What Fact-Auditing Content Costs Per Page (and Per Claim)

Fact-audit pricing goes wrong at the unit. Pages and word counts are how writing is priced, so buyers assume auditing prices the same way — but verification labor scales with checkable claims, and claims are distributed wildly unevenly across a library. This page prices the audit from our own committed ledgers: 445 claims checked on one build, 461 figures recomputed on another, and the one catch that justified all of it [our data].

What unit does audit cost actually scale with?

Claims — a claim being any statement a skeptical reader could check: a statistic, a price, a date, a derived figure, a legal assertion. Our site-wide pass on the insurance build worked through 445 claims across the 167 pages live at audit time, an average density of about 2.7 checked claims per page [our data].

The average hides the budget's real shape. Definitional and conceptual pages carry a handful of claims; figure-dense cost tables and regulation pages carry dozens each. That spread is why flat per-page audit pricing mismeasures in both directions — overcharging the glossary, undercharging the money pages — and why the model below prices claims and lets density set the page cost.

What did our audits check, and what did they catch?

Three builds, three ledgers — told in full in the fact-audit field report; the counts here are the cost-relevant part [our data]:

Two of those numbers set the economics. A 9% correction rate on already-shipped claims is the honest base rate our own content earned before auditing — the reason the stage exists at all. And 1 error in 461 recomputed figures shows what scripted verification buys: the recompute ran by code, which is why its per-figure cost collapses in the next section.

What does the math look like per claim and per page?

Set an illustrative verification rate — ours below is 5 claims per hour, covering tracing the cited source, re-reading it, and recomputing any math; replace it with your own — and the model is three multiplications [our data]:

Derived figures break the model in your favor. The 461-figure recompute on our auto-finance build ran by script — every payment figure computed by code and compared to the page — so the marginal cost per figure is near zero and the real labor is writing the recomputation once, plus investigating the 1 mismatch it surfaced [our data]. The audit budget should therefore split: script everything computable, and spend the per-claim hours where judgment is required — sources, quotes, and legal assertions.

One reconciliation, since our own per-page cost model illustrates a 2-hour audit stage per new page: these price different operations. The 2-hour stage audits a page at drafting depth — sourcing decisions included — while the per-claim model prices re-verifying shipped claims whose sources are already pinned. Budget new production with the stage model and retroactive audits with the claim model.

Why does one caught hallucination pay for the whole audit?

Because the downside of a published fabrication is not symmetric with the cost of finding it. The worst catch in our ledgers was a fully hallucinated state tax regulation on our car-leasing build — cited to a real source that says the opposite, sitting one step from publish, reading like the best-sourced sentence on its page [our data]. It was caught by refute-first auditing: trying to disprove each claim from the page's own citation rather than reading for plausibility.

Price the counterfactual qualitatively, because no honest dollar figure exists for it: a fabricated regulation published on a YMYL finance page is a trust liability with your name on it, quotable by any answer engine that picks the passage, and discoverable by exactly the skeptical readers you most need. Google's quality guidance frames the bar as reliability and first-hand depth (Google, helpful content guidance); a single catch like that one buys more of both than every clean check combined. That is the audit's actual product — the clean checks are the receipt, the caught fabrication is the return.

When is a full audit the wrong spend?

When it is priced flat against pages that carry no claims — and when it replaces, rather than follows, gated production. Audit depth should track claim density: minutes for a definition page, real hours for a figure-dense table. And an audit bought once, on content produced with no source discipline, buys a snapshot — our corrections concentrated in pages written before our content contract existed, while contract-governed pages audited clean [our data]. The process that makes audits cheap is the subject of the fact-audit process guide.

Against our own interest as sellers of exactly this work: if your library is small and your pages carry few checkable claims, a disciplined writing process plus scripted recomputation may be all the audit you need. The per-claim math above is the honest way to find out — it is the same arithmetic we run before quoting anyone, and the same standard every figure in this library is held to.