Distinguish mentions, citations and recommendations

What counts as each observation?

Read one answer before looking at totals. A brand can be named as a warning, cited as the publisher of a report or recommended as a supplier. Those are different roles, and a single answer can contain more than one.

Use a written rule for each field:

Keep citations to your own domain separate from citations to pages about your company on third-party domains. A newspaper article mentioning your brand is a newspaper citation, even if the company is its subject. A bare brand name without a link is not a citation under this definition.

How do you avoid misleading counts?

Decide whether you need answer-level presence or total occurrences. For answer-level presence, a brand repeated five times in one answer counts once. For occurrence counts, retain all repetitions but explain what that total represents. Do not compare one answer’s mention count with a whole scan’s citation count.

An illustrative answer says that Alder and Birch offer booking software, recommends Birch for a museum and cites Alder’s research article. In that answer, both brands are mentioned, Birch is recommended and Alder’s domain is cited. This does not prove that the system trusts Alder’s product less or that Birch came from training memory.

Print the number of eligible answers beside every percentage. Deduplicate URLs using a recorded rule and preserve original links. If one answer fails or contains no usable response, show that case separately so a clean percentage does not hide missing data.

Does a citation prove the claim?

Open the page and check the specific assertion. A citation may support only one part of a sentence, refer to an older version or fail to substantiate the surrounding claim. Citation-verifiability research distinguishes whether claims are cited from whether the citations actually support them.

The visible citation list also need not be a complete record of everything accessed internally. The public sampling-completeness preprint describes how observed source coverage changes with collection horizon. A short list cannot establish the full information history of a model or answer.

How do you inspect this in Rankfor?

Open Answer Trail in the relevant project, choose the available model configuration and enter the question. After the analysis completes, review the answer and surfaced sources together. Record the exact prompt, engine, date and optional region. Answer Trail is a separate analysis; do not assume every Index measurement already contains the same trail.

Use the coding sheet to identify the next task. An inaccurate mention calls for fact checking. An unsupported citation needs a claim-to-source review. An absent recommendation calls for examining the question, comparison criteria and sample. None of those observations alone establishes revenue, market share or the cause of the model’s choice.