A page is not the same as an answer unit
Page-level relevance establishes that a URL covers a topic. It does not establish that one bounded passage answers the question without requiring a reader or retrieval system to reconstruct the conclusion from several sections. A defensible audit therefore retains the exact paragraph or sentence window, its section heading, source URL, position, language, word count, visible evidence links, authorship and dates.
The goal is not to force every subject into a short paragraph. Detailed research still belongs on the page. The answer unit provides a reliable entry point: direct conclusion first, then scope, assumptions, units, evidence and limitation. The deeper method and supporting detail can follow.
Create stable passage boundaries
Use meaningful HTML structure when it exists. Associate paragraphs, list items, table cells and block quotations with their closest section heading. Divide unusually long blocks into sentence windows while preserving order. Set a maximum inventory so the crawl remains bounded, and fingerprint each passage from its URL, heading and exact text.
For documents and pasted drafts, use paragraph breaks and language-aware sentence segmentation. Do not silently translate the evidence. A Chinese, Arabic, German or Japanese passage should be evaluated in its original language and exported exactly as inspected.
Score six disclosed components
- Question relevance. How much of the meaningful question vocabulary appears in the passage and heading?
- Answer completeness. Does the unit contain enough of the answer to be useful on its own?
- Visible provenance. Are primary sources adjacent, and are author and date signals present?
- Factual specificity. Does the unit name figures, units, dates, comparisons, methods or scope?
- Self-containment. Does it name the subject instead of opening with “this result” or another unresolved reference?
- Freshness. Is there a visible date or coverage period appropriate to the question?
30% relevance, 20% completeness, 20% provenance, 15% specificity, 10% self-containment and 5% freshness. The report should expose every component rather than ask the reader to trust one opaque number.
Use diagnoses that determine the action
An answer-ready passage is relevant, sufficiently complete, self-contained and visibly supported. A relevant but unsupported passage carries useful factual content without adjacent evidence. A fragmented passage depends on surrounding copy or omits a necessary qualifier. No retrievable answer means no bounded unit clears the minimum relevance threshold. Competing passages means two pages answer the same question at similar strength, creating ambiguity or contradiction.
These states should not be collapsed. An unsupported answer needs evidence adjacency; a fragmented answer needs rewriting; a missing answer needs a new bounded unit; an internal conflict needs one canonical owner. Publishing another article for all four conditions can make the problem worse.
Distinguish adjacent evidence from page-level evidence
A bibliography somewhere on the page does not necessarily support the paragraph being scored. Record links embedded in the answer unit separately from source links elsewhere on the page. Page-level evidence is useful context, but it should receive less provenance credit and remain labelled accurately.
For material quantitative and comparative claims, the answer unit should identify the source publisher, date, scope and units close enough that a reviewer can connect the claim without guessing. Where a first-party model or dataset is used, add a concise method and version.
Find internal answer conflicts
Rank more than one passage for each question. If top passages on different URLs have similar relevance, compare their figures, periods and conclusions. A dated research page and an undated glossary may both appear credible while reporting different ranges. Keep one canonical evidence owner, convert the secondary page into a definition or summary, and link it to the current source.
Detect passage loss between crawls
Passage fingerprints make editorial regressions visible. If a previously answer-ready unit disappears, changes page, loses its source, or falls into a fragmented state, create a durable alert containing the former passage identifier, current replacement, affected URL, score movement, owner and recovery test.
Verify readiness and provider outcomes separately
After publication, rerun the same buyer question against the public-page inventory. Require a clear minimum score, relevance, provenance and distance from the closest competing page. This verifies the owned content change. If server-operated providers are available, repeat the unchanged question and provider set afterward. That second test records an observed response; it must not be replaced by the passage score.
What the professional export should contain
Export one row per question with status, movement, readiness score, passage fingerprint, URL, section heading, exact excerpt, word count, all six score components, adjacent and page-level sources, runner-up URL and relevance, measured demand when connected, precise action, and acceptance test. This lets editorial, research, legal and analytics teams review the same evidence without reverse-engineering a dashboard.