AI mentions, citations and recommendations: an evidence taxonomy

Assign the run a study ID and observation ID. Preserve the exact prompt, prompt family, language, target market, test location, platform, product surface, account or signed-out state, date and time. If the interface exposes a model or product version, retain that too.

State the classification rules before reviewing responses. Define what counts as the entity, which domains belong to it, how URL redirects are normalised, what wording qualifies as a recommendation and how ambiguous names are handled. A rule changed after seeing the answer creates a new protocol version.

Match the name against context such as the verified domain, founder, product, location, service description and organisation identity. A common name or acronym without identifying context remains an unresolved match and should not enter the mention numerator.

Keep aliases and former names in a controlled entity registry. Do not treat a partner, reseller, review site or similarly named company as the target. If the answer merges two entities, classify the mention and the factual error separately rather than silently choosing the favourable interpretation.

Save the visible response from start to finish, not a crop around the brand. Retain source chips, footnotes, expanded source panels, link destinations, position, surrounding sentence, follow-up state and any warning that changes how the answer should be read.

Resolve each link through redirects to its final public URL and record the canonical domain separately. A search result, cached copy, syndicated duplicate and original source are different records. Screenshots help with presentation evidence; selectable text and URL logs make later review possible.

Open the cited page and identify the passage, table, dataset or structured fact that could support the answer. Mark the relationship as supported, partially supported, contradicted, inaccessible or unclear. A source link displayed in the interface is still a citation observation even when support fails, but it is not a supporting citation.

Separate source selection from answer absorption. A page may be selected as a citation while its distinctive information never appears in the generated wording; conversely, a claim can resemble a source without a visible link. The public interface rarely proves the hidden causal path, so report the observable relationship only.

Check material facts: name, ownership, location, services, audience, price if public, dates, credentials, limits and the claim's original scope. Mark portrayal as accurate, partially accurate, materially wrong or unverifiable. Positive tone cannot repair a false fact.

Then read the commercial role. Is the entity merely named, included among options, preferred for a criterion, ranked, warned against or given as the next action? Preserve the actual criterion and caveat. Sentiment, comparative inclusion and recommendation are independent fields, not one label.

For important panels, have a second reviewer classify a masked record using the same rubric. Log disagreements such as mention versus namesake, citation versus search link, neutral list versus recommendation, and supported versus merely related source. Resolve them without editing the original evidence.

Publish counts with denominators, missing observations and agreement notes. Preserve raw records under appropriate access controls. A later audit should be able to reconstruct why an item entered a numerator without relying on the analyst's memory or a dashboard vendor's private score.