Why the obvious test fails

"Do you know [brand]?" triggers politeness, not knowledge. Models complete patterns: for a thin-source brand they generalize from the name, the category and similar companies — producing fluent, plausible, wrong answers. A test that can't distinguish knowledge from confabulation is worse than no test.

The 12-prompt protocol

  1. Identity (3 prompts). "What is [brand]?" / "What does [brand] sell?" / "Where is [brand] based?" Scores fact accuracy of the basic entity card.
  2. Category (3). "Best [category] providers" / "Recommend a [category] company for [use case]" / "Alternatives to [named competitor]". Scores Visibility — are you named unprompted?
  3. Comparison (2). "[Brand] vs [competitor]" / "Pros and cons of [brand]". Scores positioning and whether your differentiators registered.
  4. Reputation (2). "Is [brand] legit/reliable?" / "[Brand] reviews and complaints". Scores Brand Cognition where it hurts.
  5. Boundary (2). "What is [brand]'s pricing?" / "What products does [brand] not offer?" Boundary prompts expose hallucination — confident wrong answers here mean your source graph is sparse.

Scoring rules

  • Fresh sessions, 3 runs each, majority rules. Never screenshot one answer.
  • Log claims, not vibes. Extract each factual statement and mark it correct / wrong / stale / unverifiable.
  • Note the sources panel. Which URLs does ChatGPT cite on search-enabled answers? That inventory is your Citation Rate baseline.
  • Repeat per engine. ChatGPT's answer tells you little about Gemini's. Run the same set everywhere.

The protocol above is exactly what the GEO Checker automates — at 30 prompts with analyst review — if you'd rather have it done than do it.

Frequently asked questions

How often should I re-run the brand test?
Monthly for active brands, and immediately after any correction campaign. Answers drift as indexes refresh and models update; a January answer is not evidence about November.
ChatGPT described my competitor when asked about me. What happened?
Entity collision: the engine resolved your name to a better-documented entity. The fix is entity engineering — consistent naming, Wikidata/knowledge-panel inputs, and corroborated descriptions across independent sources.