Methodology

We measure what a vendor can demonstrate about its own testing and its own business, using material the vendor or a third party has published. We do not measure product quality, and we could not: nobody can assay a vial they have not received. What we can do is record, precisely and with dates, whether a vendor publishes enough for anyone to check it — and whether what it publishes holds up when checked.

Every factual field in this index is stored as a sourced fact rather than a bare value. Each carries the type of source, the URL we saw it at, the date we last confirmed it, and a flag for whether we corroborated it. That structure is the product. A list of purity percentages with no provenance is not a record of anything.

Every factual field carries a provenance credit that multiplies whatever merit the value itself has. Published information is taken at face value: a vendor stating its own address, pricing or payment methods, or publishing a certificate for a lot, is recording something observable, not making an unverifiable claim about a third party.

• Sourced — recorded with a checkable source we can point to. Credit 1.00.
• Unsourced — recorded without a checkable source. Credit 0.00. Such records also fail our build-time validation and are not supposed to ship.
• Not recorded — we hold nothing. Treated as unknown: left out of the score rather than counted as a zero, and it reduces coverage.

The distinction between unsourced and not recorded is deliberate. An unsourced field counts in full and earns nothing, so deleting a source can never raise a score — it costs. A field we simply do not hold is a partial gap: it counts at half weight, so a vendor that publishes what others leave blank scores higher, while a blank still costs something. And a whole criterion we could not assess at all — track record, when no independent reviews are locatable — is dropped and the remaining weights renormalised, so an unmeasurable criterion does not push every vendor down. What the score does not rest on, the coverage figure beside it makes explicit.

Coverage is a second, separate number, published beside every score. It is the weighted share of scored fields that we hold and corroborated:

coverage = Σ(signal weight × provenance credit) ÷ Σ(signal weight)

A sourced field contributes its full weight, and a missing or unsourced field contributes nothing. Coverage says nothing about whether the answers were good — only about how much of the profile the score rests on. A vendor can score badly on high coverage, which means we know a great deal and most of it is unflattering. A vendor can also score respectably on low coverage, which means the number is thin. Both readings are legitimate; neither is available if you publish only one figure.

We never show a rank without its coverage. Not on the home page, not in a table, not in the exported dataset.

Every data point carries the date we last confirmed it. A point older than 90 days is marked stale wherever it appears, flagged in our build-time data-health report, and retains 90 percent of its credit until re-checked.

The modest penalty is deliberate. Staleness is our failure to re-check, not the vendor’s failure to publish, so it reduces our confidence rather than condemning the vendor. The visible marker, not the arithmetic, is doing most of the work.

Every score on this site is computed. There is no path by which an editor types a number into a vendor’s scorecard and publishes it as though it were measured.

A score changes for one of three reasons: a data point was corrected, a data point gained or lost a checkable source, or the rubric itself changed and every score was recomputed. Nothing is edited silently.

The exported dataset carries computedScore and finalScore as separate fields so that a downstream consumer can always see whether the two have diverged.

Some facts have no URL to cite because the finding is an absence: a page that does not exist, an email that went unanswered, a certificate archive that no longer resolves. These are recorded as editorial_check data points, with a note describing what we did and when.

This is the only source type permitted to lack a URL, and it exists so that negative findings can be recorded honestly instead of being left blank. Every other source type must carry a link, enforced by validation that fails the build.

• We do not routinely commission our own assays. Most certificate records here are vendor-published documents we located and read, and are recorded as published. Where an assay was independently commissioned the record says so.
• We do not evaluate substances. We publish no dosing information, no protocols, and no health-outcome claims. Compound pages describe what published literature exists; they do not instruct use.
• We do not rate customer service, shipping speed, or packaging except where they appear in the volume and sentiment of independent reviews.
• We do not hand-enter scores. Every figure on a vendor profile is computed from the data points listed beneath it, on the weights published above.