How we evaluate every product

How a Verdict Is Decided

The answers are read against a rule tree, not weighted into a score. Every verdict on the site names the rule that produced it. A verdict is our assessment of how well a product’s ingredients fit a concern — never a claim that the product treats anything.

The scale

Every verdict word sits on one of six steps, and each step has one colour everywhere on the site. The colour is the step; the word is the reason. Learn the six once and every card reads the same way.

1. RecommendedStrong choice · Full protection · Excellent at the jobA first-line or approved active in a form that works, with no trigger for the concern.
2. GoodGood choice · Good choice (sensitivity) · Does the jobA supported active at a working level, or the right job done without a trigger.
3. SupportingSupporting role · Basic protectionHelps the routine; does not treat the concern.
4. Early evidencePromising · UnprovenA bet, not a recommendation; the dots say how much has been tested.
5. Not for thisNot for this · Falls shortResearched and found wanting, or missing what the concern needs.
6. AvoidAvoidThe formula carries a known trigger for this concern.

The words

Under each verdict sits a receipt: the document the standing rests on, the dose against the studied range, the date the row was checked, and the rule, with how many products share the reading. It opens into a full receipt page for the product. Beside it, four evidence dots say how much trial evidence stands behind the lead active for that concern — four need a guideline endorsement or a high-certainty replicated meta-analysis, three need at least one trial or a systematic review. The word answers “is it recognised for this?”; the dots answer “how much has it been tested?” They are two different questions, explained under Evidence.

Every Rule, With Its Reasoning rules 0.5.1 · evidence table ev-2026.10-r231

Rules are evaluated in order; the first that matches decides. Each is written out in words, with the condition the engine actually evaluates beneath it and the number of verdicts it decided on the current build. The register and its anchor set are editorial decisions, each with a written rationale; no advisory panel has reviewed them yet, and this page will say so until one has. Our Chief Medical Reviewer reads published pages afterwards as style and quality control — not as a gate — and nobody is named on a rule or a verdict.

Triggers that make a formula an Avoid for a concern are listed per concern in the same register; cautions — a photosensitising active, a strong acid, a single declarable allergen — add a caution line to a verdict and never decide it. A Tier 1 or 2 seal from an expert body (the National Eczema Association, the Canadian Dermatology Association and their peers) outweighs a trigger for the concern the seal covers.

Where the Evidence Comes From

Underneath every verdict sits a versioned table of ingredient–claim pairs — currently 231 reviewed rows. A row is never “niacinamide is good”; it is “niacinamide for post-inflammatory hyperpigmentation”. Each records what the primary literature measured — design, population, dose, duration, endpoint — with its sources and review date. Its tier is computed from those findings, never typed in by judgement.

Three rules keep the table honest

Two questions about every row

Standing asks whether the ingredient is recognised for the concern — and must rest on a document we opened and dated: a regulation, a guideline, a systematic review, never a person. Certainty asks how much dedicated evidence is behind it, derived by a fixed rule from trial count, replication, consistency and source quality. They are kept apart everywhere, because collapsing them would either overstate thin trials or hide real standing.

On every product page, the Key Ingredients table shows each hero’s standing and certainty for the concern it is read against, from the same table the verdicts use; the studies are one tap away. Every citation on the site is checked by machine against the source database before it publishes — the title, journal and year at the link must match the reference — because a plausible link to the wrong study is worse than no link.

Four Things You Can Check Without Asking

A method is only as good as what a sceptic can open. These four pages are generated from the same files the engine runs on, every build.

And for any product, its receipt page — linked under every verdict — holds the rule, the row, the sources and the dates behind each verdict on that product.

None of it is frozen. As formulas change and research is published, verdicts are re-derived: every new row in the evidence table updates every verdict that depends on it, across the catalogue at once.

Reading Reviews Instead of Counting Stars

A 4.7-star average tells you a product pleased people; it can’t tell you it failed everyone with rosacea. So we read review text, collected from retailer and platform APIs and passed through a fixed extraction model — measured against human coding at κ = 0.851, a level of agreement generally considered strong. It reads what people with a specific concern reported, and flags undisclosed promotional reviews — in our pilot, on some products a third to a half of reviews turned out to be seeded placements, invisible in the star average everyone shops by. Where reviews haven’t been collected for a product, its verdict rows say so in as many words, and the verdict stands on formulation and evidence rather than on a guess.

Readers can also answer for themselves. Every product page carries a short poll: have you used it, and how did it go for your concern. Answers are shown to everyone only once 25 people have given one, and they sit beside the verdict, never inside it. The live polls in the Town Square, on a dozen well-known products, show their tally straight away. A poll records what people report; it is a useful signal, not clinical evidence.

What Nothing Can Outweigh

Some findings disqualify a product however well its ingredients read. A gate can only be entered through its published criteria — never by editorial judgement — and every gate has exit criteria, a review schedule, a disputed state for findings under formal challenge, and a written appeal path. Two gates read from the ingredient list and fire today (category integrity, self-neutralising formula). The other four need a public record — a recall notice, a regulator’s finding, documented review manipulation, a settled expert consensus — and their register is public and currently empty: no product sits behind any of them. The page says so rather than letting an empty list read as a clean bill.

Hype-Aware, Never Cynical

Not every product earns the attention it gets. We apply friendly skepticism where it’s warranted — always assessing the formula on its merits, never punching down, and never mocking the people who love a product.

The Numbers You Don’t See

Beneath the verdicts, a deterministic engine still scores each formula across six factors — formula quality, safety and tolerability, real-world results, usability, value, and brand trust. No number is shown on a product page or a card: a verdict says what a number cannot, and a number invites a ranking the evidence does not support. The engine’s job is quieter — it supplies the tolerability answer inside every verdict, the category “job” reading, and the safety gates above, and it unfolds into a citable audit trail. Every fact has exactly one owner (fragrance in safety, texture in usability, price in value), and value is never derived from popularity.

What’s in Every Review

Beside the verdict, every product gets an 800–1,500 word editorial review — written by someone who understands formulation science, opening with an angle of its own rather than a template.