Share of voice in AI answers: the metric that replaces rank #1

For twenty years the question was "where do we rank?". It had a clear answer: a position on a page, checked daily, charted monthly. AI assistants break that model. There is no page, often no order worth the name, and the answer changes each time you ask. If you want to know whether you are winning in ChatGPT, Gemini or Perplexity, you need a different number.

That number is share of voice.

Why rank does not transfer

A ChatGPT answer to "best accountancy software for UK sole traders" might name four products in a paragraph, two more in a comparison, and leave one out entirely. Run the same prompt again and the four become five, with a different first name. Position exists, loosely, but it is unstable and the gap between "mentioned first" and "mentioned third" is small compared with the gap between "mentioned" and "not mentioned".

Rank also assumes a fixed query. AI prompts are long and varied: the same intent is phrased a hundred ways. Measuring one phrasing tells you very little.

So the useful questions become:

• Across the prompts my customers actually ask, how often am I mentioned at all?
• When the assistant names someone, what proportion of the time is it me rather than a competitor?
• Is that proportion going up or down, per engine, over weeks?

The definitions

Seoptist uses two related metrics, and it helps to keep them separate.

Visibility is the proportion of answers that mention your brand:

visibility = answers mentioning you / total answers sampled

If you track 30 prompts on 3 engines with 3 samples each, that is 270 answers a week. Mentioned in 81 of them gives a visibility of 30%.

Share of voice (SoV) is your share of all brand mentions, counting you and the competitors you track:

share of voice = your mentions / (your mentions + competitor mentions)

If in those 270 answers you are mentioned 81 times and your three tracked competitors are mentioned 60, 45 and 24 times, SoV is 81 / 210, or about 39%.

Visibility tells you whether you exist in the conversation. Share of voice tells you whether you are winning it. A brand can have modest visibility and a high share of voice in a niche where assistants rarely name anyone, or high visibility and a low share in a crowded market where every answer lists eight names.

Why sampling is not optional

Because answers vary, a single run is a coin toss dressed up as data. Three samples per prompt per engine is the minimum at which the week-to-week line stops jumping around for the wrong reasons; more is better when budget allows. Seoptist runs three by default and five on Enterprise, and reports a confidence range alongside the headline figure so a two-point move is not mistaken for a trend.

The same logic applies to the prompt set. Ten prompts can be moved by a single engine update to one of them. Thirty to a hundred, covering the intents that matter (discovery, comparison, local, "is X any good"), give a number you can defend to a client or a board.

Reading the number

A few patterns we see often enough to call them patterns rather than findings:

• Engines disagree. It is common for a brand to have a strong share on Perplexity, which leans on review and community sources, and a weak share on Gemini, which leans on Google's index. Treat per-engine lines as separate channels.
• Share rises before visibility. When competitors fall out of answers (often due to a stale listing or blocked crawler), your share rises even though your own mentions have not changed. Check both before celebrating.
• Position matters at the margin. Average position is a useful secondary metric for comparison prompts ("X vs Y"), where being first in a numbered list is meaningful. For discovery prompts it is noise.
• Sentiment is a modifier. Being mentioned as "cheaper but slower" is not the same as being mentioned as "the obvious choice". Weight share by sentiment when you report upward.

What to put in a report

If you report one chart, make it share of voice by engine over time, with your two nearest competitors. If you report three, add visibility (to show whether the market is being named at all) and citation rate (the proportion of answers that link to your domain, which tells you whether your own pages are doing the work or whether third parties are speaking for you).

Avoid reporting a single week's number without context. A line of eight weekly points with a confidence band is honest and still looks good when it goes up.

A worked example

A regional removals firm tracks 30 prompts across ChatGPT, Gemini and Perplexity. In week one its share of voice is 18% on ChatGPT, 31% on Perplexity and 9% on Gemini. The audit shows its site blocks Google-Extended and has no LocalBusiness schema, and that a national directory lists an old phone number.

Over six weeks it fixes the three issues. Gemini share climbs to the mid-twenties as the corrected directory listing is re-crawled; ChatGPT is flat; Perplexity drifts up slightly. The firm cannot prove causation from one case, and neither can we, but the direction and timing match the fixes, and that is a report a client can act on.

Getting a baseline

You cannot improve a number you have not measured. The free AI Visibility Check gives you a first reading across five prompts and two engines. Paid plans from £49 a month (see pricing) track 30 to 300 prompts weekly with competitors, sampling and the confidence range built in.