Credible picks

How it works

How we score

How an answer is made, step by step, and how well it works. The rules and numbers on this page come straight from the engine's code and its latest test run, so they always match what it really does.

Where the threads come from

Every answer comes from Reddit threads we saved ahead of time, from communities where people talk about these things in depth:

We read the threads through third-party services that collect public Reddit posts. People can delete or change what they wrote, so we keep checking:

How products and opinions are found

An AI reads each saved thread ahead of time, up to 150 of its most upvoted comments. For every product someone talks about, it writes down:

Then code, not the AI, checks that the quote really is in that comment, word for word. A quote that isn't is thrown away. That way the AI can never put words in anyone's mouth.

People write the same product in different ways ("CeraVe SA" and "CeraVe Renewing SA Cleanser"), so names are matched up and counted as one product.

Who counts as a credible voice

For each comment we look at who wrote it and how other people reacted to it, and count good signs and red flags.

Good signs

  • Been around: their account is at least 2 years old and they take part a lot (at least 1,000 karma, the points Reddit gives for upvotes).
  • Well liked: people usually upvote what they write, at least 5 karma for each comment or post.
  • Expert flair: the label next to their name in that community says they do it for a living, like dermatologist or chef. Anyone can type a flair, so it's a claim, not proof.
  • Well upvoted: the comment scored higher than 75% of the other comments in its thread.
  • Backed up: other people replied to say they agree.
  • Knows the topic: uses at least 3 expert words for it, like “gyuto” for knives or “niacinamide” for skincare.

Red flags

  • Brand-new account: less than 30 days old when they wrote the comment.
  • Few upvotes for how much they write: under 1 karma for each comment or post.
  • Sounds like an advert: “buy now”, “link in bio”, “use my code”.
  • Works for a brand: they say they work for, own or represent one.
  • Downvoted: more people voted the comment down than up.
  • Copies someone else: 20 words or more in a row are the same as another writer's comment in the thread, as when two accounts post the same recommendation. Copies count once: only the first one posted counts at all.

Then each comment gets a level:

Skincare formulas change, so a skincare comment older than 3 years is medium at most. And someone who recommends a product in one comment and warns against it in another, without saying what changed, counts as low.

How much a voice counts: high 1, medium 0.5, low 0.25. So a high voice counts 4 times as much as a low one.

How strong the evidence is

For each product a writer mentions, we ask: how well do they know it?

Each sign of honesty adds 10%:

How a product becomes a pick

Every mention earns points: the voice's number times the evidence's number. A recommendation adds the points, a warning takes them away, and a mention with no opinion counts nothing. For example:

A product's score is all its points added up. A mention is credible when it recommends or warns, comes from a high or medium voice, and the writer has used the product (long-term use or short-term use).

A product becomes a pick only with:

When a product has fewer than 3 credible recommendations, one whose comment has at least 10 upvotes, and that scored higher than 90% of the other comments in a thread of at least 20 comments, counts as two toward this rule, never more, and only as many as the product needs to meet it. Such a product only takes a place the picks would otherwise leave empty: it never takes the place of one that meets this rule on its own. The extra count never stands in for the second thread, upvotes never make a comment credible, and a warning always counts once. The pick's reason says which voice counted twice and for how many upvotes; every other number counts people.

When you ask for more than the product, such as a skin type, a concern or a must-have, at least one of those credible recommendations must also talk about it: in every answer for haircare and skincare, and for anything else when we read more threads than usual. An ingredient or a size you name isn't counted here: whether a product holds an ingredient is checked against its maker's list instead.

The 3 products with the highest scores are shown. If fewer have enough evidence, we show fewer and say so: we'd rather say "not enough evidence yet" than guess.

When credible voices clearly prefer one kind of product (at least 2 times the support of the next kind, say Japanese chef knives over German ones), products of that kind get a bonus of 1.

Each pick shows 2 to 3 quotes from the people who back it, each with a link to its comment and checked word for word again on Reddit, as above.

When we've looked a product up, its card links to its own page, its maker's or a big UK shop's, with the day we checked, such as “See it at John Lewis”, and may show a small image from its maker's page. We never link to a search: a product we haven't looked up yet has no link.

Products with 2 credible recommendations from at least 2 different threads aren't picks, but up to 3 of them are shown apart, under “Also worth a look”, marked as less supported than the picks, each with its own checked quotes. A product that meets the bar but ranks below the top 3 is listed there first, marked “Meets the bar for a pick”, within those 3 places.

A product can also be shown when one comment about it stands out in its thread: a credible recommendation whose comment has at least 25 upvotes, and at least 75% as many upvotes as its thread has comments, in a thread of at least 20 comments. It must name at most 5 products, and talk about what you asked for when you ask for more than the product. Its product can then take a place left empty under “Also worth a look”, even when that comment is its only credible recommendation, and with one more credible recommendation, even from the same thread, a place the picks would otherwise leave empty. It never takes the place of a product that meets the rules above. Upvotes never make a comment credible, the product needs more credible recommendations than credible warnings, and its reason says which comment stands out and by how much: “its comment has 74 upvotes in a 69-comment thread”.

Some picks have a one-line summary of what their credible voices say, after “In short:”. It is in our words, never a quote: an AI writes it ahead of time from the comments, and code checks it before it's shown. Every comment it draws on must be one of that product's credible voices in this answer, checked word for word like a quote; every number in it must be written in those comments; and it must mention a downside when credible voices warned against the product. A summary that fails a check isn't shown.

A pick can carry a label, such as “Most praised for sensitive skin”, when at least 2 of its credible voices write about that (or wrote after long-term use), and at least 2 more than for any other pick shown, or when its price, checked in your region in the last 30 days, is clearly the lowest of the picks, and it says only what those voices say, so many picks have none.

When credible voices clearly name one model of a brand's line most (at least 2 of them, and at least 2 times as many as for any other model), the pick is titled by that model, and its card says, of its credible voices, how many are about the model and how many about the line in general: the ranking is still the line's. For fast-moving tech, when the model its owners reviewed has been replaced, the card shows the current model, from its maker's page or a big UK shop's, with the day we checked.

Warnings count too. A product with at least 2 credible warnings, and more warnings than recommendations, goes on the "Skip these" list, unless it is a brand or line one of the picks belongs to ("Lodge pans" next to the pick "Lodge"): then its warnings are shown under that pick instead, so an answer never tells you to skip one of its picks. A pick that at least 2 credible voices warn against is marked "Mixed opinions".

What gets left out

Smaller clashes keep the product in, with a note under it, such as "Note: its texture is rich, which can feel heavy on oily or acne-prone skin."

A product sold in the UK only on a prescription says so wherever we show it: "Prescription-only in the UK: ask a doctor or pharmacist." It is never offered to a beginner, or when you ask for something gentle or for sensitive skin.

When we couldn't check a fact you asked for, the product can still be shown, with a note, but after the products whose facts confirm it: "Note: we couldn't confirm it's PFAS-free." A lip balm whose own name says it has SPF counts as having it. A whole brand can't be checked like this, so a brand on its own is left out when your request has a must-have like these.

Under each answer we say how many products, and how many brand or line names, were left out, and why, with a few of their names. We count names, not mentions or threads.

How well it works

We test the engine against work we checked by hand, and against targets we set before building it.

Last measured on 16 Oct 2026.

The targets we set

Our test questions

Of our 10 test questions (the ones real testers will judge in the blind test), 10 got a full answer with 3 picks and 10 got at least one pick. Those answers showed 253 quotes, and 0 quotes failed the word-for-word check.

Other checks

Your privacy

We collect as little as we can, and never your name.

What we don't collect. No names or accounts, no cookies, no adverts and no tracking across websites. We never write down your IP address ourselves, and we don't save your searches, except the request you send with feedback, and the short note about a product we haven't read about yet.

To stop anyone flooding the site, our server counts each address's requests for a short while, in memory only: it never writes them down.