Methodology
How this site produces a claim, what it refuses to publish, and what happened when it applied its own citation standard to its own corpus and withdrew 28 pages.
Every figure on this site carries a named source, a date, and a location specific enough to check. In August 2026 that standard was applied to the site's own legacy corpus, 11 sampled figures returned 0 verifiable from the URL cited, and 28 pages were withdrawn.
Most methodology pages describe a standard nobody has tested. This one describes a standard that was tested against this site's own published work, which failed it, and what happened next. That sequence is the only evidence a reader has that the standard is real rather than aspirational.
The standard
The citation standard requires four things of every figure: a named source, a deep link to the actual data table or series, a publication or retrieval date, and the figure quoted as the source states it.
A citation pointing at an institution's front door does not meet it. Neither does a figure attributed to a named organisation with no location inside that organisation's output.
The reasoning is mechanical rather than moral. A citation that cannot be followed cannot have been followed, including by the person who wrote it, because the act of checking would have replaced the front-door link with the deep one. A general citation is therefore evidence that verification did not happen, not evidence that it did.
What happened when the standard was applied here
The citation validity sample covers 11 figures across 4 pages, and it returned 0 verifiable from the URL cited. Four gave a different figure for the same measure. Seven cited a page that does not publish the measure at all.
The legacy corpus carries 465 figures, 292 of them in headline components. The citation validity sample was not designed to condemn it but to estimate it, and a zero return on eleven is not a result that supports selective repair.
28 withdrawn pages were set to noindex on 2026-08-07, rather than deleted or redirected. Redirecting would have pointed them at targets that do not exist. noindex removes search and generative-engine exposure immediately and stays reversible if a page can be rebuilt to standard.
This was a provenance failure rather than a fabrication finding. The figures may well be approximately right. Nobody can tell, which is the problem.
A provenance failure means the process that produced the figures never included verification, so the defect is the process rather than any individual number. That is why the response was a withdrawal rather than a correction pass.
What returning requires
A withdrawn page returns only when its figures meet the standard above, one figure at a time. A page is not repaired into qualifying. Where enough of its figures cannot meet the standard, it retires with the rest.
Four pages were marked for verification on the basis that they were worth the effort. That selection is explicitly a hypothesis about what is salvageable rather than a conclusion, and the verification pass is its test.
What gets refused
Three refusals are enforced by the build rather than by intention, because a rule with no gate is a preference.
Figures that cannot be traced. Where a number is widely repeated and no primary source survives tracing, the page records it as an absence claim: the figure, who it is commonly attributed to, and what tracing actually returned. The claim is not published as a fact and the tracing result is published instead. The benchmarks page is built almost entirely this way.
Claims that cannot be established from the evidence in hand. These are declared per page in an unverifiedClaims field naming the claim and what would resolve it. A page may make an argument that rests on one; it may not present it as settled.
Direction stated as magnitude. Where an effect is well supported and its size is not, the direction is stated and the number is refused. The alternative is a plausible figure that survives repetition, and a figure that survives a check is not thereby closer to correct, it is only harder to disprove.
Corrections and what does not change quietly
The correction policy states that a published number is never edited silently. Where research disagrees with a live figure, the page is left unmodified and the discrepancy is logged for per-figure review with both readings and the confidence behind them.
The reason is that a number changing without a visible reason is worse than a number that was always wrong. It makes the site's history unreliable in a way no reader can detect from outside. A figure moving materially also often signals that the original measured something adjacent rather than being wrong, which is a scope judgment and not a sourcing one.
Where a page's central premise turns out to be false rather than imprecise, it is withdrawn rather than corrected. One article reporting on the aftermath of a compliance deadline was found to be describing enforcement of a rule that Congress had barred from being enforced. There was no six-month period to review, so there was nothing to fix.
Review
Every page carries a lastReviewed date, and the date means the page was read against its sources on that date rather than that it was published then.
The currency sweep covers figures with a revision risk: certification revision numbers, compliance dates, and anything attached to a standard that revises. A figure correct on publication and wrong eighteen months later is the most common failure mode in this subject, and it does not announce itself.
How to challenge something here
The useful form is specific: the page, the figure, and either a source that says otherwise or a reason the cited source does not support it.
A challenge that identifies a figure whose citation cannot be followed will be acted on without argument, because that is the standard this site already applied to its own work. Contact routes are on the disclosure page, which also states how the site is funded, since a methodology statement from a site with an undisclosed commercial interest is worth less than one without.
The data handling behind all of this, including the tools, the dates, and how a pull is made repeatable, is set out in data and research methods. The withdrawn corpus and its disposition are in research.
Common questions
Why withdraw 28 pages over an 11-figure sample?
Because the sample returned zero verifiable. A partial failure rate would have supported selective repair. A zero return means the process that produced all 465 figures did not include verification, and the defect is the process rather than the individual numbers.
Are the withdrawn figures wrong?
Unknown, and that is the finding. They may be approximately right. Nothing in the corpus lets anyone tell, which is the condition the standard exists to prevent.
Does this site publish original data?
Yes, and it states its method to be re-run: the tool, the endpoint, the country, the date, the seed set, and the threshold. If a repeat returns something different, the page is wrong and gets corrected rather than defended.
What happens to a claim that cannot be sourced?
It is either published as an absence claim with its tracing result, or not published. It is never published as a figure with a plausible attribution attached.
Who writes this?
Editorial standards, authorship, and review are set out on the editorial standards page, and the commercial arrangement is on the disclosure page.
Sources
- 01
Last reviewed . Published by ManufacturingSEO.ai.