How we grade evidence

Every finding on this site has a letter, a direction and a list of studies. This page says what each of those means and how it is decided.

What gets a grade

We grade outcomes, not supplements. An outcome is one measured result for one supplement, such as “blood pressure” for magnesium. A supplement can have a strong grade for one outcome and a weak one for another, so there is no single score for a supplement.

Each outcome shows four things: its grade, the direction of the effect, the size of the effect when it is known, and the studies behind it.

What A, B, C and D mean

A grade rates how much good evidence there is for a finding. It is not a rating of how large the effect is, and it is not a recommendation. A finding of “no effect” can have grade A, if good trials agree that nothing happens.

Of the 17,415 outcomes in the database, 4,914 have a grade: 480 A, 783 B, 1,758 C and 1,893 D. The rest are listed without a grade because there is not enough human evidence to give one.

Where today's grades come from

The letters shown today were assigned when the database was first compiled from published evidence summaries. We are now re-deriving each letter from the studies cited on the page, using the definitions above. Until an outcome has been re-derived, read its letter as a starting point and check the studies listed under it.

When a re-check changes a grade, the page's “last updated” date changes with it.

Automated readings of the published studies

Most outcomes in the database have no grade yet. For some of them a search result or a supplement page shows an automated reading of the published studies instead, and says so. It is made by the Decosa evidence engine, which is software, not a person. It searches PubMed for the supplement and the outcome, reads the abstracts of the trials and reviews it finds with a language model, and records for each study its design, its size, who was studied and what it found, with a few words copied from the abstract. A published rubric then turns those rows into a verdict (improves, worsens, no clear difference, mixed results, or unclear) and flags any study that reported a harm. Every study listed links to its PubMed record, so you can check it.

An automated reading is not a Dr. Grey grade and never replaces or changes one. Nobody has reviewed it. It reads abstracts, not whole papers, so it cannot judge how well a trial was run, and it can miss studies. Because a model reads the abstracts, running it again over the same studies can give a different answer. So every reading is run twice, and a verdict is shown only when both runs agree on whether there is a benefit, a harm or neither. When they do not, the page says “mixed or unclear: two readings of the studies disagreed”. Today 209 supplement and outcome pairs have a reading. In 194 of them the two runs agreed on benefit, harm or neither, and in 187 they also gave the same verdict word for word.

Where the studies come from

Study references are papers indexed in PubMed, the U.S. National Library of Medicine's index of biomedical research. Each citation links to its PubMed record, so you can read the abstract yourself.

A study count on a page only includes citations that link to a paper. If an outcome has no linked study, no study count is shown for it. Today 8,647 of 17,415 outcomes (50%) have at least one linked study, from 10,895 citations in total. A page where no finding can be traced to a study is kept out of search engines until it can, unless a search engine already lists it. Those pages are first in line to have their studies linked.

How the direction of an effect is decided

“Increase” and “decrease” are not the same as better and worse. A decrease in blood pressure is usually good; a decrease in muscle strength is not. So each outcome is read in two steps.

1. Which way is better for this outcome? Risks, symptoms and markers of damage are better lower. Function, performance and recovery are better higher. A few depend on the condition: body weight going up is a good result in cachexia, where weight loss is the problem. Where we can't tell, we don't guess.
2. What did the studies find? When the outcome has a written description of the effect, the label follows that description. When it only has a stored direction, the direction is treated with caution.

The result is one of three labels:

• Improves: the change is in the better direction for that outcome.
• Worsens: the evidence describes a harm, or a reviewed source does. A stored direction alone is never enough to show this label.
• Changes; see studies: the outcome changed, but the data doesn't say clearly whether that is good or bad. The studies are listed so you can check.

Some rows have been checked by hand against a systematic review or the trial itself. Those carry a reviewed verdict, stored with its source and a quote. There are 7 such reviewed verdicts in the current rules (version 2026-09-30.1).

Effect size

Where the studies report it, an outcome also shows the size of the effect: small, moderate or large. Size and grade are separate. A small effect can be very well established, and a large effect can rest on one weak study.

Known limits

• Not every outcome has a linked study yet. Those outcomes show no study count.
• Grades have not all been re-derived from the cited studies (see above).
• New trials are published all the time; a page reflects the studies it lists, not everything ever published.
• Nothing here accounts for your health, your medicines or your circumstances. It is not medical advice.

Seen something wrong?

Tell us, with the page and the study if you have it. See the corrections policy. The data on this site was last updated on 30 September 2026.