Methodology

Every number on this site is computed directly from the NIH Dietary Supplement Label Database (DSLD), as of the January 2026 dump (214,780 labels ever ingested, 121,959 currently on-market).

Ingestion

We download DSLD's full-database JSON dump directly (confirmed link on dsld.od.nih.gov/api-guide) and parse every label into a relational schema: products, brands, ingredient rows, proprietary blends, other ingredients (excipients), and label statements. The last dump parsed 214,780 labels with a 100% parse success rate.

Normalization

Raw label ingredient names and forms are matched to a curated set of 69 canonical compounds and their common forms via exact alias matching, regex qualifier-stripping, and fuzzy matching (rapidfuzz, ≥92 similarity). Anything that doesn't clear that bar is left unmatched and routed to a review queue rather than guessed. The same pipeline matches excipients against a curated alias table of 161 common carriers, colors, and sweeteners.

Elemental amount conventions

When a label declares an ingredient by its bare element/vitamin name (the common case, e.g. "Magnesium"), we treat the declared amount as already elemental, per longstanding FDA labeling rules. When a label declares the salt itself as the ingredient name (e.g. "Magnesium Oxide"), we convert using that form's elemental fraction (computed from molecular weight stoichiometry, with the assumed hydration state documented per form). Blend members with an undisclosed amount are recorded as such — never estimated. Full conventions, including unit conversions (IU→mcg for vitamin D, DFE for folate), are documented in our reference data.

Serving columns

Some labels declare amounts for more than one serving size in separate columns (for example "2 tablets, ages 2-3" and "4 tablets, ages 4+"), sometimes as a separate ingredient row per column. Adding them together would count the same nutrient twice, so per-product dose figures use only the rows from the label's first serving column; rows from the other columns stay matched in the database but add nothing to the dose. The one exception is the Formula Twins comparison, which matches every matched ingredient row on a label together with its first serving size.

Form quality tiers

Tier 1 (best-evidenced bioavailability), Tier 2 (acceptable), and Tier 3 (low-cost/low-bioavailability) ratings are based on published comparative bioavailability or clinical-outcome literature specific to that form, cited on each compound's "forms explained" section. Where no comparative study was found, the form is left unrated (null) rather than guessed.

What DSLD is and isn't

• DSLD is survey-sourced plus voluntary manufacturer submissions — it is not a market census, and coverage of any given retailer or brand is not guaranteed to be complete.
• Off-market status can lag real-world retail availability.
• Every value is manufacturer-declared on the label as submitted to DSLD, not independently lab-tested or verified by us or by NIH.
• Statement-type bucketing (claim/warning/identity/other) is a heuristic based on DSLD's own descriptive type strings, not a legal classification.

How this site is made

Supplement Label Index is published by its operator, Alykhan Virani, and built with software and AI assistance. Being plain about which part is which:

• Computed by software, not written by AI. Every count, percentage, median, percentile, ranking and chart comes from scripts that read the NIH DSLD release and write the result straight into the pages. Charts are drawn by code from that data; no photographs or AI-generated illustrations are used.
• Written with AI assistance. Explanatory text around the numbers, glossary entries, blog posts and the reference data behind the compound and form pages (reference doses, upper limits, quality tiers, studied doses and their citations) were researched and drafted with Anthropic's Claude. In that process the model fetched the cited sources (PubMed abstracts, NIH and FDA pages, the eCFR) and was instructed to leave a field empty rather than guess.
• A second, independent check. In October 2026 separate AI agents, told to start from each claim and the primary sources rather than the existing citations, re-checked the glossary and the reference data for all 69 compounds: 842 claims in total. 616 (73%) held up, 36 (4%) were wrong, 127 (15%) were imprecise and 63 (7%) could not be verified. Every one of the 232 cited PubMed IDs exists, and every official RDA and upper limit matched its source; the errors were in citations (wrong PubMed ID or author), in claims that no human comparison of two forms existed, in dose ranges the cited papers did not support, and in glossary examples that had gone stale. All wrong and imprecise items were corrected (see the changes page). Those rates describe the drafts before correction, and this check was itself done by AI.
• What has not been done. No physician, pharmacist or dietitian has reviewed the content, and a person has not read every sentence against its source. When a reader or we find a mistake it is corrected and logged on the changes page.
• Pages built from the data. Product pages republish a DSLD record and add where its declared doses sit in the market. Pages with nothing to add beyond the record are marked noindex and are left out of the sitemap.

Spotted something wrong? Write to supplementlabel@0a0.info.

No health advice

Supplement Label Index does not rate, score, or recommend products, and nothing on this site is medical or nutritional advice. It reports what labels declare and what the market does. Consult a qualified healthcare professional for guidance on your own supplementation.

Refresh cadence

Data is refreshed monthly: a new DSLD dump is ingested, diffed against the prior month, and all metrics and reports are recomputed.