Measure AI visibility with a prompt panel: a fixed list of 20–50 real customer questions, run on a fixed schedule across each engine, logging whether you're mentioned, cited with a link, or recommended — and which competitors appear instead. Track each engine separately, because their source sets overlap far less than people assume. Then connect it to AI-referred sessions and conversions in analytics.
Two things businesses try first, both of which mislead:
Asking "do you know BAIDLABS?" This tests whether a name appears in training data. It tells you nothing about whether you'd be recommended to a customer, which is the question that matters commercially.
Checking once and drawing conclusions. AI answers are non-deterministic and retrieval runs live. The same question can name different businesses on Tuesday and Thursday. A single check is noise, not a measurement.
Write 20–50 questions a genuine prospect would ask, in their words, before they knew your name. Cover the range:
Freeze this list. Its value comes from being run unchanged over time — a panel you keep editing produces trends you can't interpret.
Score each result on three separate levels, because they mean different things:
| Level | Definition | What it tells you |
|---|---|---|
| Mentioned | Your name appears in the answer text | The engine knows you exist in this context |
| Cited | Your URL appears in the sources | Your content was actually retrieved and used |
| Recommended | You are named as a suggested option | The commercially meaningful outcome |
Also log sentiment (how you're characterised) and competitor set (who else appears). The competitor set is often the most actionable column — it shows you exactly who owns the answer you want.
Run the full panel against each engine you care about — ChatGPT, Perplexity, Google AI Overviews, Gemini, Copilot, Claude — on the same day each week or fortnight. Use a fresh session with no personalization or history, since prior conversation contaminates results.
Keep engines in separate columns. Studies have found the domains cited by different engines overlap surprisingly little, so a blended average hides the thing you need to see.
For each engine: share of voice = prompts where you are recommended ÷ total prompts. Do the same for your top three competitors. That single ratio, tracked over time, is the clearest signal of whether your programme is working.
Visibility that doesn't reach revenue is vanity. In your analytics, segment referral traffic from AI sources — chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and similar. Expect the volume to look trivially small and the conversion rate to look implausibly high; that's the documented pattern, not an error.
For most businesses running this for the first time, the honest starting picture is: mentioned on a handful of prompts, cited on almost none, recommended on zero. That is normal and it is not a failure — it's the baseline that makes every later improvement provable.
A spreadsheet run manually is entirely adequate to start, and forces you to actually read the answers — which is where the insight lives. Commercial AI-visibility platforms automate the runs and are worth it once you're tracking many prompts across many engines. Start manual, automate when the manual version becomes the bottleneck.
Book a free GEO audit: where you appear across 6+ AI engines today, the prompts competitors own, and a prioritized 90-day plan you keep either way.