Share of Voice in AI Search: Definition, Formula and a Worked Example

Ask two AI visibility tools for the same brand in the same week and you can get answers that differ by a wide margin. Both can be correct. The cause: "share of voice" for AI answers lacks one agreed formula, and most reports do not state which they chose.

The definition in one sentence

In plain terms, AI share of voice is the slice of all brand mentions, or citations, that your name takes within one fixed list of prompts, set against every brand on your tracking list. One practitioner guide defines it as the percentage of AI-generated responses that mention, cite, or recommend your brand across a defined set of category prompts. The idea comes from advertising, moved onto the answers that assistants like ChatGPT, Perplexity and Gemini write.

Because the measure is relative, it places you among named rivals. It does not size the market, and it cannot judge whether an answer was correct.

Why teams track it

Pew Research Center analysed browsing data in 2025 and found that visits with an AI summary produced a click on a result link 8% of the time, versus 15% without one, and a click on a link inside the summary about 1% of the time. When fewer decisions end in a click, the name in the answer is what a buyer sees first, and share of voice lets you compare that first sight with competitors.

Three formulas that all carry the same name

The LLM Pulse guide lists three versions. They differ in the numerator and the denominator, so they give different results from identical data.

• Mention share: your mentions divided by all tracked brands' mentions.
• Citation share: citations of your domain divided by all citations in the answers, including news sites, encyclopedias and forums.
• Position-weighted share: a mention counts for more when it comes first in the answer.

Quattr publishes a fourth, a blend that weights a citation at 1.25 and a mention at 1.0, and says plainly that these weights are its own product decision and not something the assistants report. None of this is wrong. The trouble starts when a number from one definition is compared with a number from another.

A worked example with invented numbers

Imagine four brands, 25 prompts and one assistant, with each brand counted once per answer. The counts below are invented to show the arithmetic of the mention-share formula from that same guide.

Brand	Mentions	Share of voice	Mention rate (mentions per 25 prompts)
Your brand	14	23.3%	56%
Rival B	22	36.7%	88%
Rival C	18	30.0%	72%
Rival D	6	10.0%	24%
Total	60	100%	not additive

Your share is 14 divided by 60, a little under a quarter. Your mention rate is 14 divided by 25, over half the prompts. The two figures answer different questions: the first asks how you compare, the second asks how often you appear. Rates do not add up to a whole because several brands can appear in one answer.

The example that mixed up its own denominator

Vendor examples are not always consistent. AirOps writes its formula as citations divided by the citations of all tracked brands, but its worked case divides 60 and 90 appearances by 200 prompts and reports 30% and 45%. That is a mention rate. Applying the stated formula to the same two brands would give 40% and 60%, if those two were the only brands tracked. Always read the denominator in an example before adopting its method.

Setting up a count you can trust

• Freeze the competitor list. Adding or removing a rival changes the denominator, so your share moves without anything changing on the web.
• Count once per answer. A brand named three times in one reply scores one. Write the rule where your team can see it.
• Split by assistant and by prompt type. Strong discovery and weak comparison need separate fixes. Never average per-assistant percentages into one figure.
• Check names. A name that is also an everyday word gets miscounted, so read a sample by hand.
• Verify accuracy. A mention that describes you wrongly is not a win.

If you are building the count by hand, the spreadsheet playbook is a start, and why answers change every run explains the repeat runs. For the tool question, see how to choose a tracker.

Twenty prompts is a pilot, not a sample

Beginner guides, including the AirOps tracking guide, suggest starting with 20 to 30 prompts. LLM Pulse recommends at least 50, ideally 100 to 200, repeated runs, and about four weeks before you set a baseline, and notes that a one-week swing of around five points is usually noise. We side with the larger number for any figure you plan to report upward. Over 25 prompts, a few answers shift your share by several points, which is exactly the size of change people cheer or fret about.

Use the small set to learn the method and find broken prompts. Do not put its output in a board slide. We would soften this view if your category is so narrow that only a few dozen distinct buyer questions exist; then a smaller fixed set, run many times, is the honest design.

Reporting it without overclaiming

Every figure you pass on should travel with four facts: the formula, the number of prompts, the number of repeat runs and the list of competitors. A line such as "mention share, 100 prompts, three runs each, five tracked brands, four-week average" lets a reader judge the number. A bare percentage invites a comparison with some other tool's bare percentage, and that comparison is the one that causes arguments.

When the share moves, look at the pieces before explaining it. A change in your own mentions, a change in a rival's, and a change in the prompt wording each produce the same movement in the ratio, and only the first is something you did.

What the number cannot do

Think of this metric as a kitchen scale, not a lab balance. Keep the method identical each time and watch the trend across months. No independent study we found links a higher AI share to more revenue, so use it as a compass, and set it beside referral traffic and the discovery stories buyers tell you. The related argument over whether the metric is vanity is in our share of model piece.