AI Share of Voice: When the SERP Is One Answer

Ask an assistant to recommend a tool in your category and you get three names. Not ten links and a page of options — three names, a paragraph each, and a couple of sources underneath.

If you are not one of the three, what is your share of voice? Zero, on that answer. Ask again next week and you might be one of three, which makes it thirty-three percent. Neither number means anything on its own, and that is the problem with carrying the old metric across.

Share of voice still works in AI search. It just has to be rebuilt around a different unit — so here is the definition that survives, the three numbers underneath it, and the part that decides whether any of them are worth reading.

What share of voice was built for

The metric came from advertising: your presence as a proportion of all presence in the category. Search made it concrete. Ten organic positions, a few ads, some SERP features — a fixed inventory that several brands occupy at once.

Two properties made it work. The surface holds many brands simultaneously, and being on it is a matter of degree: position eight is less visibility than position two, but it is not nothing. Visibility was divisible, so it could be shared out.

Why the metric breaks on an AI answer

An AI answer is a single artifact. It names two or three brands and stops. There is no eighth slot quietly earning you a sliver — you are in the answer or you are absent from it, and absence is the common case.

There is also no stable surface to take a share of. Google’s description of how it decides relevance is that it is “determined by hundreds of factors, which could include information such as the user’s location, language, and device (desktop or phone)”. One answer is one sample from that, not a reading of the market.

Which is why a share taken from a single answer is either 33% or 0%, and why the instinct to check by opening a chat window and typing your category is worse than useless. It returns a number that feels like measurement and behaves like a coin toss.

What AI share of voice actually measures

The working definition: AI share of voice is the proportion of answers, across a fixed set of prompts, in which your brand appears — measured against the brands that appear instead of you.

The unit moves from position on a page to presence in an answer, and the metric only exists in aggregate. One answer tells you nothing. Fifty answers to the same fifty questions, run again a month later, tell you something. Underneath the headline there are three separate numbers, and they fail in different ways:

A fourth thing is worth wanting and hard to get: prominence. Named first in the answer is not the same as appearing in a list at the bottom of it, and almost nothing reports the difference, ours included. For now that one is a manual read of the answer text on the prompts you care most about.

The prompt set is the methodology

Everything above depends on the denominator, and because there is no fixed inventory to inherit, you have to define it. A share of voice number with an undeclared prompt set is not a measurement, it is a selection.

Four rules make it hold up:

• Write the questions the way a buyer types them.Not keywords. “What’s the cheapest way to track keyword rankings for a small site” is a prompt; “rank tracker” is not.
• Fix the set and keep it fixed. Fifteen to thirty questions is enough to move the number out of coin-toss territory. Adding prompts between runs changes the metric, not your visibility.
• Declare the comparison set. Brand share is a share of a named group. Three competitors and three hundred give very different percentages from identical data.
• Read the trend, not the level. The same prompt can produce a different answer tomorrow. A single run is a sample; the direction across runs is the signal.

If you already have a keyword list, it is a reasonable starting point for the set — group it into topics first, so you can see which parts of your category the prompts actually cover. The free keyword clustering tool does that in the browser, and the clusters with no question in them are the gaps in your denominator.

Being straight about our own instrument here, since the whole argument is that a share number is only as good as its denominator: AI Visibilityreports mention counts split across Google AI Overviews and ChatGPT, the questions where you were named, the domains cited in those answers, and a citation share against up to three competitor domains you nominate. It works from an index of observed answers rather than a prompt list you hand it, so the denominator is the set of questions we have seen you in — not a set you defined. That makes it a good instrument for trend and for finding cited sources, and the wrong instrument for “what percent of my category do I own”. Nobody can honestly sell you that one.

What the number is not

It is not traffic. Even Google’s own Generative AI performance report counts “how many times links to your site were shown to a user in a generative AI feature on Google Search” — impressions, no clicks — and it is still being rolled out “to a subset of website owners”. Whatever share of voice you compute, it is a visibility reading, never a session forecast.

It is not a lever.The temptation on seeing a low number is to go and manufacture the mentions that would raise it. Google’s own optimization guide says “seeking inauthentic ‘mentions’ across the web isn’t as helpful as it might seem”, which is the short version of a longer argument made in the post on ranking in AI Overviews. The metric is a diagnosis, not a dial.

Read as a diagnosis it earns its place. The prompts where you are absent are a content brief. The domains cited on those prompts are the map: pages that already answer your buyers’ questions, that the models already trust, and that you are either not on or not beating. That list is worth more than the percentage sitting above it.

The short version

Share of voice assumed a surface that holds ten brands at once and gives partial credit for eighth place. An AI answer names two or three and stops, so the metric has to be rebuilt on presence across a fixed prompt set instead of position on a page — mention rate, brand share and citation share, read as a trend and never from a single answer.

The denominator is the whole game, which means most quoted AI share of voice figures are unfalsifiable. Ours is an approximation too, and it is labelled as one. If you want the underlying facts — the questions where your brand comes up, who gets named next to you, and which domains the answers were built from — run an AI Visibility report and start the trend line, because the first run is only ever a baseline.