How Digestery rates evidence
Fixed rules turn the studies into the short answer, so the same studies always give the same answer.
The steps
- Search. For each question we search PubMed, the main index of medical research: pooled reviews since 2012, trials in people, population studies and, for some questions, animal studies. We read up to about 50 papers per question, the ones PubMed ranks as most relevant. Large fields are sampled, not covered in full.
- Two independent reads. Claude Sonnet and Claude Haiku each read every abstract. Neither sees the other’s work. Each records the type of study, who took part, what was tested, what was found, the key numbers, and who has a stake.
- Settle and check. Claude Opus 5.5 compares the two reads. Where they differ, it reads the abstract again and decides. A script checks every number against the abstract.
- Write. The top model writes the short answer and findings. It may use only facts from the papers, every claim lists its papers, and it says what the research shows, never what you should do.
- Person check. A person checks the short answer against the sources. Until that happens, the page carries an “Under review” flag and search engines are asked not to list it.
The evidence ladder
Each page sorts its papers into steps. From strongest to weakest: trials in people that measured the question itself; trials that measured other health measures; trials that measured only lab tests or safety; population studies; blood-level and gene studies; animal studies; and studies of cells in a lab.
Reviews that pool or sum up many studies are shown separately, because they repeat studies already counted on the steps. Opinion pieces carry no weight.
Confidence levels
The rules start from GRADE, a standard system for rating evidence, written in plain language.
| Confidence | What it means |
|---|---|
| High | We are confident the answer is right. |
| Moderate | The answer is probably right, but it could change. |
| Low | The answer may well change. |
| Very low | Any answer is very uncertain. |
| Not tested in people | Only cell or animal studies exist. |
- Confidence starts high when trials in people measured the question itself, and low when the evidence in people comes only from population studies.
- It goes down one level for each serious problem: small or biased trials, results that disagree, indirect measures, or wide uncertainty.
- Population studies can go up one level for a large, consistent effect or a clear dose-response, but never above moderate.
- Hard limit: if no trial in people measured the question itself, the short answer cannot say that the thing works in people.
Funding rule
A stake is always shown. Confidence goes down only when every supporting study comes from a party with a stake.
Limits
- Pages are written from PubMed abstracts. Abstracts can leave out details, such as side effects and funding.
- AI models can make mistakes. That is why two models read every paper, a script checks the numbers, and a person checks the short answer.
- If you find an error, please flag it.