2026-09-15 — GEO · Brand Monitoring

AI share of voice: a measurement framework

“AI visibility” is useless as a vanity metric. What you can operationalize is a small set of ratios computed from repeated, geo-targeted answer collection. Here’s the framework.

The units

Every monitored prompt produces one answer. From that answer you extract three boolean/ordinal observations:

• Mentioned: does result.text (or entities[]) name your brand?
• Cited: does your domain appear in result.sources[]? At which position?
• Recommended: for commercial prompts, are you in the recommendation set (entities[], shoppingCards[], inlineProducts[])?

The metrics

Mention rate — the share of prompts in a set where you’re named:

mention_rate = prompts where brand mentioned / total prompts

Run per engine, per market, per prompt cluster (category / comparison / problem). A single blended number hides everything.

Citation share — your share of all citations in a prompt set:

citation_share = your citations / all citations in sources[] across the set

This is the closest AI analogue of SERP share of voice: it measures how much of the evidence layer you own.

Fan-out coverage — the share of fan-out queries (searchQueries[]) where you rank in classic search. It predicts future citation share: you can’t be cited by searches you don’t rank for.

Answer rank — when the answer is a list (“best X are A, B, C”), your position in entities[] order. First-mentioned brands get read; fourth ones get skimmed.

Sample sizes and cadence

Answers vary run to run, so treat each run as a sample. Rough guide: 30+ runs per prompt-market-engine cell before trusting a ratio; weekly cadence for tracking, daily only during launches or incidents. How often AI answers change covers setting the schedule.

Implementing it

Submit each prompt × market as an async task with a dated idempotencyKey, collect results by webhook, and store {prompt, market, engine, mentioned, cited_position, sources, entities} rows. The whole pipeline is one batch call and one webhook receiver — the batches guide has the mechanics, and brand-mention tracking shows the metric math on real fields.

Report trends, not snapshots: mention rate week over week, citation share by engine, fan-out coverage by cluster. Those three lines are the framework.

Keep reading