Determine whether apparently broad agreement comes from independent observations or one claim moving through a publication network.
URL counts are a poor proxy for corroboration. A press release can become a wire item, ten trade stories, a consultancy chart and a survey question while still carrying one originating observation. W3C's provenance model offers the right primitives: entities can be derived from, quoted from, revised from or linked to a primary source, while agents carry responsibility [1]. AAPOR adds a practical disclosure test: identify who sponsored and conducted research, the instrument used, and the population under study [2]. Together they support a compact ownership-and-origination graph that counts independent evidence paths rather than pages.
Represent the chain as nodes and typed edges
Create one node for each document, dataset, organization and controlling owner. Give documents a canonical URL, publication date and content hash where available. Give datasets a method, reference period and steward. Give organizations an ultimate parent plus sponsor, funder or commercial relationship relevant to the claim. The graph need not use semantic-web tooling; a table with stable IDs is enough if relationships are explicit.
Use typed edges: quoted-from, derived-from, revision-of, generated-from-data, funded-by, published-by, controlled-by and syndicated-from. Reserve primary-source for a record produced by an actor with direct knowledge or measurement, not merely the earliest page found. W3C distinguishes derivation, quotation, revision and primary-source relationships [1]; preserving those differences prevents a later summary from becoming a second observation simply because it has a different logo.
Collapse dependence before scoring diversity
For a specific claim, walk upstream until each branch reaches an observation or an unresolved origin. Collapse pages that share the same wire, report, survey panel, registry, analyst estimate or corporate owner when that relationship can influence the claim. Ownership alone does not always erase editorial independence, but it is a reason to inspect origination. Shared raw data does not erase methodological independence if analysts transform it differently; record the separate methods rather than treating the outputs as either wholly independent or identical.
Count three things separately: origin organizations, upstream datasets and measurement methods. Two reports from different firms using the same government series provide interpretive diversity, not observational replication. A regulator's filing and a company's release may be separate documents but one underlying disclosure. Conversely, a transaction series and a representative survey can disagree yet be more informative together because their error structures differ. Independence is claim-specific, not a permanent grade attached to a publisher.
Audit the graph with a worked example
Hypothetical example: a claim that 70% of firms are adopting a tool appears in eight articles. Three quote Consultancy A's survey; two copy a wire story about that survey; one vendor blog embeds Consultancy A's chart; another consultancy cites the vendor; and one trade association runs its own member poll. The graph collapses the first seven pages to one survey origin. The association poll is a second origin but may cover a different population. The honest result is two non-comparable origins, not eight confirmations.
The analyst then requests the two questionnaires, sampling frames and field dates. AAPOR's disclosure list makes these details part of interpretability rather than decoration [2]. If Consultancy A was sponsored by the vendor named in the claim, add the funding edge; do not automatically discard the survey, but label the interest. Publish the chain summary beside the conclusion: one vendor-sponsored general-business panel, one association member poll, definitions differ, no observed usage data.
- Stop at the measurement origin, not the first page you opened.
- Track owners, sponsors, upstream datasets and methods as separate dependency types.
- Report interpretive diversity separately from independent observation.
Take it into the meeting
- Count origins and methods, not domains.
- Make commercial and ownership relationships visible at claim level.
- Preserve republication chains as narrative-velocity data, not corroboration.
Sources & boundaries
Source statements are attributed; the decision process is Signal Atlas analysis. Examples marked hypothetical are teaching inputs, not observed outcomes.
- Ultimate ownership and funding relationships may be incomplete or change over time.
- Shared data can still support independent analysis; collapsing it requires judgment.
- A provenance graph describes dependence but does not establish that the originating measurement is accurate.
- PROV-O: The PROV OntologyWorld Wide Web Consortium · Source publication: 2013-04-30 · Retrieved 2026-09-19
Provenance chains can represent entities, activities and responsible agents. The model distinguishes derivation, quotation, revision and primary-source relationships.
- Transparency InitiativeAmerican Association for Public Opinion Research · Source publication: not established · Retrieved 2026-09-19
Research disclosures should identify sponsors, conductors, instruments, populations and sampling decisions. Methodological disclosure enables users to assess appropriateness but is not itself a quality certification.