1. Define the decision and sample
Choose the question the measurement should inform: whether a source is cited, whether a service is portrayed accurately, or whether an entity is shortlisted for a particular need. Freeze a written prompt panel before the comparison. Keep branded identity questions separate from unbranded buying questions.
Select products, modes, locale and collection windows deliberately. Record repeated runs rather than treating one answer as a stable rank. Repeated answers from one product may be dependent observations; more runs do not automatically create a representative audience sample.
2. Retain a record for every attempt
| Field group | What to retain |
|---|---|
| Identity | Run ID, prompt ID and the exact prompt text. |
| Conditions | Product, visible model label or “unknown,” search mode, locale, account/session context and prior conversation. |
| Time | Actual timestamp with timezone. Never the literal word “now.” |
| Receipt | Full answer text, ordered citation URLs and an accessible capture reference where sharing is permitted. |
| Collection status | Complete, blocked, failed or incomplete, plus the reason. |
| Assessment | Mention, citation, recommendation, sentiment, claim support and reviewer notes. |
Protect personal data and comply with product access terms. Keep a redaction note when a public receipt differs from the privately retained original. Do not label an inaccessible private record as publicly reproducible.
3. Score observations consistently
Define a citation as a visible source reference and a recommendation as inclusion suitable for the prompt’s stated need. Keep critical mentions separate from positive recommendations. Inspect the linked passage before scoring whether it supports the answer. See AI citations for the distinctions.
Use the same written rubric before and after a change. Have ambiguous judgments reviewed, and retain the original output so another reviewer can disagree with a score without losing the evidence.
Rate definition: qualifying complete responses divided by eligible complete responses. Report that denominator, the number of collection failures and the sampling window alongside the rate. A blocked run is missing, not evidence that an entity was absent.
4. Interpret an illustrative record
Hypothetical example, not collected data. A comparison answer names a supplier and cites its maintenance guide, but recommends another product for the reader’s budget. The first supplier receives a mention and a citation, not a positive recommendation. If the source link cannot be opened, support remains unverifiable rather than automatically passing.
A later run blocked by a login screen receives a collection-status entry and no presence score. Keeping that attempted run in the ledger prevents a report from silently excluding difficult observations.
5. Compare without inventing causality
Document exactly what changed, when it changed and when collection occurred. Compare like-for-like prompt segments and conditions. Consider untreated pages or other stable reference observations, and report uncertainty and potential product changes.
A before-and-after difference alone does not isolate the effect of a content edit. Search results, source freshness, session context and product behavior can change too. The GEO paper is research under specified benchmark conditions; it does not validate an unrelated commercial trial. Source: GEO paper.
6. Keep visibility and business outcomes separate
Google includes its AI feature traffic in the overall Web search type in Search Console. It also suggests using analytics to track conversions and time on site. Source: Google. Aggregate search reporting should not be relabeled as the citation rate of a frozen prompt panel.
When genuine analytics are available, report visits, qualified inquiries and other defined outcomes as separate measures with their own attribution limits. No analytics or engine results are supplied by this method page.
Cody Schuldt works on answer-engine measurement through the Search-to-Answer Visibility System, which he created to connect search discovery and AI-answer visibility with an explicit measurement plan. This is first-party operating context, not a claim that this protocol has produced a measured improvement. Source: original framework. The GEO and AEO methodology hub places this recording protocol within the wider workflow.