GeoPeak

How GeoPeak measures AI visibility

geopeak.ai · Last updated: Nov 19, 2026

In short

AI answers change from one run to the next. A visibility number without a margin of error says little, and a change without a significance test can be noise. GeoPeak shows both, and says "not significant" when that is the honest answer.

What we collect

  • Engines. ChatGPT, Perplexity, Gemini and Google AI Mode, and Claude on the plans that include it.
  • Answers without a web search. Claude decides for each question whether to search the web. When it answers from memory, the answer is kept, since that is what a user would see, and flagged "without web search": it has no sources, and it is counted separately in the CSV export.
  • Several answers per question. At least 10 answers per prompt and per engine at each run (3 for Claude). This depth is the same on every plan: plans differ by the number of prompts, engines, markets and domains, never by the method.
  • Cadence. Weekly at most. Asking more often adds noise and cost, not information.
  • How. Through the real interface when available, otherwise through the official API with web search enabled. The method used is stored with each answer. A model without web search is never presented as a measurement.
  • Markets. Each run uses the country and language of the markets you choose.
  • Failed answers. A collection error is recorded as an error. It is never counted as "brand not cited".

How an answer is read

  • Mention. Your brand names and aliases are matched as whole words, ignoring accents and case. Your domain is matched exactly.
  • Rank. Position in the answer's list when there is one, otherwise order of mention.
  • Sources. Pages the engine cites, and whether one of them is yours.
  • Versioned analysis. Every answer is kept as collected. When the reading rules change, the version number changes and history can be recomputed.

Visibility and its margin of error

  • Visibility is the share of valid answers that mention your brand over the period.
  • The prompt is the unit. Answers to the same prompt are correlated, so the 95% interval is computed by resampling whole prompts (cluster bootstrap), never single answers.
  • Change versus the previous period is computed on the prompts present in both periods. It is marked significant only when its 95% interval excludes zero. Otherwise it is marked not significant.
  • Below 5 prompts in a period, no interval is shown.

Measuring the effect of an action

The experiment journal tests whether a change you made (a page, a citation, a press release) moved your visibility.

  • Declared before. The action, the metric, the prompts it targets and the expected effect are recorded, and timestamped by the server, before results exist.
  • Control prompts. Prompts the action should not affect serve as a comparison group, frozen at go-live. The effect is the change in targeted prompts minus the change in control prompts (difference in differences).
  • Windows. 28 days before and after, plus a lag (7 days by default) for engines to pick up the change.
  • Tests. Permutation tests on prompts, 95% bootstrap intervals, and Benjamini-Hochberg correction (q = 0.10) when several metrics are tested.
  • Size before launch. GeoPeak computes the smallest effect the experiment can detect from your prompts and their variability. Minimums: 10 targeted and 10 control prompts, or 20 prompts without controls (before/after only, reported as weaker evidence).
  • Engine weights. Results across engines are weighted by market share (StatCounter, August 2026) for the US, Canada and France, and equally elsewhere. The version used is shown in each verdict.

Limits we state

  • Answers seen by a real person can differ (history, location, account). We measure a standardized, logged-out view.
  • A significant result shows an association in time, measured carefully. It is not proof that nothing else changed.
  • Engines change their models without notice. A model change can move every prompt at once; control prompts absorb part of it, not all.
← Back to GeoPeak