How to measure your brand's share of voice in AI answers

Why rank tracking doesn’t transfer

A rank tracker works because a results page is a stable, ordered list. An AI answer is neither: it’s generated fresh, it names three to six brands in prose, and it changes between sessions, models and phrasings. Checking it once tells you what one engine said one time — which is trivia, not measurement.

The fix is the same one search measurement made twenty years ago: stop looking at individual results and start sampling systematically.

The framework

1. Fix the prompt set

Write down the questions that define your category — the way buyers actually phrase them. “Best specialty coffee brands in Dubai”, not “specialty coffee UAE”. Separate sets per market and per language: answers in Arabic are not translations of the English ones; they’re different answers, often with different brands in them — the Arabic AI search guide shows how wide that gap runs in the GCC.

2. Run it on real engines, on a schedule

Ask every prompt on every engine you care about — ChatGPT, Gemini, Claude, AI Overviews — at a fixed cadence. Scheduled runs are what turn anecdotes into a time series; the cadence matters more than the frequency.

3. Extract three metrics per answer

• Mention — is the brand named at all? Aggregated, this is share of voice: the % of tracked answers that include you.
• Answer rank — first brand named, or fifth? Being the lead recommendation is worth far more than trailing a list.
• Citations — when the engine links sources, is your domain one of them? This is the metric your content team can move fastest.

4. Benchmark competitors from the answers, not a hand-picked list

Score every brand the engines name, not just the rivals you already watch. The most useful output is often a competitor you didn’t know was winning your prompts — and the moment a new name starts appearing is exactly when you want an alert, not a quarterly surprise. The full workflow is in competitor analysis in AI search.

5. Close the loop

Low share of voice on a prompt cluster is a content brief: it tells you which question to answer, in which market, in which language. Publish, wait a cycle, re-measure. That loop is GEO — the strategy behind it is in our GEO guide.

Pitfalls we see constantly

• Spot-checking. One flattering ChatGPT answer in a boardroom slide is not data. Sample, or don’t bother — the ChatGPT-specific loop is in tracking brand mentions in ChatGPT.
• Single-engine tunnel vision. Engines disagree. A brand can lead on Claude and be invisible on Gemini; each engine is its own audience with its own number.
• Ignoring language. If you sell in the Gulf, your Arabic share of voice is a separate battle — and usually the less contested one.
• Trusting empty answers. Pipelines that silently record a failed engine response as “zero mentions” corrupt the trend. Failed runs must be retried or excluded, never counted.

Start with a baseline

Before building anything, get your number. The free AI-visibility audit runs a first pass on your brand — real engine answers, human-reviewed — so you know your starting share of voice before you invest a dirham in moving it.