Trace five apparent sources to their evidence origin, expose circular citation and distinguish independent corroboration from repetition or transformation.
Short answer: answer: Write the exact proposition, open every cited source and trace its evidence backward until you reach a study, dataset, direct record or unsupported terminus. Draw each citation as a directed edge and collapse sources that depend on the same origin. Five pages provide one line of support when all five inherit one source; independence requires separately generated evidence, not different domains.
Search results encourage counting. A claim repeated by a university blog, newspaper, vendor and consultancy feels corroborated even when four cite the fifth—or none links beyond “research shows.” Domain diversity is not evidence diversity.
Direct citation counts are incomplete too. A source can cite a real paper that studied a narrower proposition. CDM's position is that a source earns corroboration only after its ancestry and claim fit are inspected; reputation cannot substitute for the edge between sentence and evidence.
The Citation Ancestry Trace
The Citation Ancestry Trace is a claim-specific directed graph. Nodes are documents or data; edges mean this source relies on that source for this proposition. Its six passes are normalize, extract, follow, classify, compare and decide.
Pass 1: normalize one proposition
Copy the sentence from each page, then rewrite the smallest testable claim without rhetorical variation. Humans have an eight-second attention span, attention fell from 12 to eight seconds, and people focus less than goldfish are related but different claims. Track them separately.
Record qualifiers, population, date and measure. If a page says according to Microsoft, attribution is not corroboration—it is a visible ancestry edge.
Pass 2: extract the evidence actually attached
Inspect footnotes, hyperlinks, captions, chart notes, bibliography and downloadable report. Save the link text and exact source location. A bibliography containing a relevant title does not prove the statistic came from it; identify the citation nearest the claim.
Mark no citation, self-citation, secondary citation, primary evidence or broken/archived. Use the Internet Archive or publisher archive when appropriate, but record that an archived page reflects one date and may lack attachments. Do not infer a dead page's contents from other people's descriptions.
Pass 3: follow every edge to its terminus
Open the cited source and repeat. Stop only at a primary dataset or study you can inspect, a direct official record, an explicit expert observation, or a terminus that provides no verifiable support. A report citing industry research is not a terminus; it is a missing edge.
DOIs and Crossref Cited-by metadata help discover connections, but Crossref explains that its links depend on references deposited by members. Missing metadata does not prove no citation exists, and a citation link does not prove support. Read the source.
Pass 4: classify transformations
For each edge, label how the claim changed: faithful quotation, rounding, population expansion, period removal, correlation-to-causation, forecast-to-actual, comparison invention or unsupported paraphrase. Preserve the before-and-after wording.
This is where circularity becomes visible. Source C may cite B, which cites A, while A cites C for a different sentence. None supplies an origin. Or five pages may independently cite one original survey; that is one study with five repeaters, not a circle but still not five corroborations.
Pass 5: test genuine independence
Two studies are not independent merely because they have different titles. Check authors, sponsor, dataset, sample, collection dates and method. A press release, report and dashboard produced from one survey are one evidence family. A systematic review can synthesize many studies, but its conclusions depend on selection and quality; record that structure rather than counting its references as your own verification.
Independence also does not guarantee quality. Two weak nonprobability polls do not become strong because their samples differ. Evaluate method after ancestry.
Pass 6: assign the editorial decision
Use four outcomes: corroborated by independent fit sources; supported by one source with attribution; context narrowed to match origin; or remove/unverified. Document uncertainty. Do not launder an unsupported claim with widely reported unless the fact of widespread reporting is itself the relevant claim.
Three real documented source families
These traces rely on published records; they are not claims that CDM recreated every historical web page.
Family A: the eight-second attention-span chain
In 2015, Time reported that a Microsoft study showed a human attention span of eight seconds and a goldfish's of nine. The Microsoft Canada report circulated the figures, but its table attributed them to Statistic Brain rather than the report's own EEG and survey work. A 2017 BBC investigation reported trying to trace the figures through the named sources and finding no supporting record.
The ancestry is therefore:
``text later marketing page -> Time or Microsoft report Microsoft report's eight-second table -> Statistic Brain Statistic Brain's stated verification chain -> unresolved in BBC investigation ``
The safe conclusion is not attention spans are definitely longer. It is that these pages are not five independent measurements of an eight-second human attention span. The claim should be removed or presented as a disputed popular statistic with this trace.
Family B: the 7–38–55 communication rule
The familiar universal phrasing often collapses to two 1967 experiments. PubMed's record for Mehrabian and Wiener's Decoding of Inconsistent Communications identifies a study about inconsistent communication of attitude using words and tone. The APA-published Mehrabian and Ferris study examined attitudes inferred from facial and vocal channels, with vocal content placed over a neutral word.
Those primary scopes do not support only 7% of all communication is verbal as a universal rule. Five training pages pointing to Mehrabian still represent one narrow experimental family, and a page that drops inconsistency and attitude has transformed the claim. The appropriate editorial move is to state the narrow context or omit the percentage.
Family C: a measured citation network
Greenberg's 2009 BMJ study constructed a complete citation network for a biomedical belief about beta amyloid and inclusion-body myositis. The paper documented how citation distortion and selective citation could create authority. This is not a marketing claim family, but it is primary research showing why counting citations without reading the supporting path can mislead.
CDM did not reproduce Greenberg's network or extend its biomedical conclusion. The operational inference is narrower: editorial teams should map ancestry and transformations before calling repeated sources corroboration.
Reader asset: citation ancestry worksheet
| Node | Exact proposition | Cites for proposition | Source class | Transformation | Independent family? | Decision |
|---|---|---|---|---|---|---|
| A | “Humans average eight seconds” | Microsoft report | publisher repeater | drops source uncertainty | no | do not count |
| B | table lists 8 seconds | Statistic Brain | corporate report/secondary | presents compiled value | no | trace backward |
| C | same value, no inspectable method | named downstream sources | unsupported terminus | method missing | no | remove/unverified |
| D | narrow attitude experiment | DOI 10.1037/h0024532 | primary study | none at origin | family 1 | supports only narrow scope |
For a real review, add URL, archive date, locator, author, sponsor, dataset, access status and reviewer. Never put two different propositions in one row.
Related guides
Frequently asked questions
How many independent sources are enough for a marketing claim?
There is no universal count. One authoritative primary record can be sufficient for a directly observed fact; a causal, comparative or high-consequence claim may need multiple strong studies or the level experts and regulators expect. First confirm that sources are independent and fit the exact proposition. Five repeaters are still one family. Document why each source is fit and independent.
The caveat is jurisdiction and claim type: advertising, health, finance and environmental claims can carry specific substantiation expectations that an editorial source-count rule cannot replace.
Does a different domain count as an independent source?
No. A newspaper, university blog and vendor page may all repeat one press release or dataset. Trace their citation edges, authors, sponsors, samples and collection periods. They are independent only if their relevant evidence was generated separately, not because their logos differ. A secondary source can add independent analysis, but make that contribution explicit.
Record the shared origin. The caveat is that several independent analyses of one dataset may test robustness while still sharing the data family; label both analytical and data dependence.
What should I do when the original source is unavailable?
Use a publisher or web archive to recover the dated document, search identifiers and contact the source owner when stakes justify it. Record what you actually inspected and what remains secondhand.
Do not reconstruct methods from later summaries. If support cannot be verified, narrow the wording to the documented fact that a source reported the claim, or remove it. The caveat is that attribution does not make a misleading claim safe; a high-consequence or promotional statement may still require adequate underlying evidence.
Can citation software detect circular citations automatically?
It can surface reference links and graph cycles, but it cannot reliably determine which reference supports a particular sentence or whether meaning changed. Crossref Cited-by depends on deposited metadata and does not judge evidentiary fit. Use tools to collect candidate edges, then inspect the text, table and method.
Language models can propose ancestry with source locators, but hallucinated links are a known risk. The caveat is scale: structured DOI families permit useful automation, while web pages, deleted reports and uncited claims require manual research.
Is a systematic review an independent source or many sources?
Treat it as a synthesis node connected to the included studies, not as an automatic multiplier. Its value comes from explicit search, selection, quality assessment and synthesis; the primary studies remain the underlying evidence families. Check whether several reviews include the same studies and whether their conclusions fit your population and claim.
A high-quality review may be more useful than counting individual papers yourself. The caveat is recency and publication bias: inspect the review date, inclusion criteria and reported limitations.
How should the ancestry trace appear in a published article?
Usually cite the strongest inspectable origin and explain material limits in plain language. If the tracing problem is central, publish a compact chain or methodology note showing which sources converge and where support ends. Do not overwhelm readers with every repeater. Keep the full worksheet internally so corrections are reproducible. Keep citations close to the supported sentence.
The caveat is source safety, confidentiality and copyright: identify public records and short relevant passages, but do not expose protected material or reproduce entire reports merely to demonstrate diligence.
Next decision: Run the statistic integrity gate before publishing a number
Related reading: Build a claim evidence register · Set evidence refresh dates · Correct one fact across multiple articles
Sources and research notes
- Greenberg, *BMJ*, “How citation distortions create unfounded authority” — checked 26 September 2026. Primary citation-network study; CDM did not reproduce its biomedical analysis.
- Crossref, Cited-by documentation — checked 26 September 2026. Explains metadata-derived citation links and their scope.
- PubMed, Mehrabian and Wiener, “Decoding of inconsistent communications” — checked 26 September 2026. Primary bibliographic record and DOI for the narrow 1967 study.
- APA-published study record, Mehrabian and Ferris, DOI 10.1037/h0024648 — checked 26 September 2026. Original study family; full text may require access.
- Time, “You Now Have a Shorter Attention Span Than a Goldfish” — checked 26 September 2026. Publisher node in the real media chain, not primary measurement evidence.
- BBC, “Busting the attention span myth” archived PDF — checked 26 September 2026. Archived copy of the publisher investigation; BBC's original page was not directly accessible in this research environment.
Research limitation: the examples are documented source families, not a systematic census of every page repeating either claim. The Microsoft report was available through circulating archived copies rather than a stable current Microsoft URL, so the trace reports that access limitation explicitly.
This article is editorial guidance. Apply the principles in proportion to your market, evidence, and responsibilities.
