Count evidence origins for each claim, not links in the bibliography. An original study, its press release and a syndicated story can all be useful documents while providing only one underlying set of observations. Conversely, a shared publisher does not prove that two investigations used the same evidence.
The useful question is what the next source adds: a copy, a new interpretation, an updated record or a new collection. This reference maps those relationships without treating all secondary reporting as worthless or every new domain as independent confirmation.
- 01Separate reports from observations.A new URL or title can describe the same underlying work.
- 02Keep useful follow-ups.A correction or secondary report can contain information absent from the first version.
- 03Assess the claim.A source may add direct evidence for one sentence while repeating another.
01 — Classify what the second source addsClassify what the second source adds
Use the table for a particular assertion, not as a permanent label for a publisher. “New collection” records a different acquisition activity; it does not guarantee statistical independence. Common samples, instruments or upstream feeds may still connect the results.
| Relationship and group | What may be added | How to treat the evidence |
|---|---|---|
| Reproduction: Exact mirror | Another location for the same document | One report; retain a usable location without adding corroboration. |
| Reproduction: Syndicated article | A new publisher or headline | Trace the shared byline and originating report. |
| Reproduction: Translated report | Access in another language | Preserve translation differences; the underlying study is unchanged. |
| Reproduction: Summary of a study | A shorter explanation | Cite the original for its result; check omitted qualifications. |
| Reproduction: Same press release repeated | Distribution across outlets | Count the originating announcement once for the repeated claim. |
| New interpretation: Reanalysis of shared data | Different analytical choices or calculations | Compare methods; do not call it a newly collected sample. |
| New interpretation: Fact check of existing evidence | A check of an assertion against already available evidence | Credit what was checked; classify newly collected evidence separately. |
| New interpretation: Expert commentary | Interpretation and domain context | Distinguish opinion from an additional observation. |
| New interpretation: Review of multiple studies | Synthesis across underlying studies | Trace included studies before combining with the originals. |
| Evolving record: Preprint and journal version | Revision, review or fuller reporting | Link versions and inspect whether data or conclusions changed. |
| Evolving record: Corrected report | A correction to the existing result | Use the correction for the current claim; preserve the history. |
| Evolving record: Follow-up of the same cohort | Later observations of the same participants | Retain new timing while disclosing the shared cohort. |
| New collection: Replication with a new sample | Fresh observations under a related method | Check sample overlap, protocol and claim comparability. |
| New collection: Separate direct interview | A separately collected account | Check whether the witness has firsthand knowledge or repeats another source. |
| New collection: Independent document inspection | A separate examination of primary records | Credit the inspection; disclose common underlying records. |
| New collection: Separate instrument or dataset | A different observation path | Check whether both depend on the same upstream data feed. |
02 — Why one study can have several useful reportsWhy one study can have several useful reports
The Cochrane Handbook’s search and selection chapter separates studies from reports of studies. It advises linking reports of the same study while retaining secondary reports that may contain additional information. We adapt that distinction to business research; this article is not a clinical-review protocol.
The W3C PROV primer models derivation and revision separately. Its corrected-dataset example connects a new chart to the data version that produced it. That is a useful basis for recording whether a later document adds observations or changes an earlier representation.
Our four groups are editorial categories above those source concepts. They are not probability weights, and a chart counting these rows cannot estimate how often research agents double-count evidence.
03 — Build a source family for a consequential claimBuild a source family for a consequential claim
Choose the assertion that drives the recommendation. Open its cited document and follow the attribution to the underlying study, interview or record. Record the original title, authors or institution, observation period and any stable identifier. Then inspect the other documents for those same details.
An illustrative bundle might contain a laboratory report, a press release, a news story quoting that release and a translated version of the story. Keep the four documents in one source family for the laboratory result. If the news story also interviews an outside researcher, that interview can add interpretation without becoming a second experiment.
Similar numbers and wording are clues, not proof. Two separate studies can share a method and an institution. If the connection cannot be established from inspected material, mark the relationship UNVERIFIED and avoid claiming independent corroboration.
04 — Handle reanalysis and mixed reporting carefullyHandle reanalysis and mixed reporting carefully
A reanalysis can challenge a finding without collecting new data. It may expose a calculation error, change an exclusion rule or use a different model of uncertainty. Those are substantive contributions. Describe them as a new analysis of the same data, and keep the method disagreement visible.
A follow-up may add a later outcome while observing the same participants. It is neither a mere copy nor a wholly unrelated population. The appropriate description depends on the claim: evidence about change over time can be new even when the cohort is shared.
An article can also combine repeated claims with original interviews or records. Split its contribution at the claim level. The citation-check reference helps test whether each passage supports the wording; this reference helps describe the relationship between the passages.
05 — Report the evidence without inflating agreementReport the evidence without inflating agreement
Replace a vague statement such as “several sources confirm the result” with a precise account of the evidence you inspected. Name the original study, identify a reanalysis as such and say what any separate investigation checked. Do not convert a document count into a confidence percentage.
Keep the source-family map beside the draft so updates propagate to the dependent claims. A correction to the originating report can affect every summary that relied on it. The presence of unchanged copies online does not restore the old finding.
For a related but different problem, see why AI reviewers can agree on a wrong answer . For revisions to published prose, use the fact-preservation guide .
- Scope
- 16 source relationships across 4 evidence groups. The complete selected reference appears above; no claim of exhaustive coverage of all systems.
- As-of date
- September 7, 2026. This is the actual source collection and review date; publication is assigned to the September 6 batch.
- Collection
- Read W3C PROV derivation/revision guidance and Cochrane Handbook sections on studies versus reports. Select document relationships an editor can investigate. Classify contribution to a specified claim, separating reused data, revised data and fresh collection. Do not infer independence solely from domains or authors.
- Counting
- Each row is one editorial case and belongs to its displayed group. Chart widths use 45 SVG units per entry. Group sizes describe our selection, not a measured distribution.
- Sources and interpretation
- W3C PROV supplies derivation and revision concepts. Cochrane distinguishes studies from their reports and says secondary reports may add useful information. Applying that distinction to AI-generated business research is our editorial adaptation; it is not a claim of systematic-review compliance.
- Exclusions
- No vendor census, model benchmark, search-volume estimate or observed failure rate. Examples are hypothetical; no customer operations were tested.
- Gaps and limitations
- UNVERIFIED means the evidence has not been inspected or the relationship remains unresolved after inspection. Similar cases can overlap in practice; classify the particular claim or operation, and retain uncertainty when the distinction cannot be established.
06 — DecisionWhat to do next
Describe the contribution of each source.
Keep copies, corrections and commentary when they help the reader, but explain what they add. Independent corroboration requires tracing evidence, not finding a second website with the same sentence.
For implementation support, explore our AI transformation services.