Our method
Studies are pulled from PubMed, extracted into structured fields, matched to benefits by a judge model, filtered by a fixed threshold, and graded by a formula.
The process
1
Build the corpus
Each supplement has its own PubMed search, covering its name and its synonyms. The search is repeated over time, so trials published after the first pass are picked up rather than missed permanently.
2
Extract each study into structured data
A model processes one paper at a time and returns fields rather than prose: the number of participants, the outcome measured, the direction of the change, and a strength rating from 1 to 5 with a short justification. A record is saved only when every field is present.
3
Establish the benefit list first
The model is asked what a supplement is claimed to do before it has seen any of the studies. Deriving that list from the search results would introduce confirmation bias, because the benefits would then reflect whatever the search happened to return.
4
Score each study against each benefit
A judge model (LLM-as-judge) rates the relevance of every study to every candidate benefit from 1 to 100 and records a short justification alongside the score. A single paper may support several benefits, or none of them.
5
Apply a fixed threshold
What is published is decided by fixed thresholds rather than by the model itself. A study needs 50 to appear under a single benefit, and 60 when it is matched to several benefits of the same supplement. Studies with no score, or with fewer than 30 participants, are excluded.
6
Grade the claim
The final score is derived from two figures: the number of qualifying studies and the total number of participants across them. Because it is arithmetic rather than a model, the same set of studies always produces the same score.
Where the data comes from
PubMed
NCBI E-utilities
The literature itself: records, abstracts, and full text from PubMed Central where the licence allows. Every claim on this site traces back to PMIDs that you can open and read yourself.
ClinicalTrials.gov
API v2
Recruiting, active and completed trials by condition and intervention, so a sparse evidence base can also show what is currently being tested.
openFDA
enforcement reports
The safety notices: recalls of dietary supplements in the agency's own words, pulled directly from the FDA enforcement feed rather than from press coverage of it.
NIH Office of Dietary Supplements
reference intakes
The official US reference values: the Recommended Dietary Allowance, the Tolerable Upper Intake Level, and the adequate intakes beside them, transcribed from the published tables.
FDA substance registration
UNII identifiers
Each substance is tied to its Unique Ingredient Identifier where the FDA registry holds one, so two spellings of the same compound resolve to a single entity instead of two.
NIH Dietary Supplement Label Database
substance vocabulary
The names that substances are actually printed under, read from the labels the database publishes, so an ingredient page describes what appears on a bottle rather than a name we chose ourselves.
Found a study we read wrong, a match that makes no sense, or a threshold you would argue with? Write to us.