Methodology

How we find the evidence.

Most health content starts with a claim and looks for support. We start with the literature and find out whether the claim survives it.

Searching, not citing

Every topic begins as a question, not an answer. Specialist AI research agents work the literature databases — PubMed and Europe PMC — for the strongest available evidence on that question, and they are looking for a specific thing: a meta-analysis or systematic review that pools the primary studies. Not a single trial, not a press release, not another article citing an article.

What comes back gets compared rather than accepted. Candidate reviews are weighed against each other on design, size and how recently they were run, and the strongest is chosen as the anchor the topic rests on. Then a second agent goes back to the source and tries to break it — re-checking every figure against the paper itself, independently of the one that found it.

Wherever the full text is reachable we read the full text, not the abstract. That is not pedantry: abstracts round, omit, and occasionally contradict the paper they summarise. On one recent topic the abstract stated a threshold that the paper's own methods section and figures contradicted — we published the version the arithmetic supports and said so.

The checks that catch the subtle errors

• Every figure carries its own denominator. A review reporting several outcomes usually pooled a different set of studies for each one. Showing a result beside the review's overall participant count would overstate it, so each number is bound to the sample it actually came from.
• Two reviews of the same literature are not two pieces of evidence. Before any two reviews appear side by side we check whether they share their underlying studies. In one case, 93% of the cohorts behind one review sat inside the other — presented together they would have looked like independent agreement. Only one was used.
• The measured outcome is the claim. A change in a blood-pressure reading or a pain score is reported as exactly that, never inflated into a disease or mortality claim.
• No number is written from memory. Every figure is read off the source at the moment it is used, and links to a resolvable DOI you can open yourself.
• Observational findings say "associated with," never "causes."
• Every fix is free and behavioral. Nothing we publish requires buying anything.

What happens when the evidence isn't there

Often the honest answer is that we searched for the evidence behind a popular piece of advice and did not find it. That is a result, and we publish it rather than quietly picking a different angle.

Three worked examples, each stated as what we searched for and what we found — the honest form of a negative result, and the one we hold ourselves to:

• We went looking for the source of the "60% volume for 60 minutes" headphone rule and could not find one. Every "60/60 rule" we could locate in the peer-reviewed literature is a cochlear-implant referral criterion that shares the nickname and nothing else. The standard that does exist, WHO–ITU Recommendation H.870, is written in decibels and hours — not as a percentage of a volume control, which is a fraction of a maximum that differs between one phone and the next.
• The controlled experiment that set out to test the 20-20-20 screen-break rule found no effect on symptoms, reading speed or accuracy, and its authors wrote that there is relatively little peer-reviewed evidence for the rule. The one positive study we found was uncontrolled, in 29 people, and its benefit had gone a week after the breaks stopped.
• We found no study that has directly tested stopping alcohol three or four hours before bed. The one meta-analysis that set out to quantify the timing of intake reported that its influence could not be established — nearly every trial dosed within about three hours of bedtime, so there was almost no variation in timing left to measure.

Note the shape of all three: we searched and did not find, not it does not exist. The second is a claim about the literature that a search cannot support, and it is exactly the kind of overreach this page exists to rule out.

Six topics have been dropped outright rather than softened — three because no anchor cleared the evidence bar, two on editorial judgement, and one after the video was already made. A topic that fails is dropped, not weakened until it passes.

Two human gates

On top of all of that, a scientist signs off twice: once on the facts, the design and the full script before anything is produced, and again on the finished video before it can be published. A figure that cannot be reached in the source is reported as a gap — never inferred, rounded, or taken from a secondary source.

The role of AI, stated plainly

We use AI throughout: to search the literature, to draft, to generate these pages, and for the narration voice. Every video is labelled AI-narrated and disclosed as such on every platform. You should know that without having to guess.

What AI does not do is decide. No claim, number, source or fix is published without a working scientist reading it against the literature it cites and confirming that the two match. That is an independent evidence check, not a professional endorsement of the advice, and it is deliberately the narrower claim. Editorial responsibility for everything published here and on our channels rests with the site's operator, named in the legal notice.

When better evidence supersedes something we have published, we correct it in the open rather than quietly editing — see corrections. What this content is and is not is set out in the Medical Disclaimer.

Colophon

Narration uses Kokoro-82M, an open-weight text-to-speech model licensed under Apache 2.0. Videos are rendered with Remotion. Type is Archivo and IBM Plex Mono (SIL Open Font License). Kokoro's training data includes small CC-BY corpora, among them Koniwa and SIWIS, credited here alongside the model. This site is a static build generated from our own knowledge base; it sets no cookies and builds no profile of you. The only measurement is an aggregate, cookieless page count that cannot follow you between sites. What is running, and on what legal basis, is set out in the privacy policy.