Trace the finding to an original answer
Open Chats or follow a supporting-passage link from Overview. Read the full answer, not just its preview. Confirm the exact buyer question, source label, collection time, market and language. A short excerpt may omit a qualification or a stronger recommendation elsewhere in the answer.
Check how the business is named: exact name, alias or domain. Then examine the passage that drives the classification. If a named product has several meanings, resolve the ambiguity before repeating the finding in a client report. Retain the answer’s identifier or report link so another reviewer can inspect the same record.
Move from domains to exact pages
- Use Domains to find websites appearing in the retained source relationships. Read whether a count represents citing answers, cited pages or retrieval occurrences.
- Open URLs to identify the exact page, title and related answer. A large domain can contain many unrelated topics.
- Inspect the actual public page where appropriate. Verify the author, date, claim and context before turning its appearance into a content recommendation.
- Distinguish a retrieved page from a cited page. A provider may explore a page without including it in the final answer.
- Preserve response-level attribution when that is all the provider supplied. A source list is not automatic proof that each URL supports every sentence.
Use gaps to choose an investigation
Gap analysis highlights sources and competing brands within the selected evidence. A domain cited when your own website is absent is a useful research lead. It does not establish that the publisher owes you coverage, that a backlink will produce a citation or that copying the page will improve performance.
Read the cited page’s job: comparison, definition, product documentation, review, data or practical advice. Ask whether your own material gives a reader equally clear evidence for the relevant decision. If not, the action should identify the missing explanation or proof, rather than simply target a publisher’s name.
Interpret Fanouts carefully
Fanouts shows research-query information returned by supported sources. It can reveal adjacent questions and terminology that deserve investigation. Coverage varies by source, and a missing fanout record is not proof that a model did no research. Do not turn these records into claims about hidden reasoning or private prompt demand.
Useful follow-up questions can be saved for a future run. Keep their origin and purpose in the register. If you change the wording, label it a new test instead of appending it to an unchanged trend.
Create an evidence packet
For each proposed action, retain the original question and answer, exact supporting passage, cited URL, collection date, selected scope and proposed change. Include what remains uncertain. A content writer needs the question and evidence; a developer needs the affected URL and technical acceptance criteria; a stakeholder needs the business decision and a concise explanation.
If the finding depends on one source or a very small sample, say so. Compare another relevant question or repeat the same scope when the decision justifies using another credit. More data is useful only if it answers the uncertainty you identified.
Your next step
Build one evidence packet that another person can inspect without asking you where the conclusion came from.
Continue to the next guide → · All Academy lessons