Syndicated AI sources: is the evidence independent?
An AI answer may link to several pages repeating the same claim. Counting them as independent confirmations requires something the URL list does not show: the origin of each passage's evidence. Nor should an entire article be discarded as a copy if part of it adds a new check.
This guide proposes provenance work for analysts already retaining answers, citations and documents. Its output is a claim, passage and version record, with justified relationships and unresolved origins. It does not calculate concentration, organize press campaigns or reconstruct the AI's internal path. This is a review methodology proposed by Mentio, not an automated detector or independence certification.
Define the claim needing corroboration
Start with a specific claim in the answer. Preserve entity, scope, date, quantity and conditions; topical similarity does not mean two pages support the same fact. Split a sentence containing facts that may have different origins while retaining their relationship to the original wording.
Define what you need to establish about provenance. Finding the document publishing a company's own terms differs from checking a comparison between companies. A primary source may be adequate without several confirmations, while a comparative claim may need further review.
Keep two questions separate: whether the passage supports the claim and where it gets its evidence. The citation-to-claim support audit addresses the first. This review addresses the second; a copy may faithfully preserve its origin's statement without adding new corroboration.
Create passage-and-version nodes, not just URLs
Assign identifiers to the claim and each passage addressing it. Retain cited URL, consulted URL, title, declared author or organization, displayed publication or update date and your review date. Record the passage's location and a permitted copy or sufficient reference for checking it again.
A URL can change. Create a new node or explicit version when reviewing another edition; do not overwrite the passage that justified an earlier relationship. If only the current page is available, state that limit rather than asserting what it contained when the AI answered.
Use domain and editorial organization as attributes, not substitutes for the evidence's origin. Different publications may use the same document, and one publication may produce several separate checks. The record should distinguish those cases before grouping them.
Justify each derivation relationship
To connect two passages, retain evidence of the link: express attribution, a link to the material used, a republication note or a method description. Distinguish what the text declares from what you could verify. Similar wording must not become a confirmed relationship by itself.
The W3C PROV-O ontology distinguishes entities, attribution and derivation, with specific quotation and revision relationships. That separation provides a conceptual reference. This guide's procedure does not require a PROV-O implementation, and its review labels are not an official W3C or search-engine classification.
Keep direction clear: the derived passage points to the material it comes from. Record relationship type, affected claim, reason, reviewer and verification state. Where there are only clues, leave a candidate relationship pending rather than using it as a fact to deduplicate the whole set.
Separate copied material from new contributions
Review each link's scope within the article. A report may repeat a figure and add an interview; the interview is not thereby independent evidence for the figure. Equally, reused text does not automatically invalidate a new measurement documented in another passage.
| Passage finding | Provenance record | What remains to check |
|---|---|---|
| Attributed republication of a claim | Relationship to the named article or document | Whether the consulted version preserves that fact |
| Commentary on another report | Dependence on the report for its data | Whether it adds evidence or only interpretation |
| Described original check | Material and method it declares using | Whether the evidence is distinct and sufficient for the fact |
| Similar wording without verifiable attribution | Candidate relationship, not yet confirmed | Origin, direction and scope of possible derivation |
| Article with passages from several origins | Separate relationships by claim | Each passage's contribution without extending its independence to the rest |
Do not label an entire URL independent because it contains an original section. Do not group all its claims under one citation's origin either. The useful output is a reviewable passage classification, not a permanent label for a publication.
Follow the chain without inventing the first origin
Follow verifiable attributions to the document or evidence you can inspect. Retain intermediate steps: they help locate a transformation or lost condition. A broken link or inaccessible document is a limit to the chain, not a gap to fill with the most similar page.
The earliest date you find does not establish authorship or absolute priority. A page may have been updated, migrated or republished with incomplete timing information. Likewise, a canonical is a technical signal you can record, but it does not replace checking a claim's origin.
If two articles cite each other, preserve the cycle and look for the evidence they actually refer to. Do not create a fictional root source to make the diagram a tree. If no evidence outside the cycle is found, show that the origin remains unresolved.
Assess evidence independence, not just the publisher
Ask which observation, document, testimony or dataset supports the passage. A different byline or new wording may add editorial review without establishing an independent empirical basis. Record who publishes separately from the evidence used.
A new interpretation of the same data may be valuable without being an independent measurement. Several texts repeating the same testimony do not become different testimonies because their domains differ. Retain those specific dependencies without dismissing content for its distribution format.
To substantiate new corroboration, explain what additional evidence was reviewed and which part of the claim it supports. If a source only says it verified a fact without exposing enough material for your review, record the declaration and its limit. Do not complete the record by assigning an unknown method to it.
Adjudicate disagreements and retain unknowns
Require another reviewer to reconstruct each relationship from the passage and evidence. Retain reasons for disagreement: a version may be missing, attribution ambiguous or a check's distinct contribution unclear. Resolving disagreement does not mean choosing whichever label produces more independent sources.
Separate "verified derivation," "documented additional evidence," "candidate relationship" and "origin not verifiable" within your protocol. These labels describe completed work and limits; they are not universal confidence percentages. An unproven link does not substantiate independence either.
Record corrections and version changes without deleting earlier reviews. If a source withdraws a figure or changes attribution, revisit affected relationships. Do not propagate that correction automatically to passages that never depended on it.
Deliver the graph with evidence and limits
Include claims, versioned nodes, directed relationships, evidence for each link, reviewers and pending cases. Linked tables are sufficient; an attractive graph does not improve an unjustified relationship. Check that another analyst can trace a conclusion to its supporting passage.
Deliver grouping rules and exceptions too. The source concentration analysis can subsequently use these relationships to examine dependencies; this guide does not recalculate its indicators or simulate source removal. Preserve uncertainty during the handoff rather than turning all pending cases into copies or independent sources.
If the communications team uses the findings to seek corroboration, refer it to Digital PR for AI visibility. More publications do not replace new evidence. This review closes with a narrower result: which provenance has been substantiated for each claim and what still needs checking.
Frequently asked questions
Do two different domains provide two independent corroborations? Not necessarily. They may repeat an article, testimony or the same data. Review the evidence's origin for each claim and retain domain, publisher and provenance separately.
Is a syndicated article always useless? No. It can improve access and preserve information accurately. The review determines whether it adds new evidence for a particular claim, not whether distribution has value in itself.
Is similar wording enough to group sources? No. Similarity can guide review but does not substantiate a derivation's origin, direction or scope. Keep the relationship as a candidate until evidence justifies it.
What if an article mixes reused text and original checks? Separate passages and claims. Record relationships for reused material and check what new evidence the other part adds, without applying one label to the whole URL.
Can I consider a source independent if I cannot find its origin? No. Record the origin as not verifiable from available material. The absence of a demonstrated relationship neither proves independence nor authorizes inventing a root source.
Want to know if AI mentions your brand?
Discover your visibility in ChatGPT, Claude and Gemini in minutes.
Related articles
AI Source Concentration Risk: How to Measure Dependency
Measure source concentration by domain and claim, calculate Top 1, Top 3 and HHI, and test which evidence would be lost if a source became unavailable.
GEO AnalyticsDoes the citation support the answer? Auditing AI claims and passages
Check whether AI citations support each claim using versioned passages, complete, partial, absent or contradictory support, and a clear review record.
GEO StrategyDigital PR for AI Visibility: How to Earn Mentions in Authoritative Sources
Turn authority gaps into measurable Digital PR campaigns with stories, media targeting, outreach, editorial coverage and AI-response tracking.