Why screenshots lie

AI answers are probabilistic — the same prompt can name you on one run and omit you on the next. They are also personalized and session-dependent. A screenshot proves an answer happened once; it says nothing about appearance rate. Vendors selling GEO on screenshot evidence are selling you weather and calling it climate.

The minimum viable protocol

  1. 30–50 graded prompts covering your category's question matrix (20% head / 40% mid / 40% long-tail), including reputation and comparison prompts.
  2. Three fresh-session runs per prompt per engine; majority scoring.
  3. Log four things per answer: were you named; which URLs were cited; what factual claims were made about you; what was the sentiment.
  4. Report per engine first, composite second. Trends monthly; alerts on negative flips immediately.

Tooling reality

AI-visibility trackers are maturing but diverge wildly in prompt selection and sampling rigor — insist on seeing the prompt set any score is computed from. The full methodology chapter publishes our protocol openly; the GEO Checker runs a lightweight version free, and the monitoring service runs it on a daily/weekly cadence with alerts.

Reporting that survives a CFO

Report movement against baseline (points gained per dimension per engine), not absolute scores without context. Pair every number with the prompt set version and sample date. The moment AI-visibility reporting is more rigorous than the SEO reporting it sits beside, budget conversations get easy.

Frequently asked questions

How do I track AI visibility without expensive tools?
The protocol matters more than the platform: a spreadsheet, 30 prompts, 5 engines, 3 runs each, done manually once a month, beats an expensive dashboard with an unauditable score. Automate only after the manual version proves which questions matter.
What's a realistic visibility improvement timeline?
Technical-access fixes can move Perplexity and live-retrieval engines within weeks. Index-dependent engines (AI Overviews, Copilot) follow index refresh. Training-memory effects take a model generation. Report on 90-day cycles; expect the first measurable movement in 30–60 days.