# How to Measure AI Share of Voice Across Engines

> Measure AI share of voice with a fixed prompt set, repeat runs and per-engine splits. A worked example, sample sizes and where the numbers mislead.

- Author: Sheetal Shakshi, Founder's Office, Nurdd
- Published: 2026-11-05
- Publisher: Noma by Nurdd (https://nomahq.app)
- Original page: https://nomahq.app/blog/measure-ai-share-of-voice

To measure AI share of voice, run the same set of unbranded buyer questions on each AI engine several times, count the answers that name your brand and each competitor, and divide your mentions by all tracked brand mentions. Report every engine on its own line, state the denominator, and ignore moves smaller than your sampling margin.

## Key takeaways

- Fix 20 to 30 unbranded prompts and do not edit them mid-quarter. Changing the prompts changes the metric.
- Run each prompt at least 5 times per engine. One run per prompt tells you almost nothing, because answers vary from run to run.
- Report share of voice per engine first, then a pooled figure. An average hides the engine where you are absent.
- State two denominators every time: share of all brand mentions, and mention rate across all answers. They answer different questions.
- Call a change real only if it exceeds the margin of error and holds across several weekly readings.
- Keep brand mentions and citations in separate columns. Citations churn faster than mentions.

## What does AI share of voice measure?

AI share of voice is the portion of brand mentions in AI answers that belong to you, within a defined prompt set, engine set and period. It is a ratio against competitors, not a count of your own mentions. Existing guides cover the definition; see our explainer on [what AI share of voice is and how it is calculated](https://nomahq.app/blog/what-is-ai-share-of-voice). This post covers the measurement workflow.

Published definitions differ slightly. [Single Grain, 2026](https://www.singlegrain.com/artificial-intelligence/measuring-share-of-voice-inside-ai-answer-engines/) describes it as the percentage of answers where your brand appears for a query set, engines and time period, optionally weighted by prominence. [HG Insights, 2026](https://hginsights.com/blog/ai-share-of-voice/) frames it as the percentage of AI-generated answers that mention or cite your brand compared with competitors in the category. Both agree on the inputs: a query set, a list of engines, a time window and a competitor set. Fix all four before you count anything.

## How do I measure my brand's share of voice in AI answers?

Use these seven steps. The first four decide whether the number means anything; the last three decide whether anyone trusts it.

1. **Name the competitor set.** Pick 3 to 6 brands a buyer would put on the same shortlist, including the ones that appear in AI answers even if you do not consider them rivals. Add alternate spellings and product names so a mention is not missed.
2. **Write the prompt set.** Use unbranded questions a buyer asks before choosing: category discovery, comparison, use-case fit and price. Leave out prompts that contain your own brand name, because they guarantee a mention and inflate share. Our guide on [which prompts to track](https://nomahq.app/blog/which-prompts-to-track) covers the five prompt types.
3. **Choose the engines.** List each engine you will report. Keep ChatGPT, Gemini, Perplexity and Claude as separate lines, and add Google AI Overviews or AI Mode if your buyers use them.
4. **Freeze the protocol.** Record the exact prompt wording, the engine and model, the country, whether web search was on, and the run date. Start each run in a fresh session so earlier answers do not shape the next.
5. **Run every prompt several times.** Five repeats per prompt per engine is a workable minimum for a weekly reading. Save the full answer text, not only the tick mark.
6. **Count and divide.** For each engine, count answers naming each brand, sum the mentions, and divide yours by the total. Also record your mention rate: answers naming you divided by all answers.
7. **Report with context.** Show the per-engine split, the pooled figure, the sample size and the margin. Add the sources cited in answers where you were absent, because that is where the work comes from.

For a first pass without a tool, the free [Noma brand audit](https://nomahq.app/brand-audit) shows how AI engines describe your brand today, which gives you a baseline to compare your manual counts against.

## How many prompts do I need to track AI share of voice?

For a stable category-level reading, 20 to 30 unbranded prompts, each run at least 5 times per engine, is a sensible starting point. That gives 100 to 150 answers per engine. Fewer than 10 prompts makes the figure depend on the wording of a handful of questions rather than on your category.

The margin of error shrinks with the number of answers. The table below uses the standard approximation for a proportion, 1.96 times the square root of p(1-p)/n, with a mention rate of 20%. It assumes independent answers, which repeated runs of the same prompt are not, so treat it as the best case. Real uncertainty is wider.

| Prompts | Runs per prompt | Answers per engine | Approximate margin at a 20% mention rate |
| --- | --- | --- | --- |
| 10 | 5 | 50 | about ±11 points |
| 30 | 5 | 150 | about ±6.4 points |
| 30 | 10 | 300 | about ±4.5 points |

The practical reading: with 30 prompts and 5 runs, a move from 20% to 24% on one engine is inside the noise. Wait for a second or third weekly reading in the same direction before you report it as a trend. For help sizing the questions in your category, see [how to estimate what people ask AI engines](https://nomahq.app/blog/ai-prompt-volume).

## Why do you need repeat runs?

AI answers are not stable lists. [Ecommerce Fastlane, 2026](https://ecommercefastlane.com/ai-assistants-change-top-pick-weekly/), summarising research by Rand Fishkin of SparkToro and Patrick O'Donnell of Gumshoe published in January 2026, reports that each of 12 brand-recommendation prompts was run 60 to 100 times per platform. The finding: there was less than a 1 in 100 chance that ChatGPT or Google's AI would return the same list of brands twice.

Citations are even less steady than mentions. [Ahrefs, October 2026](https://ahrefs.com/blog/how-stable-are-ai-citations/) followed about 37,500 citations across five AI tools for 20 days and found that 50 to 83% of citations were gone the next day, though about 75% of those came back within 20 days. Ahrefs' own advice is blunt: ["Don't judge AI visibility from one check." (Ahrefs, October 2026)](https://ahrefs.com/blog/how-stable-are-ai-citations/)

Two consequences follow for your workflow. First, a single manual check can swing you from "we are recommended" to "we are invisible" with no real change. Second, Ahrefs notes that brand mentions are more stable than citations, so lead your report with mentions and keep citations as a supporting column. Our post on [AI citation stability](https://nomahq.app/blog/ai-citation-stability) goes through the numbers in more detail.

## Why split share of voice by engine?

Engines disagree, so a pooled number can look healthy while one engine ignores you. [PPC Land, November 2026](https://ppc.land/chatgpt-and-google-differ-on-top-brand-in-35-of-buying-topics-bluefish-says/), reporting on Bluefish data, says ChatGPT and Google put a different brand at the top of their answers in 35% of 5,810 buying topics. The same pattern shows up week to week. In BusySeed's 2026 study, as reported by [Ecommerce Fastlane, 2026](https://ecommercefastlane.com/ai-assistants-change-top-pick-weekly/), the brand named first changed in 61.2% of weekly transitions on Gemini, 58.9% on ChatGPT and 49.1% on Claude.

Different engines also have different volatility, so they need different thresholds. A 5-point move on Claude and a 5-point move on Gemini do not carry the same weight. Report each engine against its own history. Our post on [why ChatGPT, Gemini, Claude and Perplexity disagree about your brand](https://nomahq.app/blog/why-engines-disagree) explains the mechanics.

## A worked example with two denominators

The numbers below are illustrative, not real data. A fictional brand, Brand A, is tracked against three competitors with 25 prompts run 5 times per engine, so 125 answers per engine. Each cell is the number of answers naming that brand.

| Engine | Brand A | Brand B | Brand C | Brand D | All mentions | Brand A share of voice | Brand A mention rate (of 125) |
| --- | --- | --- | --- | --- | --- | --- | --- |
| ChatGPT | 30 | 55 | 40 | 25 | 150 | 20.0% | 24.0% |
| Gemini | 12 | 50 | 45 | 33 | 140 | 8.6% | 9.6% |
| Perplexity | 38 | 40 | 30 | 22 | 130 | 29.2% | 30.4% |
| Claude | 25 | 45 | 35 | 20 | 125 | 20.0% | 20.0% |
| **Pooled** | 105 | 190 | 150 | 100 | 545 | 19.3% | 21.0% (of 500) |

Read it three ways. The pooled share of 19.3% looks like a mid-table result. The engine split shows the real story: Brand A is at 8.6% on Gemini, under a third of its Perplexity share (roughly 29% of it). And the two columns differ because an answer can name several brands, so share of voice (a slice of mentions) and mention rate (a slice of answers) are different measures. Label which one you are quoting.

## Where do share-of-voice figures mislead?

Check any figure, yours or a vendor's, against this table before you present it.

| Distortion | What it does | How to check or fix it |
| --- | --- | --- |
| Branded prompts in the set | Guarantees a mention and inflates your share | Keep only unbranded buyer questions in the headline metric |
| Single run per prompt | Reports noise as signal | Run 5 or more repeats and state the margin |
| Pooling engines | Hides the engine where you are absent | Report per engine, then pooled |
| Shifting competitor set | Share falls when you add a rival, with no real loss | Freeze the set; log changes as a break in the series |
| Edited prompts | Breaks comparison with earlier weeks | Version the prompt set and never edit in place |
| Counting mentions and citations as one | Mixes a stable measure with a volatile one | Two columns; lead with mentions |
| Position weighting with no stated formula | Makes two tools' figures incomparable | Ask for the formula; record raw counts too |
| Ignoring sentiment | Counts a warning as a recommendation | Read a sample of answers; track sentiment beside share |

One more caution: share of voice measures presence in answers, not sales. For the revenue link, see [AI search attribution](https://nomahq.app/blog/ai-search-attribution-guide).

## How often should you run and report it?

Run daily or weekly, report monthly, and read trends over at least four weekly points. Manual teams can run monthly, as [Single Grain, 2026](https://www.singlegrain.com/artificial-intelligence/measuring-share-of-voice-inside-ai-answer-engines/) suggests for early-stage setups, but a monthly manual check gives you 12 data points a year and a wide margin on each.

A simple cadence:

- **Weekly:** run the full set, log per-engine share and mention rate, flag any move above the margin.
- **Monthly:** report the four-week average per engine, the pooled figure, the top competitor movers and the sources cited where you were absent.
- **Quarterly:** review the prompt set. If you change it, start a new series and note the break.

Pair the quantitative report with a handful of full answers. A number of 19% means little to an executive; the sentence an engine used about you means a lot. Our [guide to reporting AEO to a CFO](https://nomahq.app/blog/report-aeo-to-a-cfo) shows how to frame it without overclaiming.

## Checklist before you publish the number

- Prompt set has 20 to 30 unbranded prompts, versioned and dated.
- Each prompt ran at least 5 times per engine in fresh sessions.
- Competitor set and alternate spellings are frozen.
- Per-engine figures sit above the pooled figure.
- Both denominators are labelled.
- Margin of error and sample size are stated.
- Mentions and citations are in separate columns.
- Raw answers are saved so anyone can audit a count.

## Where Noma fits

Noma by Nurdd is an answer engine optimisation platform that asks 13 AI engines, including ChatGPT, Gemini, Perplexity and Claude, the questions your buyers ask, every day. It reports visibility, position and sentiment per engine, shows each competitor's share of voice, and lists the sources each answer cites, so you can run this workflow without copying answers by hand. Every plan starts with a 3-day trial that includes 25 tracked prompts.

See it on your own brand: [book a demo](https://nomahq.app/demo) or [start the 3-day free trial](https://platform.nomahq.app/signup). A card is taken at signup so billing can start on day 3; cancel before then and nothing is charged. Plans are on the [pricing page](https://nomahq.app/pricing).

## How we researched this

We reviewed published AI share-of-voice guides from analytics vendors and agencies, Ahrefs' citation-stability study, coverage of SparkToro and Gumshoe's consistency research, BusySeed's weekly-change data as reported by a trade publication, and coverage of Bluefish's cross-engine data. Statistics are quoted as stated by each source and linked with their month of publication. Facts and prices were checked on 2026-11-04. The worked example is illustrative arithmetic, not measured data, and the margin-of-error figures use a standard textbook formula.

## What to do next

Write your prompt set this week, run it five times per engine, and record the per-engine split before you change any pages. Then compare the sources cited where you are missing against the gaps in your content. For a faster start, [book a demo](https://nomahq.app/demo) or [start the free trial](https://platform.nomahq.app/signup), and read our [step-by-step guide to tracking your brand across AI engines](https://nomahq.app/blog/track-your-brand-across-ai-engines) alongside it.