The method, in full.

Frequency, not ranking.

The stable, meaningful quantity is not where a product appears in an answer but how often it is recommended for a specific buying need. List position is close to random across repeats; appearance frequency is stable. So the core metric is frequency by intent, and cross-engine agreement is reported as a separate number.

Four independent, web-grounded engines.

We measure across ChatGPT, Perplexity, Gemini and Claude in modes that browse the web and return citations. Parametric (no-web) answers are treated as internal exploration only — never as evidence for a badge.

• Neutral buying questions, many paraphrases, no brand names in the prompt.
• Observations are spaced over time, not taken in a single burst.

It beats chance, or it isn't a claim.

Each intent has its own null baseline — the frequency you'd expect by chance given how many brands plausibly populate that need. A product's observed frequency must clear that baseline, and we attach a conservative bootstrap confidence interval so the separation is statistical, not anecdotal.

Worked example (our own measurement): for office / dressy the null sits near 24%; the example record clears it at 77% across 4/4 engines.

No echo chambers.

A signal that comes from one repeated citation is fragile. We measure the concentration of citing domains (an HHI over the cited sources); low diversity is flagged as a weakness and can disqualify a record. This stops a single affiliate page from manufacturing a recommendation. Known limit: Gemini currently returns citations as opaque redirect URLs, so its source-diversity score is not yet reliable — we treat Gemini's grounding as present but its domains as unresolved, and never rest a claim on Gemini alone.

Signal that drifts is revoked, not kept.

• Freshness: evidence supporting a record must be within a 90-day window.
• Revalidation: records are recomputed on a cadence; stability over consecutive rounds is required.
• Revocation & expiry: if a product falls below its baseline or loses cross-engine consensus, the record is automatically and auditably revoked and the store badge stops showing. If evidence merely ages out of the 90-day window without a contradicting round, the badge shows its date — a dated fact — rather than a current claim, and stops showing only past the category horizon; it never displays a negative mark.

All must hold at once.

A (product × intent) pair is eligible only if every criterion below holds on fresh evidence. Thresholds are marked [provisional] and are recalibrated as the program matures.

You cannot pay to change any of these. Payment covers measurement, display and upkeep — never eligibility.

Reproducible from the record.

Every eligibility decision is reproducible from the per-query log: given a round id, an outside auditor can recompute C1–C7 and confirm the status. The method version travels with the badge.

The limits, stated.

Status: methodology v0.2.2 · grounded-only scope · numeric thresholds provisional, recalibrated in phase 1. Example figures are RecommendedByAI's own measured data for Carets · office.