Share of Voice in AI Answers: How to Measure It Properly

Someone in your leadership team has asked what share of AI answers your brand is showing up in. It’s a fair question, and most attempts to answer it produce a number that cannot be defended. Share of voice in AI answers is measurable, but only if you define the denominator before you count anything.

What does share of voice mean when there is no results page?

In traditional search, share of voice borrowed from advertising: your visibility relative to competitors across a defined keyword set, usually weighted by search volume and position. There were ten blue links and a stable structure to measure against.

AI answers have no fixed slot count. One answer may cite six sources, the next may cite none and simply assert a fact. Your brand may be named in the answer text without any link, or linked without being named. The same prompt run twice can return different sources. Any share-of-voice metric that ignores these properties will be unstable and unpersuasive.

The workable definition: share of voice is the proportion of a defined set of buyer-relevant prompts in which your brand appears, relative to the appearance rates of the competitor set you are measuring against. Everything hinges on “defined set” and “competitor set”.

Build the prompt set first, and freeze it

The prompt set is the denominator. If it changes between measurement rounds, your trend line means nothing.

Construct it from questions your buyers actually ask, not from keyword tools. Practical sources: discovery-call transcripts, RFP questions, support tickets, and the comparison questions a buyer asks when they have narrowed to a shortlist. Write each prompt as a complete natural-language question.

Then structure it deliberately:

• Category prompts — questions about the problem space, where no vendor is named. These are the fairest test of visibility.
• Evaluative prompts — “who are the leading providers of…”, “what should I look for when choosing…”. These are where share of voice is competitively meaningful.
• Brand prompts — questions naming you directly. Useful for accuracy checking, but exclude them from share-of-voice maths. You will always appear, and including them inflates the number.

Aim for 40 to 80 prompts. Freeze the wording. Version the set, and record the date whenever you change it.

Define the competitor set honestly

Share of voice is a ratio, so the choice of comparison set determines the answer. Two failure modes are common. Selecting only weaker competitors produces a flattering number that collapses the moment someone tests it. Selecting every company in the category produces a number so small it’s useless for decision-making.

The defensible approach is to derive the competitor set from your own baseline data. Run the prompt set once, record every organisation named across all answers, and count frequency. The organisations that appear most often are your competitive set in AI answers — regardless of whether they are the ones your sales team names. This often surfaces surprises: aggregators, directories, professional bodies and publishers frequently occupy more answer space than any vendor.

How to compute the metrics properly

Record, for every prompt and every run, three separate binary outcomes for each organisation in your set: named in the answer text, cited by URL, or both. Then compute:

• Appearance rate = prompts where the brand appears (named or cited) / total prompts tested.
• Citation rate = prompts where one of the brand’s URLs is listed as a source / total prompts tested.
• Share of voice = your appearance count / total appearance count across all brands in the defined set.

Note what that last formula means. If ten prompts each name three companies, the total appearance count is thirty, not ten. Share of voice is a share of slots, not of prompts. Getting this wrong is the single most common calculation error, and it makes cross-period comparison invalid.

An alternative, sometimes more intuitive for executives, is average share per prompt: for each prompt, your share of the brands named in that answer, averaged across prompts. This gives every prompt equal weight regardless of how many brands the answer happened to name. Pick one method, document it, and never silently switch.

Handle variance, or your trend is noise

AI answers are non-deterministic. The same prompt can produce different sources on consecutive runs. If you measure once per period, you are measuring randomness as much as performance.

Three controls make the number trustworthy:

1. Repeat each prompt at least three times per measurement round and treat the brand as appearing if it appears in the majority of runs, or record a fractional appearance score. Either is defensible; mixing them isn’t.
2. Hold conditions constant. Same assistant, same model version where visible, no personalisation, no chat history, no location signal you haven’t recorded. Personalised sessions aren’t measurement.
3. Report a range, not a point. A share of voice reported as a single decimal implies a precision the underlying system doesn’t have.

Segment the number or it will mislead you

A single blended share-of-voice figure hides everything actionable. At minimum, split by:

• Assistant. Different systems use different retrieval sources. Strong performance in one and absence in another is common, and it points to different fixes.
• Prompt type. Category, evaluative and comparison prompts usually show very different profiles. Being visible on informational prompts while absent on evaluative ones is a specific, addressable problem.
• Mention versus citation. These are different phenomena with different causes, and they should never be summed into one figure.

What good measurement looks like in practice

A credible share-of-voice report states: the prompt set version and size, the assistants and dates tested, the number of runs per prompt, the competitor set and how it was derived, the formula used, and the confidence range. It shows the trend over at least three periods before drawing any conclusion about direction.

Anything less is a screenshot exercise. Our ARIA citation tracker is built to hold these conditions constant across runs, which is the hard part of doing this by hand. If the underlying issue turns out to be that your pages are hard to extract from rather than hard to find, our AI citability scorer is the better starting point.

What to do with a low number

Don’t respond to a low share of voice with more publishing. Diagnose first. If competitors are cited and you’re absent everywhere, the problem is usually access or corroboration. If you appear on brand prompts but not category prompts, your content is about you rather than about the buyer’s problem. If you’re named but never cited, your ideas are circulating without your URL attached — an attribution problem, not a discovery one.

Each of these has a different remedy, and the segmented data tells you which one you’ve. That’s the entire reason to measure properly rather than approximately. More on the wider approach on our AI visibility page.

Frequently asked questions

How many prompts do I need for a reliable share of voice?

Forty to eighty for most organisations. Below forty, a single prompt swings the percentage too much to distinguish real change from variance. The prompt set matters more than its size — buyer-realistic questions beat volume.

Should brand-name prompts be included?

No, not in the share-of-voice calculation. You’ll nearly always appear when named, which inflates the metric. Track them separately as an accuracy check on how assistants describe you.

Why does my share of voice change without me doing anything?

AI answers are non-deterministic and retrieval indexes update continuously. Repeat each prompt several times per round and read the trend across periods rather than reacting to single-round movement.

Can I compare share of voice across different assistants?

You can report them side by side, but not combine them into one figure. Each assistant uses different retrieval sources and citation behaviour, so the numbers measure different things.