Why a graph is more useful than a page score
A page score can indicate that a report contains authorship, dates and external references. It cannot prove that the reference beside one statistic supports that statistic, that the named author matches structured data, or that another page does not publish a conflicting figure. These are relationship questions. A graph represents each observed object as a node and each published relationship as an edge.
The minimum useful node set includes canonical pages, extractable material claims, external evidence sources, visible or structured authors, and explicitly declared JSON-LD entities. Edges describe facts such as a page publishing a claim, a claim linking to an adjacent source, a page declaring an entity, or a page naming an author. Keep the exact URL and evidence class on every edge.
Extract structured identity before removing scripts
JSON-LD is commonly published inside script elements. Parse it before removing scripts from the readable document. Traverse arrays and @graph objects with a bounded queue, retain types, names, @id values, URLs, sameAs references, authors and publishers, then separately strip scripts when creating visible page text.
This order matters. A crawler that removes scripts first may report structured data as absent even when it exists. Conversely, merely finding a type is not enough. Retain the identity fields needed to test whether the same organisation, person, product or dataset is declared consistently across the site.
Map each claim to its nearest evidence
Find the passage that most closely contains each extracted material claim. Source links inside that passage create an adjacent-evidence edge. External sources elsewhere on the page create a weaker page-level relationship. If neither exists, classify the claim as unsupported. These states must remain separate in the report and export.
A bibliography proves that a page cites sources. It does not prove which source supports a particular material claim. Adjacent evidence receives the strongest classification because the published relationship is explicit.
Prioritise broken claim edges using business evidence when available. If the owning page maps to imported search demand, show that demand beside the exact unsupported claim—but label it correctly as search-platform exposure, not AI prompt volume.
Detect possible factual conflicts without declaring errors
Compare quantitative claims on different pages only when their non-numeric language is sufficiently related. If the related claims contain different extracted values, create a verification candidate with both exact sentences, pages, values, similarity and source relationships. Do not automatically declare one claim false.
Different values can both be valid when their period, geography, sample, unit, method or scenario differs. The prescribed action is to verify the primary evidence, keep one current canonical record, and state the distinguishing scope wherever both values remain public. The completion test is not merely deleting one sentence; it is eliminating unexplained contradiction.
Audit visible and structured identity
Compare visible author signals with Article or Report author declarations. Flag a mismatch when the names differ, and a visibility gap when structured authorship exists without a corresponding visible signal. For content types that require accountability, verify that the structured record names a publisher.
Group structured entities by normalized name. The same entity using multiple @id values can fragment the site’s identity record. Incompatible type declarations—for example, the same name represented as both a Person and an Organization—require review. A recurring entity with neither a stable @id nor a truthful sameAs reference may need a canonical identity record.
Measure evidence-source concentration carefully
Count the domains supporting material claims and calculate concentration only within the analysed claim-source edges. A high HHI can expose dependence on one evidence publisher. That is a resilience and research-quality prompt, not a command to add arbitrary links. Every additional source must directly support the claim and improve the evidence basis.
Monitor evidence loss across releases
Fingerprint each claim with its page. On the next comparable crawl, detect whether the same claim moved from adjacent support to page-level or unsupported. A template migration can remove footnotes while leaving figures visible; an evidence-loss alert should record the exact claim, prior and current support state, affected page, owner and automatic recrawl test.
Turn graph findings into controlled work
- Research and legal: verify related public claims with different values and document why any legitimate difference remains.
- Editorial: attach primary evidence beside material claims and expose scope, units, period and limitations.
- Web operations: align visible authors, publisher relationships, schema types and stable entity identifiers.
- Analytics: join imported demand to the owning page without double-counting it across claims.
- Release management: treat evidence loss on high-value pages as a regression with an automatic retest.
Keep the evidence boundary explicit
An evidence graph describes the public pages that were analysed. It can prove that a claim and link were published together, or that two structured records disagree. It cannot prove that an answer engine trusted the source, retrieved the passage, used it for training, or will cite it later. Provider observations remain a separate timestamped evidence layer.
What to include in the professional export
Export claim ID, exact claim, extracted values, page URL, page title, support state, passage fingerprint, section heading, passage match, adjacent source URLs, page-level sources, measured demand, authors, entities, conflict candidates, identity issues and historical support movement. A professional reviewer should be able to reproduce the recommendation without accepting a black-box score.