Answers / Measurement

How to measure AI visibility (metrics that mean something)

By Hoss · Published July 13, 2026 · Sources linked at the end

What are the three metrics that matter?

Mention, recommendation, and citation — each defined precisely enough that two people scoring the same answer get the same result. These mirror the definitions in our published methodology.

The gap between the three is where the diagnosis lives. Mentioned but never recommended is a positioning problem. Recommended but never cited means the engine is learning about you entirely from third parties. Not mentioned at all is an existence problem — start with the self-check.

Why do you have to run each prompt more than once?

Because the answers are stochastic. Ask an engine the same question twice and you can get different businesses, different sources, different phrasing — nothing about the market changed between the runs. A single check is one sample from a distribution: it can show you present on Monday and absent on Tuesday while your actual visibility held perfectly still. Industry guidance on manual audits says the same thing — run each query multiple times and record the pattern, not the run [1]. Repetition is what turns "ChatGPT mentioned us once" into "we're mentioned in 40% of runs" — a rate you can compare next month.

Why report per engine instead of one blended score?

Because the engines barely read the same web. One study of 127,198 citations across five engines found only 2.7% of sources were cited by all five, while roughly seven in ten were cited by just one [2]. Visibility is not one market — it's several, sharing a label. Strong ChatGPT numbers say almost nothing about Gemini, and a blended "AI visibility score" averages away the only actionable fact in the data: which engine is sending you buyers and which one has never heard of you. Report each engine as its own row, every month.

How do you design the prompt suite?

• Use buyer language, not marketing language. The suite should be the questions real buyers ask — "best [service] in [city]", "is [product category] worth it" — not your brand terms. If you're not sure what buyers ask, their words are recoverable from reviews, sales calls, and community threads.
• Freeze the prompts. Month-over-month comparison only means something if the questions are identical. Version the suite; when you must change it, note the change and expect a discontinuity in the trend line.
• Verticalize. A med spa's suite and a SaaS vendor's suite share a structure, not questions. Generic prompts measure a market you're not in.
• Cover the engines your buyers use, not just ChatGPT — and expect the results to differ sharply between them [2].

Do you need a tool, or can you measure manually?

Either works; they answer to the same method. A tool category now exists for exactly this — Profound, Otterly.AI, and Peec AI are the names industry roundups put on it [1], and platforms like SE Ranking have productized ChatGPT visibility tracking: mentions, how you're described, links, and competitors [3]. Tools buy you scale and consistency. Manually, the process is: fresh chat per query, record whether you're mentioned, your position, how you're described, and which competitors appear, then repeat two to three times per query [1]. At small scale that's an afternoon a month — our self-check guide walks through it step by step. What neither buys you is different truth: a tool running an unfrozen, one-run, single-engine suite produces the same noise a human would.

What are the common measurement mistakes?

• One-run checks. A single ask per prompt samples a distribution once and calls it a measurement. The resulting "trend" is mostly variance.
• Changing prompts between months. New questions each month means every data point is month one. Nothing is comparable to anything.
• Testing only one engine. Given a 2.7% source overlap across five engines [2], a ChatGPT-only number silently claims the other engines behave the same. They don't.
• Using search rankings as a proxy. Ranking well and being named by AI assistants are different outcomes with different causes — sites that rank well are routinely absent from AI answers [4]. If rankings measured AI visibility, none of this would need measuring.