1. Freeze the audit scope
Write the research question, public surface, route, locale, account or session rule, observation window and question list before capture. Define what counts as a visible source: inline citation, linked source panel, expandable source or another interface element.
State what the audit cannot see. It does not recover private query fan-out, retrieval candidates, ranking weights or model reasoning.
2. Capture the public record
Assign an observation ID, record an offset timestamp, preserve the exact submitted question and capture the visible answer state. Store every visible source URL as shown, with its location or visibility tier when that distinction is reliable.
When the answer or source panel is unavailable, record the missing reason. Do not rerun only failed cases until a preferred result appears unless the retry rule was defined in advance.
3. Normalize and verify without erasing provenance
Derive normalized URLs and registered domains while preserving the original values. Document parameter removal, redirect handling and canonical selection. Then open a bounded, stated sample of cited pages to check status, title, topic match and whether the page is actually accessible.
A cited URL is not automatically evidence for the answer’s claim. The audit can distinguish source presence from source support, but claim-level verification needs a separate coding guide and reviewer process.
4. Report coverage before conclusions
Publish scheduled and completed observations, answers with visible citations, total citation occurrences, unique URLs and domains, missing reasons and the verification sample size. Then report source recurrence, concentration or category mix if those measures were preregistered.
Include the source list or a rights-safe representation when permitted. Name capture limitations, possible personalization, interface changes and any pages that could not be verified.
- Keep raw capture and analysis tables separate.
- Use one rule for every eligible observation.
- Show denominators beside percentages.
- Treat common ownership and conflicts explicitly.
- Publish corrections without deleting the old record.
Source notes
These sources support the definitions, standards or project boundaries named in this reference. They do not prove that a public observation dataset exists.
- portfolio-dossierCanonical ai-fanout.com domain dossierOwner record
Confirmed ownership, accepted public Evidence Lab purpose, named Research Owner, indexable website launch and separately gated provider research.
- google-query-fanoutAI features and your websiteOpen
Google describes query fan-out publicly without exposing a general private-query inspection interface.
- w3c-prov-oPROV-O: The PROV OntologyOpen
Provides provenance concepts for entities, activities, agents, derivations, sources and versions.
- rfc-3339RFC 3339: Date and Time on the InternetOpen
Supports an interoperable timestamp representation tied to UTC.
- fair-principlesThe FAIR Data PrinciplesOpen
Supports reusable research data with metadata, provenance and clear usage licenses.