GEO has an evidence problem: answers are non-deterministic, and most “visibility scores” are reported without a denominator. The way out is to keep three small records and be boring about them.

1. A prompt set you wrote down

Pick 15–30 questions a real buyer or reader would type. Write them exactly as they’d be typed, not as keywords. Keep the list in a file, version it, and do not add questions after seeing results — that is how a set stops measuring anything.

prompts/2026-11.txt
what is generative engine optimization
how do I get cited by chatgpt
does llms.txt help with ai search
geo vs seo for a small site

2. A citation log

For each prompt, record: engine, date, whether your domain appears, and which URL. One row per observation, in a CSV. Do not average across engines — they retrieve differently.

date engine prompt cited url
2026-11-18 example-engine what is geo yes /guides/what-is-geo/

The value is in the trend over weeks, not in a single run. Answers vary between runs; a single missing citation means little, a month of missing citations means something.

3. A crawl log

This is the measurement almost everyone skips, and the one that explains the other two. Keep 30 days of access logs for your host and ask three questions:

  • Which AI crawlers reached the site at all?
  • Which paths did they fetch, and how often?
  • Did any of them receive a non-200 status?

If a crawler never arrives, no amount of rewriting will produce citations. Fix access first.

Metrics that look fine and mean little

  • “AI visibility score” without a published method and sample size.
  • A single prompt that now cites you — one run is noise.
  • Impressions in a search console — those are classic search, not answers.
  • Referral traffic from AI assistants — useful commercially, useless as a proxy for being cited: most citations are read without a click.

A cadence that fits in an hour a month

  1. Re-run the prompt set (same engines, same phrasing), append to the citation log.
  2. Diff the crawl log against last month: new crawlers, new 4xx/5xx, paths that stopped being fetched.
  3. For any prompt where a competitor is cited and you are not, open the cited URL and compare one thing only: how directly its first sentence answers the question.

That last step is the whole method. The rest is bookkeeping.