page.
• Methodology on every score, not in a footnote covers where the caveat appears in the product and why it is attached to the record rather than published once.
• Model-version tagging goes through how a model update is distinguished from a real competitive move, which is the mechanism behind the drift section above.
• The LLM answer log covers the evidence layer: the stored, scrubbed answers behind each count, and what you can read in them.