Measuring AI share of voice without fooling yourself
Pick five to ten fixed category prompts, run them against each engine on the same schedule, and record whether you were named and how you were described. Never compare two ad-hoc screenshots: model answers vary between runs and shift when engines re-crawl or retrain, independent of anything you changed.
Fix the prompts before you measure
Write prompts a buyer would actually type, without your brand name in them. If the brand is in the prompt, you are measuring recall of a name you supplied, which tells you nothing.
Keep the set stable for at least two quarters. Changing prompts resets your baseline.
Record three things per run
Mentioned or not. Sentiment of the mention. Which competitors appeared alongside you. The third is the most actionable — it tells you which sources the engine trusts for your category, and those sources are your outreach list.
Expect noise, and say so
The same prompt on the same day can return different companies. Treat any single run as one data point in a series. We report a mention rate across a prompt set, never a single answer, and we label which engine produced it.
Want the version of this measured against your own site? Run a free audit.