Editorial discipline and sourcing
Summary
Overview
Editorial discipline is the operational layer that makes a knowledge base, a public site, and client deliverables defensible under scrutiny. It is built from converging professional-journalism standards (BBC, Reuters, SPJ), an institutional confidence taxonomy (US Intelligence Community), a working internet-scale example (Wikipedia), legal substantiation doctrine (FTC), and link-rot empirics (Zittrain 2014, NYT 2021). This page consolidates a seven-label confidence taxonomy, an eight-pattern citation library, a "what to source" decision matrix, a nine-step link-rot mitigation plan, a four-magnitude retraction/correction playbook, the operational rules that govern a research-disciplined public artifact, and a catalogue of statistics that are not supported by primary evidence and therefore should not be quoted.
The discipline is summarised by a single test: if a skeptical reader asks "where did that come from?" and the writer cannot answer in one sentence with a name and a date, the claim needs sourcing. Under this standard, every objective claim in client deliverables, KB entries, and public writing carries a named source, a publication or retrieval date, a verbatim quote of twenty-five words or fewer where possible, and a confidence label drawn from the seven-label taxonomy. This is the operational expression of Wikipedia's verifiability-not-truth policy for business writing, and it is also the practical defence against AI-amplified misinformation in 2026.
See Agency methodology for small-business website projects for the original research-discipline framing this page extends. Wikipedia's three interlocking core content policies — verifiability, no original research, and neutral point of view — are the foundational reference; the discipline below adapts the verifiability policy specifically.
Seven-label confidence taxonomy (Candid 2026)
A documented practitioner confidence-label taxonomy. Many existing KB entries use an informal version of this; the taxonomy below is the formal source-of-truth going forward.
| Label | Definition | Example |
|---|---|---|
| Verified | Primary source quoted directly; URL + date + verbatim quote ≤25 words; re-checked within 6 months. | "BBC requires sourcing 'as appropriate to subject and nature' (BBC §3.2.2, 2019 ed.; verified May 2026)." |
| Industry-consensus | Multiple independent reputable sources concur; no authoritative contradiction. | "Core Web Vitals influence ranking. (Google Search Central + multiple 2024-2025 industry studies.)" |
| Single-source | One credible source, not yet corroborated. | "Wikipedia is 7.8% of ChatGPT citations. (Profound, 680M citations Aug 2024-Jun 2025.)" |
| Estimated | Author's calculation/synthesis from data; not a direct citation. | "AI Overviews appeared in 88.1% of informational queries at initial measurement; share has since shifted (Semrush, 2025)." |
| Author's view | Opinion, recommendation, or interpretation. | "We recommend KB-backed sites over brochure builds. (Author's view.)" |
| Contested | Reputable sources disagree. | "Whether E-E-A-T is a direct ranking factor is contested. (Google: no. Some SEOs: yes.)" |
| Stale / under review | Was Verified/Industry-consensus, source >18 months old, awaiting re-check. | "Link-rot rate in legal citations was ~50% after a decade (Zittrain 2014 — methodology likely still valid)." |
Lineage
The three-tier base (High / Moderate / Low confidence) is drawn from the US Intelligence Community three-tier confidence taxonomy (see the IC section below). The seven-label adaptation extends that base with four additional labels needed for business writing that the IC framework does not need:
- Estimated — bridges the gap between Industry-consensus and Single-source when the writer is doing arithmetic on cited inputs.
- Author's view — makes opinions visibly opinions, not stealth-sourced claims.
- Contested — explicitly distinguishes "honest disagreement among credible sources" from "no good source."
- Stale — the time-decay label that turns the KB into a living document.
How to apply
The labels are used inline at the research stage (Stage 1 of the research-first workflow). In derived public articles, the labels are smoothed into prose ("according to verified sources" / "industry estimates suggest") but the underlying KB entry preserves the explicit label. Marketing pages link back to the article; a reader pursuing one click sees the labeled source.
The seven-label discipline is the operational expression of Wikipedia's "verifiability, not truth" policy for business writing. WP:V identifies four claim types that always require inline citations — direct quotations, contentious material about living persons, anything challenged or likely to be challenged, and statistics — and the seven-label taxonomy applies a confidence flag to each.
Citation pattern library (eight patterns, 2026)
A documented citation pattern library. The pattern is chosen to match the artifact, not the writer's preference.
| Pattern | When to use | Tradeoff |
|---|---|---|
| Inline link in running prose | Default for 1-2 citations per paragraph | Lowest friction; risk of "link to a blog repeating the claim" = not sourcing |
| Inline parenthetical with source + date | KB articles, research briefs (recommended default) | Heaviest signal of rigor; slightly heavier prose |
| Hover footnotes (Wikipedia Reference Tooltips style, sidenotes.css) | Long-form analytical content | Excellent on desktop; requires keyboard/screen-reader accessibility; needs tap-to-expand on mobile |
| Numbered footnotes with backlinks | Academic-style depth, 15+ citations | Familiar; requires scrolling. Tufte-style margin sidenotes preserve eye-line |
| End-of-section "Sources" block | Topical clusters of citations | Cleaner prose; pairs well with claim numbering |
| Dedicated "Sources & Methodology" page | Marketing pages making research-derived claims | Marketing page breathes; research page carries rigor — marketing page MUST link to source page |
Schema.org structured data (Article.citation, ClaimReview, Claim) |
Machine readers / AI engines | Invisible to humans, high AI-citation value. ClaimReview no longer surfaces as a Google rich result (June 2025) but remains useful for non-Google ingestion (see the ClaimReview section below) |
| Archive-link pairs ("Original | Archived [date]") | Time-sensitive web citations | Extra link cost; immunizes against link rot (see the NYT link half-life section below) |
Documented defaults by artifact:
- Inline parenthetical with source + date in research briefs and KB articles.
- Hover footnotes or end-of-section "Sources" blocks in long-form public articles.
- Dedicated Sources page for marketing pages derived from KB research.
- Always include archive links for time-sensitive citations.
What-to-source checklist
A "what needs sourcing" decision matrix.
Must source
- Statistics, percentages, dollar amounts, dates, counts.
- Direct or paraphrased quotes from named people or organizations.
- Attributions to "studies" / "research" / "experts" — these are FTC triggers under the FTC reasonable-basis doctrine (see the FTC section below).
- Claims about competitors, vendors, regulators, platforms.
- Historical events, regulatory rules, technical specs, industry standards.
- Performance claims ("X% faster" — FTC-actionable).
- Contested or non-obvious claims.
- Anything that, if wrong, damages a person, a company, or the reader.
Should source
- Industry-consensus statements with specific numbers.
- Best-practice recommendations from named experts.
- Definitions where multiple definitions exist.
Doesn't require sourcing
- Common knowledge ("Kitchener-Waterloo is in Ontario").
- Clearly labeled author opinions (Author's view label from the seven-label taxonomy).
- Clearly framed first-hand experience.
- Generic descriptions of the publisher's own services (subject to FTC truthfulness / substantiation if comparative).
- Things the reader can verify by looking at the page.
The test
If a skeptical reader asks "where did that come from?" and the writer cannot answer in one sentence with a name and a date, it needs sourcing.
Link rot empirics
NYT 15-year half-life, 13% content drift
A 2021 study of New York Times external links from 1996-2019 found a ~15-year half-life of working links, with 13% of "live" links no longer pointing to original content (content drift, not just 404s). The source is the Wikipedia synthesis citing the 2021 NYT-link study at https://en.wikipedia.org/wiki/Link_rot. Confidence: Industry-consensus.
Zittrain 2014 (companion)
Zittrain, Albert and Lessig (Harvard Law Review, 2014) studied citation rot in legal scholarship and US Supreme Court opinions. The verbatim finding:
"More than 70% of the URLs within the above mentioned journals, and 50% of the URLs within U.S. Supreme Court opinions suffer reference rot."
Source: Zittrain, Albert and Lessig, Harvard Law Review, 2014. Confidence: Verified (foundational academic study). The older study reports larger numbers in the same direction as the NYT 2021 finding above.
Empirical assumption for sourcing
Any web citation has a meaningful probability of breaking within 5 years. The defense is not "don't cite the web" — it is capture + quote. See the nine-step plan immediately below.
Link-rot mitigation: 9-step plan
- Capture an archive snapshot at citation time — archive.org Save Page Now, Perma.cc. Store the archive URL alongside the original.
- Prefer persistent identifiers: DOIs, ISBNs, government document numbers, regulatory filing IDs.
- Cite the most stable host: Wikipedia (revision permalink), .gov, .edu, established publishers.
- Quote the source verbatim, don't just link. A ≤25-word direct quote in the citation means the claim survives link death and can be searched in archives. Highest-leverage practice.
- Quarterly automated link-checker on the KB.
- Quarterly manual spot-check of high-importance citations.
- Own-domain mirror (PDF) of critical external sources where copyright/fair-use permits.
- Date every citation ("as of May 2026" / "retrieved May 2026").
- Update
dateModifiedin structured data when sources are re-verified or replaced.
Self-published research outputs
Get a DOI (free via Zenodo) for any self-authored research artifact (white papers, downloadable reports, methodology documents). The DOI persists even if the original URL changes.
Why verbatim quote is the highest-leverage step
When a URL dies, a footnote like "Source: example.com/article-123" leaves no recovery path. A footnote like "Per BBC §3.2.2: 'All BBC output...must be well sourced, based on sound evidence, thoroughly tested.' (BBC Editorial Guidelines, 2019 ed.)" is searchable in any archive even if the URL is gone. The verbatim quote is the durability layer; the URL is the convenience layer.
Professional-journalism converging standards
Three independent professional-journalism standards converge on the same point: name your sources, prefer named to anonymous, attribute the uncorroborated.
BBC Editorial Guidelines §3.2.2 — sound evidence
"All BBC output, as appropriate to its subject and nature, must be well sourced, based on sound evidence, thoroughly tested and presented in clear, precise language."
Source: https://www.bbc.co.uk/editorialguidelines/guidelines/accuracy (2019 edition). Confidence: Verified.
Companion §3.3.4 (single-source reluctance):
"We should be reluctant to rely on a single source. If we do rely on a single source, it should be credible, and a named, on-the-record source is always preferable."
Companion §3.2.3 (attribution required when uncorroborated):
"Claims, allegations, material facts and other content that cannot be corroborated should normally be attributed."
The BBC §3 cluster is the cleanest single editorial-standards source. Three rules, all transferable to marketing content: sound evidence basis; prefer named on-the-record sources to single anonymous sources; attribute the uncorroborated.
SPJ Code of Ethics (2014) — identify sources clearly
"Identify sources clearly. The public is entitled to as much information as possible to judge the reliability and motivations of sources."
Source: https://spj.org/spj-code-of-ethics. Confidence: Verified.
Companion: the Washington Post / two-source norm (December 2017): "There must be two sources with the same firsthand knowledge" before publication of contested claims. Industry-consensus in working newsrooms.
The "reliability and motivations" framing matters more than the citation mechanic. Readers need to evaluate not just what was said but who said it and why they might say it. A named source with credentials, affiliation, and date does all three. An anonymous source does none.
Reuters Handbook — anonymous sources are the weakest
"Accuracy entails honesty in sourcing. Our reputation... rests on the credibility of our sources."
"A named source is always preferable to an unnamed source... Anonymous sources are the weakest sources."
Source: https://handbook.reuters.com (last comprehensively revised 2012; principles remain in force at Reuters). Confidence: Verified.
Caveat on dating: the Reuters Handbook is older than the default 2023 freshness threshold. Justification for inclusion: it is the primary editorial-standards document for Reuters and remains authoritative; no successor document has replaced it.
Reuters Trust Principles (adopted 1941, reaffirmed at Thomson-Reuters merger April 2008):
"That the integrity, independence and freedom from bias of Thomson Reuters shall at all times be fully preserved... That Thomson Reuters shall at no time pass into the hands of any one interest, group or faction."
The operational application: "Anonymous sources are the weakest" is the test to apply to every "experts say" or "studies show" in marketing copy. Marketing content routinely uses anonymous sources where journalism explicitly refuses to. The fix is named sources, dated, with verbatim quote.
Vendor sources and concession-against-interest
A vendor source is by default weaker than an independent one because its commercial incentive distorts its claims. The exception — and a recognised credibility signal in editorial practice — is concession against interest: when a vendor states a fact that cuts against its own commercial incentive, the statement gains credibility precisely because the incentive points the other way. Four converging examples in the faceted-search literature illustrate the pattern.
HawkSearch (a search-software vendor) states:
"If your online catalog has just a few dozen products, basic search and navigation tools might be adequate."
Source: HawkSearch, "If They Can't Find It, They Can't Buy It" (accessed June 2026). Confidence: Industry-consensus (vendor-against-interest). Caveat: Vendor sells search software; incentive is to inflate need, here conceded against interest.
Prefixbox (faceted-search software) places a more specific cutoff:
"Your store has a small catalog (e.g., <200 products). Basic category filters solve most findability needs. The implementation cost outweighs value for your UX."
Source: Prefixbox, "Faceted Filtering and Faceted Search: Complete Guide" (accessed June 2026). Confidence: Single-source / vendor. Caveat: Same vendor incentive as above; corroborated by HawkSearch and Luigi's Box.
Luigi's Box (search software) reaches the same cutoff independently:
"smaller catalogs with only a few hundred products may not require this level of complexity. Simpler filtering systems often suffice in such cases, making faceted search an unnecessary investment in time and resources."
Source: Luigi's Box, "Faceted Search: What Is It and Why Your E-Shop Needs It" (accessed June 2026). Confidence: Industry-consensus — two independent vendors state the same cutoff. Caveat: Vendor source; credibility derives from concession against interest.
The independent corroboration comes from the Nielsen Norman Group:
"the extra power of faceted navigation also adds interaction cost by presenting users with more options to comprehend and manipulate" — and vocabulary/metadata maintenance "consume significant financial and human resources."
Source: NN/g via Taylor & Francis library-science article, "Musings on Faceted Search, Metadata, and Library Discovery Interfaces," 2023. Confidence: Verified (independent / academic).
When three competing vendors and one independent research body converge on the same cutoff against the vendors' own commercial interest, the convergence approaches industry-consensus rather than vendor marketing.
Wikipedia: verifiability not truth
"The threshold for inclusion in Wikipedia is verifiability, not truth — whether readers can check that material in Wikipedia has already been published by a reliable source."
Source: https://en.wikipedia.org/wiki/Wikipedia:Verifiability. Confidence: Verified.
Four claim types that always need inline citations (per WP:V)
- Direct quotations.
- Contentious material about living persons.
- Anything challenged or likely to be challenged.
- Statistics.
The companion verbatim policy text:
"All quotations and any material challenged or likely to be challenged must be attributed to a reliable published source using an inline citation."
Source: Wikipedia:Verifiability, https://en.wikipedia.org/wiki/Wikipedia:Verifiability. Confidence: Verified.
Why "verifiability not truth" is the right model for business writing: the writer does not have to be omniscient — only honest about where claims come from. Wikipedia is the largest-scale working example of "every claim → reliable source" at internet scale, and verifiability is one of three interlocking core content policies (verifiability, no original research, neutral point of view) that together produce a functioning encyclopedia at internet scale.
Institutional confidence taxonomy (US IC)
The US Intelligence Community uses a three-tier confidence taxonomy codified in Intelligence Community Directives 203 and 206, used in National Intelligence Estimates since at least the 2007 NIE Iran: Nuclear Intentions and Capabilities:
- High confidence: high-quality information, solid judgment.
- Moderate confidence: "credibly sourced and plausible but not of sufficient quality or corroborated sufficiently to warrant a higher level of confidence."
- Low confidence: credibility uncertain, source information scant or fragmented.
Source: NPR summarizing the 2007 NIE convention; Harvard Belfer Center analysis; ICD 203/206. Confidence: Verified (institutional documentation, long-standing practice).
Tetlock companion
Tetlock's superforecasters (Good Judgment Project) demonstrate empirically that calibrated probability statements with frequent small updates beat categorical assertions. The 10 commandments for aspiring superforecasters (Tetlock & Gardner, Superforecasting, 2015) translate directly into editorial practice: focus on questions where evidence pays off, distinguish probability from confidence, update incrementally.
CIS adaptation
The Center for Internet Security adopts the same logic for cyber threat intelligence. "Moderate Confidence generally means that the information is credibly sourced and interpreted to be plausible but is not of sufficient quality or corroboration..." The three-tier base is the source pattern for the seven-label taxonomy above.
Epistemic-status labels lower the publishing threshold
Devon Zuegel's framing:
"[Epistemic status labels are] a hack in order to publish half-baked ideas that I'd otherwise not feel comfortable sharing."
Source: Devon Zuegel personal essays on epistemic status practice. Confidence: Verified.
Confidence labeling is a publishing-threshold lowerer, not an academic affectation. Without labels, a writer faces a binary: publish at full conviction or do not publish. With labels, partial findings, in-progress thinking, and contested claims can all ship — with the reader appropriately calibrated.
Chris Krycho framing (companion): "a nice little tool to signify that kind of humility." The same impulse, less hack-ier framing.
FTC reasonable-basis doctrine (1984)
The FTC's Policy Statement on Advertising Substantiation (1984, building on a 1971 line of decisions) requires that advertisers possess "the amount and type of substantiation the ad actually communicates to consumers" for any objective ad claim. Section 5 of the FTC Act remains the enforcement basis.
Source: https://ftc.gov Substantiation Statement; FTC training materials. Confidence: Verified.
Specific phrases that trigger heightened substantiation thresholds
"Tests Prove... Doctors Recommend... Studies Show..."
Source: FTC training materials — https://ftc.gov/sites/default/files/attachments/training-materials/substantiation.pdf.
The operational implication: the phrase "studies show" in marketing copy without a specific named study is not just a sourcing failure — it is potentially FTC-actionable. The reader's instinct that "studies show" is a tell of unsourced claims aligns with FTC enforcement reality.
Status in 2026
Xlear v. FTC litigation (filed late 2024, citing Loper Bright Enterprises v. Raimondo, 2024) is challenging the doctrine. Absent injunction, the doctrine remains in full force as of May 2026. Monitor for outcome.
Canadian portability
FTC enforcement is US-only, but the substantiation principle applies via comparable Canadian rules (Competition Bureau's "general impression test" under the Competition Act, plus CRTC and provincial consumer-protection statutes). The discipline is portable; the legal mechanism differs.
Google retired ClaimReview rich results (June 2025)
Per Poynter (June 2025):
"Google decided to retire ClaimReview, a program little known outside the fact-checking world that highlighted fact checks from reputable news sources."
Source: Poynter (IFCN), June 2025 — https://poynter.org/ifcn/2025. Confidence: Verified.
Important nuance: the schema itself remains an open standard at schema.org and is still being used by fact-checkers "to provide cues to new chatbots and other AI applications" (Bill Adair, Duke Reporters' Lab).
The operational read for practitioners: do not overinvest in ClaimReview / FactCheck specifically for Google rich-result visibility — that signal is gone. But do implement broader sourcing schema (Article.citation, dateModified, Author/Person with credentials) for AI engines that index more openly. The Schema.org Article foundation remains a high-value signal.
LLMs hallucinate with high certainty (Simhi et al., Feb 2025)
Quote (Simhi, Itzhak, Barez, Stanovsky, Belinkov, arXiv:2502.12964, February 2025):
"Models can hallucinate with high certainty even when they have the correct knowledge."
Source: https://arxiv.org/abs/2502.12964. Confidence: Single-source (peer-reviewed preprint); the broader finding (LLMs produce confidently wrong output) is Industry-consensus across the hallucination literature.
Companion: Vectara 2025-2026 hallucination leaderboard shows top models at 0.7-10%+ hallucination on summarization tasks, with rates over 50% on fact recall about specific people.
Correction to circulating attribution
Many SEO and marketing blogs cite "MIT research, January 2025" for a "LLMs are 34% more confident when wrong" finding. The closest verifiable primary source is Simhi et al. (Technion/Oxford/Hebrew University), arXiv:2502.12964, February 2025. The "34% / MIT / January 2025" institutional attribution is Single-source / Contested; the core finding (LLMs hallucinate with high certainty) is Industry-consensus.
Why this matters for sourcing
Citation discipline is the only practical defense against AI-amplified misinformation. When a writer cites a verbatim source with URL + date + archive, the reader can verify; when an AI writes "studies show" without citation, the reader cannot. The asymmetry is the operational case for the seven-label taxonomy.
Retraction and correction playbook
Adapted from COPE's August 2025 guidelines + Trust Project + AP standards. The playbook divides the response surface into four magnitudes.
| Magnitude | Response | Visible? |
|---|---|---|
| Typo, broken link, formatting | Silent edit | No |
| Substantive (fact wrong, conclusion stands) | Inline dated correction note | Yes |
| Material (conclusion changes) | Top-of-page notice + dated changelog | Yes, above the fold |
| Fundamentally wrong | Retraction notice replaces body; original archived & linked | Yes; original URL serves the retraction |
Templates
Correction template:
Correction (DATE): An earlier version stated [WRONG CLAIM]. The accurate statement is [CORRECT CLAIM], sourced from [NAMED SOURCE, URL, DATE]. We thank [READER] for flagging this.
Retraction template:
Retraction (DATE): This article has been retracted because [REASON]. The original text is preserved at [ARCHIVE LINK] for transparency. We regret the error and have updated our editorial process to [SPECIFIC CHANGE].
Surface the correction
- Add to a public Corrections log (linked from the footer).
- Update
dateModifiedin structured data. - Notify any syndicators of the wrong claim.
Then learn
A correction without a process change is theater. Every material correction must trace to a specific update in the editorial process — a new check, a new source-verification step, a new automated guard.
COPE framing (Aug 2025)
"The purpose of retraction is to correct the literature and ensure its integrity, not to punish the authors."
Source: https://publicationethics.org. Confidence: Verified.
A correction without a process change is theater. A retraction stigmatized as failure produces under-disclosure. The COPE framing — corrections exist to fix the record, not punish — is the right disposition to import into editorial process. The four-magnitude retraction discipline (typo / substantive / material / fundamentally wrong) explicitly adopts the COPE framing.
Stats Candid will not quote (2026 catalogue)
The following statistics are commonly cited in marketing and SEO blogs but are not supported by primary evidence at 2026, or are misapplied beyond their original scope. Under this discipline they should not be quoted in sales conversations or marketing copy.
"A 1-second delay reduces conversions by 7% across all sites." Traces to Akamai/SOASTA 2017 retail data. Defensible as a retail e-commerce benchmark from 2017, not a universal law. For a KW lawyer or HVAC contractor, this specific elasticity has not been demonstrated.
"Amazon loses 1% revenue per 100ms of latency." Reportedly Amazon internal research circa 2006 (Greg Linden). No public methodology, approximately twenty years old, Amazon-scale operations.
"Core Web Vitals doubled in ranking importance" / "Google made CWV twice as important." Cited by some marketing blogs but not in any Google primary statement. Google has consistently described CWV as one of many page-experience signals, tiebreaker class.
"53% of mobile users abandon in 3 seconds" applied to 2026 baselines. Original study (Google/DoubleClick/SOASTA, 2016) is nearly a decade old. Networks, devices and tolerance baselines have shifted. Quote it as historical context, not a current 2026 number.
"Faster site = better AI Overview citation rate." No defensible primary research supports this as of May 2026. The one direct analysis found the opposite of what an SEO marketer would assume.
Agency "case studies" claiming "redesign → +40% conversions" as performance wins. A redesign confounds content, layout, CTA copy, navigation, trust signals, and performance. Lift cannot be attributed to speed without an isolated A/B test. Reject these.
Conversion-rate elasticity for "professional services" specifically. No major isolated study (Deloitte, Akamai, web.dev cases, Portent, Contentsquare) breaks out elasticity for lawyers, accountants, dentists or consultants in a defensible way. Extrapolate cautiously from Deloitte lead-gen and Portent B2B — but flag as "general B2B," not "professional services-specific."
"88% of online consumers are less likely to return to a site after a bad experience" (often paired with "89% shop with a competitor after a poor experience"). The figure originates with HubSpot, not with the maintenance-vendor blogs that repeat it. Source: HubSpot original; repeated by e.g. Hyperping. Confidence: Single-source — HubSpot original. Caveat: Common laundering pattern — a single vendor figure becomes "the industry says…" through repetition. When the stat is used, the citation should name HubSpot explicitly and disclose the vendor incentive.
"McKinsey: 83% of B2B customers value transparency over brand reputation." The widely-repeated figure is not verifiable to any primary McKinsey publication. It appears only in vendor blogs (e.g., Togai, DealHub) that re-cite each other. Source: Repeated by Togai / DealHub (pricing-software vendors). Confidence: Single-source / vendor-laundered — treat as unsourced.
"Princeton GEO: 30–40% higher visibility from statistics / citations." Numerous SEO vendors repeat this figure; all trace back to the same arXiv paper (Aggarwal et al.). The vendor chorus is repetition, not independent confirmation. Source: Cross-vendor observation, 2024–2026; primary is the Aggarwal et al. KDD 2024 paper. Confidence: Single-source (one primary, many echoes). Caveat: When multiple vendors all quote the same figure, the appropriate check is whether they are citing different studies — here they are not.
Recycled customer-retention magnitudes. The figures "5×–25× cheaper to retain," "5% retention → 25–95% profit," "80% of profits from 20% of customers," and "AI churn prediction → 20–30% retention improvement" circulate vendor-to-vendor across the analytics blogosphere without traceable primary citation. The 25–95% figure ultimately traces to Reichheld / Bain-era work, but the qualifying conditions have eroded through the citation chain. Treating any of these as facts inherits the credibility erosion of the chain. Confidence: Single-source / vendor-recycled. Caveat: Quarantine until primary sources are located and original conditions re-read.
Effect-size and methodology caveats from primary research
A separate failure mode is citing a primary research finding without its methodological qualifiers. Five examples from the 2014–2025 psychology literature illustrate how qualifying conditions get stripped in transit.
Flow measurement is contested. A 2025 systematic review (Wonders, Human Behavior and Emerging Technologies) found that most studies fail to screen for flow-proneness or match task difficulty to skill, undermining confidence in flow findings. Source: Wonders 2025 systematic review; Løvoll & Vittersø 2014; PMC 2020 review "Investigating the 'Flow' Experience." Confidence: Verified (methodological limit). Caveat: No single instrument cleanly identifies in/out of flow; operationalisation is criticised as imprecise.
Challenge-skill balance is shakier than the textbook version. Løvoll & Vittersø (2014), Social Indicators Research, found that neither flow indicator peaked at challenge-skill balance and that the data support an imbalance model. Engeser & Rheinberg (2008) found flow not always optimised by balance, and Abuhamdeh & Csikszentmihalyi reported challenge-enjoyment relations as unstable across people and activities. Source: Løvoll & Vittersø (2014), SIR; Engeser & Rheinberg (2008). Confidence: Verified (limit). Caveat: Sits alongside Fong et al.'s (2015) meta-analytic finding of a moderate balance-flow link — the antecedent matters but is not the decisive switch the early theory implied.
Self-Determination Theory has a cultural-universality critique. Hagger et al. (2013) found that collectivist-culture participants sometimes show higher intrinsic motivation under authority direction, suggesting that autonomy's primacy may reflect Western individualism. SDT defenders reply that autonomy ≠ independence. Methodological critiques include social-desirability vulnerability in self-report measures and weak psychometrics on the Relative Autonomy Index (Chemolli & Gagné 2014). Source: Hagger et al. (2013); Chemolli & Gagné (2014). Confidence: Verified (limit / contested). Caveat: Prescriptions apply most cleanly in Western individualist markets; flag the qualifier for materially collectivist audiences.
The generation effect has a length ceiling. A 2023 text-generation meta-analysis (Educational Psychology Review) reported Hedges g ≈ 0.41 not attributable to time-on-task, with the effect largest for 301–600 word texts and no effect beyond ~900 words. Source: Educational Psychology Review (2023). Confidence: Verified (meta-analysis). Caveat: Length ceiling is the critical practical limit; quoting the headline g without the ceiling overstates the finding.
The generation effect does not reliably transfer to expository text. A 2025 conceptual replication (Cognitive Research: Principles and Implications) found that the generation effect did not reliably transfer to learning from expository text, and some experiments showed a disadvantage. Source: Cognitive Research: Principles and Implications (2025). Confidence: Verified (limit / scaling). Caveat: The lab-robust word-level effect does not automatically scale to complex real-world content; citations of the d ≈ 0.40 generation-effect meta-analyses (Bertsch 2007, McCurdy 2020) should preserve the word-level scope.
Sales conversations are won on credibility. A quoted stat that does not survive scrutiny loses the room. See Agency methodology for small-business website projects for the broader sourcing discipline.
Operational rules
The discipline above is enforced by four operational rules that govern every research-disciplined public artifact, plus two source-hierarchy rules that govern which magnitudes are quotable in the first place.
Rule 1 — Every objective claim carries a named source + date + verbatim quote + confidence label
Every objective claim in client deliverables, KB entries, and public writing carries:
- A named source (person + institution where applicable).
- A publication or retrieval date.
- A verbatim quote ≤25 words where possible (the link-rot survival layer).
- A confidence label from the seven-label taxonomy.
Why: reader trust default in 2026 is distrust (Edelman 2025 Trust Barometer — 7 in 10 believe leaders mislead). BBC §3.2.2, the Reuters Handbook and the SPJ Code all converge on named-source-preferred. Wikipedia's "verifiability not truth" model is the largest working example. FTC reasonable-basis doctrine makes unsourced objective claims legally actionable. AI engines preferentially cite content with structured citations, named sources and explicit dates. The verbatim quote is the highest-leverage link-rot defense.
How to apply: use the what-to-source checklist as the decision matrix for what needs sourcing. Use the citation pattern library for the format (recommended default: inline parenthetical for research / KB, hover footnotes for long-form, dedicated source page for marketing). The test: if a skeptical reader asks "where did that come from?" and the writer cannot answer in one sentence with a name and a date, the claim needs sourcing. For LLM-assisted drafting: every claim the LLM produces must be traced to a named source before it ships — the Simhi et al. finding is the reason.
Rule 2 — Cite with a named source and URL, confidence flagged honestly
Every quantitative claim, every quoted statement, and every attributed framing carries a citation with: named author or institution, publication date, and verifiable URL. Each claim also carries an explicit confidence flag — Verified (primary source quoted), Industry-consensus (multiple independent corroborating sources), or Single-source (one source only — note).
This operationalizes Wikipedia's verifiability policy for the agency context. The same discipline that makes Wikipedia the #1 cited domain in Google AI Mode makes research-disciplined content extractable, defensible and citable. It also catches errors before publication — the act of finding the URL for a citation is the act of catching the citation that does not exist.
How to apply: every claim node in this KB carries a Source line and a Confidence line. Public-facing articles use inline markdown links with the citation immediately next to the claim. "Studies show…" without a citation is banned under this discipline. Either find the study or drop the claim. Re-using a vendor stat? Disclose the methodology and the original publication date.
Rule 3 — Capture an archive snapshot at citation time
At the moment of citing any web source, capture an archive snapshot (Perma.cc preferred; archive.org Save Page Now as fallback). Include both URLs in the citation. Quote the source verbatim (≤25 words) so the claim survives link death.
Why: NYT link half-life is ~15 years with 13% content drift even on "live" links. Zittrain (Harvard Law 2014): 50% of US Supreme Court opinion links, 70% of Harvard Law Review links broken. A footnote with "source: example.com/article-123" leaves no recovery path when the URL dies; a footnote with a verbatim quote remains searchable in archives forever.
How to apply: the workflow is cite → snapshot (one click in Perma.cc browser extension) → verbatim quote into the citation. Format: "...[verbatim quote]..." (Source, Date — original URL | archived). For research-brief and KB entries: archive every primary source on first use. For marketing pages citing sources via the KB: the KB entry is the archive layer; marketing pages link to the KB. Quarterly link-checker pass per the nine-step plan above. Self-authored research artifacts get DOIs via Zenodo (free).
Rule 4 — Publish a public corrections log and retraction policy
Every public artifact (website, KB, client deliverable) is governed by a public corrections log + retraction policy following the four-magnitude playbook. Material corrections trigger a specific editorial-process change — never a silent edit pretending the original was correct.
Why: Trust Project / Reach Plc showed trust in The Mirror jumped 8 points after adding Trust Indicators including a corrections protocol. COPE August 2025: "The purpose of retraction is to correct the literature... not to punish the authors" — the right disposition to import. Edelman 2025: 7 in 10 readers expect to be misled — visible corrections are the differentiation signal. AP / Trust Project standards make corrections visible as a feature, not a bug.
How to apply:
- Publish a public Corrections log linked from the site footer.
- Categorize corrections by playbook magnitude (typo / substantive / material / fundamentally wrong).
- Update
dateModifiedin structured data on every correction. - For material corrections: top-of-page notice that persists for at least 90 days.
- For retractions: original URL serves the retraction notice; original text preserved in archive.
- Process discipline: every material correction traces to a specific editorial-process change. New check, new source-verification step, new automated guard. Document it in the corrections log.
- Quarterly transparency report: corrections issued, sources re-verified, link-rot rate in the KB.
Rule 5 — Prefer peer-reviewed / award-vetted magnitudes over vendor-recycled figures
When a magnitude figure is available in both a peer-reviewed or award-vetted source and a vendor-recycled secondary form, the peer-reviewed or award-vetted source is preferred — even when the secondary is more rhetorically convenient.
Why: Award-vetted sources (e.g., the INFORMS Edelman Award) and primary academic sources (NBER working papers; foundational economic-theory papers like Stigler 1961 and Akerlof 1970) carry their qualifying conditions intact; vendor-recycled versions tend to strip them. The source-incentive asymmetry is well-documented across the SMB-advice cluster.
How to apply: the default sourcing hierarchy is (1) peer-reviewed primary, (2) award-vetted (Edelman or equivalent), (3) multilateral / Crown-corporation with named incentive flag, (4) trade press citing named executives, (5) vendor self-report — only with explicit attribution and only as direction, never as magnitude. Material below level 5 is dropped.
Rule 6 — Quarantine recycled retention magnitudes until primary-sourced
The figures "5×–25× cheaper to retain," "5% retention → 25–95% profit," "80% of profits from 20% of customers," and "AI churn prediction → 20–30% retention improvement" are not cited under this discipline until the primary source has been located and the original conditions re-read.
Why: These four magnitudes circulate vendor-to-vendor across the analytics blogosphere without traceable primary citation. The 25–95% figure ultimately traces to Reichheld / Bain-era work, but the qualifying conditions have eroded through the citation chain. The 5–25× has a frequently-misattributed primary source. Treating them as facts inherits the credibility erosion of the chain.
How to apply: when a draft uses any of these magnitudes, the figure is replaced with a capability claim (e.g., from the INFORMS Analytics Magazine churn / binary-classification literature) or dropped. If the primary source is later located and re-read, the claim is restated with the original's qualifying conditions intact.
Sources and confidence
| Item | Source | Confidence |
|---|---|---|
| 7-label confidence taxonomy | Candid Creative formal source-of-truth, 2026 | Verified (canonical) |
| 8-pattern citation library | Candid Creative, 2026 | Verified (canonical) |
| What-to-source checklist | Candid Creative, 2026 | Verified (canonical) |
| NYT 15-year link half-life, 13% content drift | Wikipedia synthesis citing 2021 NYT-link study — https://en.wikipedia.org/wiki/Link_rot | Industry-consensus |
| Zittrain reference rot — 50% Supreme Court links, 70% Harvard Law Review links | Zittrain, Albert and Lessig, Harvard Law Review, 2014 | Verified |
| 9-step link-rot mitigation plan | Candid Creative synthesis from Zittrain + NYT empirics, 2026 | Verified (canonical) |
| 4-magnitude retraction playbook | Candid Creative adaptation of COPE + Trust Project + AP standards, 2026 | Verified (canonical) |
| COPE retraction purpose ("correct...not punish") | Committee on Publication Ethics, retraction guidelines, August 2025 update — https://publicationethics.org | Verified |
| BBC §3.2.2 "well sourced, sound evidence, thoroughly tested" | BBC Editorial Guidelines, 2019 ed. — https://www.bbc.co.uk/editorialguidelines/guidelines/accuracy | Verified |
| BBC §3.3.4 single-source reluctance / §3.2.3 attribution | BBC Editorial Guidelines, 2019 ed. | Verified |
| SPJ "identify sources clearly" | SPJ Code of Ethics, 2014 revision — https://spj.org/spj-code-of-ethics | Verified |
| Washington Post two-source norm | Washington Post, December 2017 | Industry-consensus |
| Reuters Handbook — anonymous sources weakest | Reuters Handbook of Journalism, last comprehensively revised 2012 — https://handbook.reuters.com | Verified |
| Reuters Trust Principles (1941; reaffirmed April 2008) | Thomson-Reuters merger documentation | Verified |
| HawkSearch vendor-against-interest cutoff (a few dozen products) | HawkSearch, "If They Can't Find It, They Can't Buy It," accessed June 2026 | Industry-consensus (vendor-against-interest) |
| Prefixbox vendor-against-interest cutoff (<200 products) | Prefixbox, "Faceted Filtering and Faceted Search: Complete Guide," accessed June 2026 | Single-source / vendor |
| Luigi's Box vendor-against-interest cutoff (few hundred products) | Luigi's Box, "Faceted Search: What Is It and Why Your E-Shop Needs It," accessed June 2026 | Industry-consensus |
| NN/g — faceted navigation interaction cost + metadata maintenance | NN/g via Taylor & Francis, "Musings on Faceted Search, Metadata, and Library Discovery Interfaces," 2023 | Verified |
| Wikipedia verifiability not truth + four claim types | Wikipedia WP:V — https://en.wikipedia.org/wiki/Wikipedia:Verifiability | Verified |
| Wikipedia verifiability — companion inline-citation policy text | Wikipedia WP:V — https://en.wikipedia.org/wiki/Wikipedia:Verifiability | Verified |
| US Intelligence Community High/Moderate/Low taxonomy | ICD 203/206; 2007 NIE Iran: Nuclear Intentions and Capabilities; NPR + Harvard Belfer | Verified |
| Tetlock superforecasters / 10 commandments | Tetlock & Gardner, Superforecasting, 2015 | Verified |
| CIS Confidence adaptation for cyber threat intel | Center for Internet Security | Verified |
| Zuegel epistemic-status as publishing-threshold lowerer | Devon Zuegel personal essays | Verified |
| Krycho framing — humility tool | Chris Krycho personal essays | Verified |
| FTC reasonable-basis doctrine — "the amount and type of substantiation the ad actually communicates" | FTC Policy Statement on Advertising Substantiation, 1984; FTC training materials — https://ftc.gov | Verified |
| FTC "Studies Show / Tests Prove / Doctors Recommend" trigger phrases | FTC training materials — https://ftc.gov/sites/default/files/attachments/training-materials/substantiation.pdf | Verified |
| Xlear v. FTC challenge, citing Loper Bright (2024) | Litigation status as of May 2026 | Single-source / under review |
| Canadian portability — Competition Act general impression test | Competition Bureau; CRTC; provincial consumer-protection statutes | Industry-consensus |
| Google retired ClaimReview rich results (June 2025) | Poynter (IFCN), June 2025 — https://poynter.org/ifcn/2025 | Verified |
| ClaimReview schema persists for non-Google AI ingestion | Bill Adair, Duke Reporters' Lab, via Poynter June 2025 | Verified |
| Simhi et al. — "Models can hallucinate with high certainty even when they have the correct knowledge" | Simhi, Itzhak, Barez, Stanovsky, Belinkov — arXiv:2502.12964, February 2025 — https://arxiv.org/abs/2502.12964 | Single-source |
| Broader LLM-hallucinate-confidently finding | Hallucination literature, multiple 2024-2025 sources | Industry-consensus |
| Vectara hallucination leaderboard (>50% on person fact recall) | Vectara, 2025-2026 | Industry-consensus |
| "34% / MIT / January 2025" institutional attribution | Circulating in SEO/marketing blogs; not located in primary source | Single-source / Contested |
| HubSpot 88% "less likely to return after bad experience" | HubSpot original; repeated by Hyperping and other maintenance vendors | Single-source (vendor original) |
| "McKinsey 83% B2B value transparency over brand reputation" | Repeated by Togai / DealHub; not located in primary McKinsey publication | Single-source / vendor-laundered |
| "Princeton GEO 30–40% higher visibility" vendor echoes | Cross-vendor observation, 2024–2026; primary is Aggarwal et al. KDD 2024 | Single-source (one primary, many echoes) |
| Recycled customer-retention magnitudes (5×–25×, 25–95%, 80/20, AI churn 20–30%) | Vendor-recycled across analytics blogosphere; primary chain to Reichheld / Bain-era work eroded | Single-source / vendor-recycled |
| Flow measurement contested — 2025 systematic review | Wonders 2025, Human Behavior and Emerging Technologies; Løvoll & Vittersø 2014; PMC 2020 review | Verified (methodological limit) |
| Challenge-skill balance shaky at the strong end | Løvoll & Vittersø (2014), Social Indicators Research; Engeser & Rheinberg (2008) | Verified (limit) |
| SDT cultural-universality critique | Hagger et al. (2013); Chemolli & Gagné (2014) | Verified (limit / contested) |
| Text-generation effect ceiling beyond ~900 words | Educational Psychology Review (2023) meta-analysis | Verified (meta-analysis) |
| Generation effect doesn't transfer to expository text | Cognitive Research: Principles and Implications (2025) conceptual replication | Verified (limit / scaling) |
| Stats not supported by evidence catalogue (1-11) | Candid Creative synthesis, 2026 | Verified (canonical) |
| Rule — every objective claim sourced with confidence label | Candid Creative operational rule, 2026 | Verified (canonical) |
| Rule — cite with named source and URL | Candid Creative operational rule, 2026 | Verified (canonical) |
| Rule — capture archive snapshot at citation time | Candid Creative operational rule, 2026 | Verified (canonical) |
| Rule — publish corrections log and retraction policy | Candid Creative operational rule, 2026 | Verified (canonical) |
| Rule — prefer peer-reviewed / award-vetted over vendor-recycled magnitudes | Candid Creative operational rule, 2026 | Verified (canonical) |
| Rule — quarantine recycled retention magnitudes until primary-sourced | Candid Creative operational rule, 2026 | Verified (canonical) |
Linked standalone entries kept outside this page
- Agency methodology for small-business website projects — the original research-discipline brief this page extends.