The short answer
Really Good GEO's citation-test protocol: test the same representative questions across named AI engines, record whether web search occurred and which domains were cited, then repeat the runs. Compare a baseline with results after a controlled page change. Repetition distinguishes signal from noise; a single result is not treated as a stable ranking.
Choose questions before seeing results
| Question type | Example | What it measures |
|---|---|---|
| Brand-free discovery | Which tools help distributed finance teams close the books? | Category consideration |
| Comparison | What are alternatives to [competitor] for a small team? | Competitive retrieval |
| Known brand | What is Acme and who is it for? | Entity retrieval |
Repeat the runs
A single result can change with retrieval behavior, model variation, location, timing, or product updates. Repeat each question enough times to label the brand consistently cited, sometimes cited, or absent.
Record the observable facts
- Engine and product tested
- Date and exact question
- Whether visible web search occurred
- Domains and pages cited
- Whether the brand appeared without being named
- Competitors repeatedly surfaced instead
Interpret cautiously
Visible citations are an outcome at a moment in time, not a permanent ranking. No-search responses and invisible retrieval should be labeled separately rather than counted as ordinary absence.