Probe methodology
Page score vs probe score, and why recognition / mention / recommendation / citation must not become one AI-visibility number
Page scores answer whether models can read you. Probes answer whether models will say you. Public pages: methodology, cases (protocol examples). How-to: AI Answer Probes.
Page score vs probe score
| Input (page GEO) | Output (answer probes) | |
|---|---|---|
| Question | Can a model read and extract this page? | After asking, will it mention, recommend, or cite you? |
| Signals | Princeton GEO heuristics, Schema, robots / llms.txt | Multi-model raw answers + structured judgments |
| Where | GEO Optimization → Overview / AI visibility | GEO Optimization → Answer Probe |
Do not weight the two. SoloKit keeps input optimization and adds output probes. NiubiGEO focuses on named output audits. Profound / Peec are closer to consumer-UI share.
Four rates, four questions
| Rate | Protocol | Judgment |
|---|---|---|
| Recognition | D (domain-only prompt) | Can the model recap the named site / brand? |
| Mention | K (keyword-only prompt) | Brand name or official host appears |
| Recommendation | K | Structured relation=positive: explicit preferred option |
| Citation | K | Answer URLs hit the official domain |
Being able to describe a named domain is not unprompted mention. Named without a recommendation must keep recommendation at 0.
Citation vs content writeback
- Citation rate: do links in the answer land on the official host?
- Published content hits: do those URLs equal a distribution
published_url(exact / canonical)?
Only the latter means this shipped article was used.
Same-condition retest
Freeze aliases, keywords, and competitors first. Compare in percentage points (pp) and keep the previous original answers. After you change the word list, do not compare against old rounds.
Not one AI-visibility score
Recognition / mention / recommendation / citation have different denominators and questions. A blended score hides “recognized but not recommended” or “mentioned but not citing your URL”.
D domains and K terms on the public cases page are protocol examples, not live rankings. Product-line alignment: GEO Insights and 榆关 / Pen.