Methodology: how Evidencing answers

Sources

We search PubMed (MEDLINE), Europe PMC and OpenAlex, three open bibliographic databases covering the biomedical and health sciences literature. We use each study’s title, abstract and publication data; the results of the three databases are merged and duplicates removed.

Search

The AI turns your question, written in any language, into its PICO elements (population, intervention or exposure, comparison and outcome) and into several searches with MeSH terms and synonyms: a precise one with the whole question, a broader one without the setting or population, one for systematic reviews and meta-analyses, and one for outcomes and adverse effects.

Study selection

From up to 100 studies found, the AI reads each abstract, identifies its design and sample size and scores from 0 to 10 whether it answers the question. Unrelated ones are discarded, and the answer is written from the most relevant (up to 20), preferring the strongest designs.

Answer and citations

The answer may only use the selected studies, and every claim carries its study’s citation (author and year). Citations to studies that were not among those found are removed before the answer is shown, and each study’s supporting sentence is checked word for word against its abstract.

Strength of evidence

Each finding gets a score from 1 to 10 for how many studies support it, what kind they are and whether they agree. The strongest design sets the ceiling: meta-analyses and systematic reviews up to 10, randomised trials 9, cohort and other observational studies 7, cross-sectional studies 6, case reports and narrative reviews 4, animal and in vitro studies 3.

Consensus meter

For yes/no questions, it shows what share of the studies that answer say yes, possibly, mixed or no. It counts studies without weighting them by quality, which is why it is read together with the strength of evidence.

Certainty and agreement

Certainty (high, moderate, low or very low) sums up how far the conclusion can be trusted, and agreement shows whether the studies agree, give mixed results, contradict each other or are insufficient.

Limitations

Answers are based on abstracts, not on the full text of the studies, and are written by an AI that can be wrong or oversimplify. They do not replace a systematic review or clinical judgement, and they are not medical advice. When there are not enough studies for a question, Evidencing says so instead of making up an answer.

Who is behind it

Evidencing is the free tool of SysRevAI, the open-source (AGPL-3.0) platform for AI-assisted systematic reviews in health. In SysRevAI the same answers come with follow-up questions, a deep mode that reads the open-access full text of the studies, and export to PDF and Word.

Developed by Octavi Rodríguez · sysrevai.com

Privacy

No account is needed to ask. The question is stored anonymously to show you the answer and deleted after 30 days unless you move it to a SysRevAI account. Of your IP address only an anonymous fingerprint (hash) is kept for the daily limits, and we use technical cookies only.