Fact sourcing: repair unsupported claims for AI search
Fact sourcing is a claim-by-claim repair: inventory the material statements on one existing page, attach the most direct authoritative source available, record exactly what that source supports, and retain, narrow, rewrite, or remove each claim. Keep visible copy, links, dates, and structured data aligned. Acceptance is complete when every retained material claim passes that support check; a later AI citation is not required. Site traffic is a separate outcome and is not required to complete this fix.
What is the scope of a fact-sourcing fix?
Use this fix after a citation-gap analysis identifies one page whose material facts are missing support, exceed their sources, use stale versions, or hide the relevant evidence from readers. Bound the work to that page, its visible claims, its structured data, and a dated claim-to-source record.
Sourcing can make a claim inspectable and correct. It does not reveal an answer system's selection formula, make markup an AI citation signal, or guarantee crawling, indexing, ranking, a mention, a recommendation, a citation, a referral, or traffic.
Which prerequisites must be attached before editing?
Capture the before state and define what counts as a material claim before changing the page.
- Canonical page URL, saved before version, visible text, metadata, structured data, publication dates, and every current outbound source link.
- The exact claim or field identified by the Find record, its page location, and the observed source comparison that motivated inspection.
- Claim types and source-selection rules: first-party product facts, public records, standards, research findings, derived calculations, estimates, and opinions are not interchangeable.
- A source owner, current version or effective date, checked date, support scope, correction path, deployment check, and rollback plan.
- The matched prompt-and-surface cells reserved for later proof, without making their outcome an acceptance condition for this edit.
How should sources match different claim types?
Choose the source closest to the fact and preserve the boundary between a source statement and your inference.
| Claim type | Preferred evidence | Required boundary |
|---|---|---|
| First-party product fact | Current product documentation, changelog, or directly inspected behavior | State plan, version, date, and inspection scope |
| Law, policy, or public record | The issuing authority's current text or dataset | State jurisdiction, effective date, and exceptions |
| Technical standard | The standard or steward's specification | State version and distinguish requirements from implementation advice |
| Research result | Original paper, report, and accessible data or method | State population, collection date, denominator, and limits |
| Derived calculation | Visible inputs plus the declared formula | Show arithmetic, exclusions, rounding, and owner |
| Opinion or recommendation | Named author and disclosed reasoning | Do not relabel judgment as measured fact |
What does the worked synthetic claim-to-source artifact contain?
SYNTHETIC EXAMPLE — fictional brand, surfaces, sources, and observations; not RankEcho data, customer data, or a provider result.
Northstar S1 is fictional. Its source names are plain-text placeholders, not live evidence. The bounded after block keeps the supported interval and protocol facts, adds the documented exception, and removes the unsupported superlative.
| Claim ID and before text | Fictional source and exact scope | Action | Bounded after text |
|---|---|---|---|
| F01 · Calibration interval: 12 months | Synthetic Manual v3 §4.2 · only under standard conditions | Qualify and cite | Under standard conditions, the fictional S1 manual specifies a 12-month interval. |
| F02 · No shock exception stated | Synthetic Bulletin B-2026-04 §2 · recalibrate after shock | Add exception | Recalibrate after a shock under the fictional bulletin's stated conditions. |
| F03 · Most accurate sensor available | No market, measure, period, or independent comparison | Remove | Claim removed; no replacement superlative. |
| F04 · Works with every control system | Synthetic Manual v3 §7 · Modbus TCP and OPC UA only | Narrow and cite | The fictional S1 supports Modbus TCP and OPC UA. |
How do you rewrite a claim to match its source?
Open the source, locate the exact supporting passage or data, and test subject, predicate, quantity, population, time, geography, version, and exceptions. Quote only when the wording matters; otherwise paraphrase without broadening. Label inferences and derived calculations as such.
If the source supports only part of a sentence, split the sentence. If the source is secondary, volatile, inaccessible, or contradicted by a more direct current record, find a better source or narrow the claim. If no adequate source exists, remove the material factual assertion.
How must visible copy and structured data stay aligned?
Readers must be able to see the supported claim, its relevant qualifier, and the link or source context. Metadata, FAQ text, comparison tables, accessibility labels, and JSON-LD must not introduce a broader version of the claim than the rendered page.
Use structured data only when it represents the visible content and applicable type. Valid markup can support eligibility for applicable rich results; it is not a special AI citation mechanism and does not guarantee display.
Which deterministic validation checks must pass?
Complete the fix when the page and its evidence record pass these checks. Do not wait for organic crawler traffic, a citation, a referral, or a RankEcho outcome threshold.
| Required check | Pass condition | Evidence to retain |
|---|---|---|
| Claim coverage | Every retained material factual claim has one or more adequate sources | Claim identifier, page location, source URL, and action |
| Exact support | Subject, quantity, scope, date, version, and exceptions do not exceed the source | Supporting passage or data location and checked date |
| Source delivery | Every visible source link resolves to the intended authoritative record | Final URL, status, title, and checked time |
| Presentation parity | HTML, metadata, FAQ, accessibility text, and structured data make no broader claim | Rendered checks and structured-data diff |
| Correction safety | Removed and narrowed claims disappear from all public surfaces and can be restored from the saved before version | Search results, route render, diff, and rollback version |
Which false positives and risks should stop the fix?
Stop when the proposed source merely repeats the claim, describes another plan or date, supports only a neighboring concept, or is cited because a competitor page uses it. A link is not support unless the opened record matches the exact assertion.
Do not add citations decoratively, hide qualifications in footnotes, publish fabricated statistics, create schema-only claims, or turn provider guidance into a universal selection rule. Preserve uncertainty, conflicting primary records, and source limitations where they affect meaning.
What is the rollback and correction plan?
Save the before version and reverse the deployment if a source link is wrong, a material qualifier disappears, markup exceeds visible copy, the page breaks, or an unrelated claim changes. Restore the last supported version, then repeat the same checks.
When a source changes later, update the claim, checked date, and visible evidence together. Do not refresh a page date without rechecking the material sources that justify it.
How does the fix hand off to proof?
Record the deployment time and exact semantic delta, then collect the pre-registered prompt-and-surface cells under the same protocol. Keep the content acceptance result separate from later answer outcomes.
Report no movement, adverse movement, and unavailable or failed cells. A changed manual cell-level owned-link rate after deployment is descriptive evidence under the matched design, not proof that fact sourcing caused the change.
Who maintains this fact-sourcing guide?
Written and maintained by Abiot Y. Derbie. Published 2026-09-07; sources checked 2026-09-06; last updated 2026-09-07. The worked artifact is synthetic and does not represent a customer page.
Recheck source ownership, effective dates, product versions, visible-copy parity, and every material claim before applying the method. Send corrections through the contact page.
Sources reviewed
Provider eligibility and measurement claims below were checked against primary documentation. These records do not establish a universal selection formula, causation, or a guaranteed ranking, impression, recommendation, or citation.
4 claim-level source records
| Claim reviewed | Official source | Review record |
|---|---|---|
| Google recommends helpful, reliable, people-first content and encourages clear authorship, appropriate production-method disclosure, and a people-first purpose while demonstrating first-hand expertise and supporting trust. | Google helpful-content guidance | Checked 2026-09-06 · Current people-first content and authorship guidance · Primary-source documentation check; these editorial questions guide claim support and transparency but are not represented as a disclosed ranking formula or citation guarantee. · Confidence: High |
| Google requires structured data to represent visible page content accurately and says valid markup makes a page eligible for supported rich results rather than guaranteeing that a feature will appear. | Google structured-data general guidelines | Checked 2026-09-06 · Guidelines last updated July 10, 2026 · Primary-source documentation check; visible-copy and markup parity is an acceptance check, not evidence that an answer system will select or cite the page. · Confidence: High |
| Google says its generative AI Search features use core Search systems and require no special AI file, content chunking, writing style, or AI-specific structured data. Search eligibility does not guarantee display in a generative AI feature. | Google guide to generative AI Search optimization | Checked 2026-09-06 · Google Search guidance updated July 10, 2026 · Primary-source documentation check; ordinary Google Search guidance is not generalized into a formula for other answer products or a guaranteed citation outcome. · Confidence: High |
| OpenAI says ChatGPT responses that use web search may include citations and may expose cited sources and other relevant links in a Sources view. It also warns that search results and citations can be incomplete, outdated, or incorrect and should be checked against the opened source. | OpenAI Help: searching the web with ChatGPT | Checked 2026-09-06 · Current ChatGPT search source guidance · Primary-source documentation check; the observed visible citation is retained as evidence for that response, while source support and answer accuracy are checked separately. · Confidence: High |
Frequently asked questions
No. It makes the page's claims inspectable and bounded. Crawling, indexing, retrieval, selection, citation, referral, and traffic remain separate later observations.
No. Inventory material factual claims and source them according to risk, volatility, specificity, and reader need. First-party experience, derived analysis, and opinion should be labeled rather than decorated with unrelated links.
No. Google requires structured data to represent visible page content accurately. Keep the claim and its qualifiers aligned across HTML, metadata, FAQ text, accessibility text, and JSON-LD.
Split or narrow the claim to the supported subject, quantity, population, date, version, and exceptions. Remove the unsupported remainder or find a more direct source.
No. Accept the content fix when every retained material claim and every public representation passes the declared support checks. Treat later outcomes as learning signals.