Things we published that turned out to be wrong, and what we did about them. Each entry says what the error was, what it affected, and how long it was live.
Our nattokinase page cited a MedlinePlus entry that has 404ed since MedlinePlus retired its natural-products database in July 2025 — the citation was dead while the page still claimed the backing. The dose, evidence and safety text is now sourced to two NIH-hosted peer-reviewed reviews, and the bleeding caution is labelled for what it is: mechanism-based, not trial-based.
Every source listed on a nutrient hub was attributed to the NIH Office of Dietary Supplements, including the ones published by NCCIH, LiverTox, StatPearls and PubMed Central. Publishers are now named correctly, on the page and in the structured data.
The author byline on 97 guides linked to a GitHub profile that does not exist, and the same dead URL was published as the author’s verifiable-identity link in our structured data. The link has been removed rather than pointed somewhere convenient.
Our third-party-testing guide linked to a USP registry page that has since been taken down. It now points at the live USP Verified product listing. Every source we cite is checked against a script on a schedule so a dead citation gets caught instead of ageing quietly.
Combination products were ranked using the single label dose stored against the listing, which belongs to whichever ingredient the product is built around. That put an L-tyrosine bottle at the top of the cheapest-omega-3 ranking, and cayenne pepper at the top of berberine — 55 of 194 shelves had a wrong leader. Rankings now only use a dose the shelf’s own nutrient owns.
Our price-spread study headlined a 2,556x spread for vitamin D3. Both endpoints were listings later quarantined as scrape glitches, so the figure was wrong the moment they were removed — the real spread was 107x. The study, the guides quoting it and the press materials were rewritten onto p90/p10, which one bad row cannot move.
The listing parser read “Pack of 2” as a serving count, so 24 listings showed impossible per-serving costs — and those rows had already fed the published bottle-size and gummy-premium studies (86% became 80%, 2.1x became 1.9x). The rows were quarantined, the parser fixed with regression tests, and the studies recomputed.
Non-error changes — new pages, coverage, method updates — are on data updates. Our sourcing rules are in the editorial standards.