Our Methodology
Every claim on this site carries a number out of 100. This page is the whole rubric behind that number — where it comes from, what can move it, and what it can never be. If a score looks wrong to you, everything you need to audit it is on this page.
Last reviewed: 2026-08-04.
The 30-second version
- 1. We score the source, not the sentence. A published study starts at 95. A seller’s own product page starts at 40.
- 2. A claim carries its source’s score, and that score can never rise above the ceiling for the tier it came from.
- 3. A claim can never score above its source’s ceiling. No amount of confident wording gets a marketing page to 95.
- 4. Under 40, it never publishes at all — it goes to a review queue.
1. Every claim has a source
1. Every claim has a sourceNothing appears on this site unless it came from a page we actually read. Each benefit, routine and side effect keeps a link back to the document it came from — the study, the regulator’s page, the reference article. Click any ↗ source link and check us.
We do not write claims from a language model’s memory. A model reads the source page and pulls structured claims out of it; anything it cannot point at in that page is dropped. That is the difference between a summary and an invention, and it is the whole reason this site exists.
2. The eight source tiers
Every source we crawl is classified into one of eight tiers. The tier sets the claim’s starting score — and its maximum. These are the exact values our scorer uses, not a simplified version of them.
| Score | Tier | What it means |
|---|---|---|
| 95 | Peer-reviewed study | Published in a peer-reviewed journal. The gold standard — other experts checked the methods before it ran.e.g. PubMed, PMC, DOI-linked journals |
| 90 | Regulator or clinical registry | Issued by a regulator or a clinical trial registry. Carries the weight of an official health authority.e.g. FDA, EMA, NHS, ClinicalTrials.gov |
| 72 | Established reference | Reputable reference outlets and newsrooms with editorial oversight and fact-checking.e.g. Mayo Clinic, examine.com, Healthline, Reuters Health |
| 60 | Specialist community | Expert forums, practitioner communities, and industry trade press. Useful primary signal, short of peer review.e.g. ExcelMale, FiercePharma, press releases |
| 50 | Unclassified sourcenot published here | A page our classifier could not place in a stronger tier. Kept on the ledger at a neutral score rather than guessed at.e.g. miscellaneous sites still under review |
| 40 | Vendor / seller pagenot published here | A product page whose job is to sell. May describe real products, but its claims exist to convert buyers.e.g. peptide shops, 'add to cart' pages |
| 30 | Forum or blognot published here | Anecdote and opinion from forums and blogs. Genuine experience, but unverified and often second-hand.e.g. Reddit threads, wellness blogs |
| 12 | Marketing copynot published here | Pure promotion — written to sell, not to inform.e.g. landing pages, sponsored copy |
4 of these 8 tiers can appear as a claim on this site. Claims whose strongest source sits in the bottom 4 — unclassified source, vendor / seller page, forum or blog, marketing copy — are held back from this consumer site, though our research knowledge base keeps them.
3. How a score is calculated
A claim starts at the score the model extracts from its source page. Two things then happen to that number, in code, before it is ever stored or shown:
- The extracted score
- The model assigns a score as it reads the page. That is the claim’s starting number — not a number we invent, and not a number the page can dictate on its own.
- Clamped to 0–100
- A reading can never go below 0 or above 100, whatever the page implies.
- Capped at the tier ceiling
- The single rule that does the real work: the score is then hard-capped at the ceiling for the source’s tier. This is the step that stops a confident page from buying a score it did not earn — covered in full just below.
You can check this yourself on any claim. Hover a score badge anywhere on the site — or tab to it with a keyboard — and it opens the breakdown for that one claim: the source tier, that tier’s ceiling, how many independent sources back it, and the published number. On a phone, tapping a badge brings you to this page instead.
4. Tier capping: a claim can never score above its source
This is the part we would most like you to test.
Each tier’s score is also a hard ceiling. A claim taken from a vendor page cannot score above 40. A claim from a marketing page cannot score above 12. The cap uses the source page’s actual domain tier, not the tier it claims for itself — so a page can’t lift its own score by declaring a higher tier, however confidently it describes its evidence.
The cap is applied in code, after extraction and before anything is stored. The claim itself proposes a score; the cap overrules it. So the ceiling holds even when a source page is actively trying to game us — and pages selling things do try. It is the reason a 95 on this site means something narrower and more useful than “we felt confident”: it means a peer-reviewed source said this.
The reverse is also true, and less flattering to us: a well-established fact that we have only found on a low-tier page will show a low score here. The score rates our evidence for the claim, not the claim’s truth in the world.
5. What goes to review
A claim is held back for human review, rather than published, when any of these are true:
- It scores below 40. That is our publish threshold. A weakly-sourced claim waits in a queue instead of appearing on a page.
- The extraction was not confident. Every page we read comes back with a confidence figure between 0 and 1. For research and reference sources it has to reach 0.85 — otherwise the entire page is queued, not just the shakiest claim on it.
- The output failed validation. If the structured result from a page is malformed, or a field looks manipulated, we keep the raw output for inspection and publish none of it.
- The page tried to manipulate us. Crawled pages are treated as untrusted input and screened for instructions aimed at our extraction step. A page that looks like it is trying goes to review, or is dropped.
What we refuse to put in front of you
- Anything with no traceable source. No source URL, no publication. A claim we cannot point you at is indistinguishable from one we made up, so we treat it as one.
- Citations we cannot resolve. A reference that is just “Ma et al.” with nothing to look up gets dropped rather than displayed as evidence.
- Promotional copy dressed as description. Vendor sales language is rejected wholesale from summaries rather than lightly edited.
- Claims whose strongest source is one of the bottom 4 tiers. If the best source we have for a claim is an unclassified, vendor, forum or marketing page, this site does not carry the claim. Our research knowledge base keeps it; this consumer site does not.
6. If a score looks wrong
Then one of a few things happened, and this page is enough to tell them apart: the source is tiered wrongly, the claim is attached to the wrong source, or we are missing better evidence that exists. Open the badge, compare it against the rubric above, and follow the ↗ source link. If it still does not add up, tell us which claim and which step — that is a fixable bug at a specific line, not a matter of opinion.
7. Affiliate transparency
We may earn a commission when you buy through our “Where to Buy” links. This does not change the price you pay, and it does not touch any score on this site. Scoring runs on source tiers in an automated pipeline that has no idea which vendors we have a relationship with — and as the tier table shows, a vendor page is one of the lowest-scoring things we can read.
Policies & trust
The commitments behind every page on this site.