A domain in the sources does not prove the claim

A large site can contain many sections, dates, and conflicting formulations. When an answer cites a domain, the reader still does not know which page was used or whether it supports the stated conclusion.

Even the exact page may be only background: it mentions the topic but does not support the number, condition, or limitation. A domain is therefore a starting point for checking, not proof by itself.

First rewrite the answer as short statements that can each be supported or challenged separately. Do not keep a fact, causal explanation, and recommendation in one sentence: they may need different sources and different levels of confidence.

For each claim, record subject, action, object, date, and condition. This quickly reveals what was lost in paraphrase and where the system added meaning absent from the page.

Open the page and locate the sentence or table that could support the answer. Record the exact fragment, section heading, update date, and surrounding context. A headline often promises more than the body confirms.

If the page contains only a related topic, mark it “unclear,” not “supported.” A partial match may be useful as a possible connection, but it should not be presented as support for the main conclusion.

A fact may have changed after a system accessed the page. Save the observation date and source-check date, and for unstable pages keep a snapshot or other safe artifact. This does not prove that this exact fragment was read, but it makes the audit reproducible.

Do not refresh an article date merely because you reopened the page. A new date is justified by a substantive change, correction, or new check that changes the conclusion.

A useful record contains the original answer, atomic claim, URL, fragment, date, support status, and a short boundary note. Store visible link and domain mention separately: they are different observations answering different questions.

If the fragment cannot be found, do not substitute another page from the same site merely to fill the row. A negative or unknown result can be correct; it protects the editorial team from false confidence.

In a public article, separate “domain shown,” “link visible,” “claim supported,” “quote accurate,” and “reader can act.” This ladder is clearer than one aggregate metric and prevents a technical log from becoming a claim of trust.

The editorial recommendation is to show not only strong matches but also cases where the domain appeared while the needed claim was unsupported. A clear boundary of knowledge is more useful than a promotional list of wins.