How we research and rank

Where the rating comes from

A product's star rating is the mean of the numeric ratings listed in the Rating sources box on its page, each normalised to a 5-point scale (a 9/10 becomes 4.5, an 85/100 becomes 4.25). Every source counts equally — we don't up-weight bigger publications or the reviewer we happen to agree with. That trade-off is real: on a two-source product one outlier moves the score a long way. The rule is at least visible and identical in every category.

Each publisher's score is put on one scale first. Publishers don't score alike: across the site, some labs that test everything side by side average well under 4 out of 5, while some magazines average above 4.5 for the products they choose to cover. A raw average would then say as much about who happened to review a product as about the product. So before averaging, each printed score is shifted by the gap between that publisher's average and the average of every printed score on the site. A publisher with only a few scores on file gets only part of that shift, because a few scores are not enough to say how strictly it scores. Under each affected score on a product page we print the number it counts as and the publisher's average, so you can see the adjustment. We tested this before turning it on: hiding one publisher's score and predicting it from the rest got about 15% closer with the adjustment than without it. The table of publisher averages is fixed between planned updates, so a product's score does not move because some other page changed. Derived scores (below) are not adjusted.

This is checkable. Add up the numbers each score counts as on any product page, divide by how many there are, and you should get the score we print. Some entries in that box read “Reviewed — no score”: those are outlets that reviewed the product and published no number, so they carry a link and a quote but are not part of the arithmetic. The line under the box tells you how many were. Two more cases are left out. First, when a publisher printed a score for the product, ratings we derived from other reviewers' prose (marked *) are not averaged in: a printed number beats our reading of someone's words. Second, when a product has at least two scores read from a publisher's structured data, a score from any other kind of source that sits more than 1.5 points from their median is treated as a likely misread. Both stay in the box, marked “Not in the average”, so you can see what was set aside. Rankings then follow the scores mechanically — no editor reorders a list. One rule weighs the depth of evidence: a product whose score rests on a single review sits one place lower than its score alone would put it, behind the next product that has two or more (as long as that product is rated 3.5 or higher), and its card says so. Apart from that one step, across all 442 guides there is no product ranked above one with a higher score. If you find a page where the math doesn't reconcile, that's a bug; tell us through the contact form and we'll recompute it.

Two kinds of number, and how to tell them apart

Plenty of good reviewers never print a score. They write a clear verdict in prose and leave it there. We use both kinds of review, and we label which is which.

Where a publisher states a number, we record it. That number is read off the review page itself — usually from the schema.org structured data publishers embed for search engines, sometimes from a printed score in the body copy. Across the 8,542 citations currently on the site, 3,135 (37%) are of this kind.

Where a publisher reviewed the product and printed nothing we could turn into a number, we say that too. 3,783 citations (44%) sit in that third category: a real review, linked and quoted, contributing no figure to any average. They appear in the sources box as “Reviewed — no score”. We would rather show you the review and admit there is no number than drop the outlet or invent one.

Where a publisher reviews a product without scoring it, we derive a rating from the review's content. A language model reads the full article or video transcript and returns a 1.0–5.0 rating plus a verbatim quote capturing the reviewer's verdict. That is 1,624 citations (19%) — a large share, and we'd rather say so than bury it. These are marked with an asterisk (*) everywhere they appear, and 67% of them display the quote they were drawn from, right next to the score. The quote isn't decoration: it is string-matched against the text we fetched from that page, so a quote that doesn't actually appear in the source is rejected before publication. Click through, read the reviewer's own words, and judge whether our number matches their conclusion.

A derived rating is an interpretation, not a measurement, so it only counts when there is nothing better: if any publisher printed a score for the product, the derived ratings stay in the source list as quoted context, marked “Not in the average”, and the score is the mean of the printed numbers alone. We measured why: on products that have both kinds, a derived rating misses the printed one by close to half a star on average, and lands on exactly 4.5 about twice as often. A check also runs the other way — a structured-data score that contradicts several quote-backed reads by more than 2 points is treated as a mis-scrape and dropped, because publishers' own markup is wrong more often than you'd hope.

How many sources sit behind a score

It varies by product, and the honest answer is that some scores rest on very little. Of the 2,256 products on the site right now:

• 1,728 (77%) have three or more cited reviews
• 398 (18%) have exactly two
• 130 (6%) rest on a single review

Two rules produce that shape. Our research stops looking once it has four usable citations for a product, which is why most land at three or four rather than a dozen. And a product now needs at least two citations to clear the publishing gate — the single-source products predate that gate and are still on the site pending re-research.

We don't hide the count. The source total appears in the byline of every product page, in the comparison table of every guide, and the full list of publications with their individual scores is on the product page. A 4.6 from one review and a 4.6 from six are different things, and you can see which one you're looking at before you spend money.

What this method can't see

Syndication. We count one score per publication — no publication is cited twice on the same product. (Two independent YouTube reviewers can both appear; they share a generic label but are different creators with different links.) What we can't catch is two publications reporting the same underlying test: a magazine re-running a sister title's lab results is two domains and one test. We count by domain, so some multi-source scores are thinner than the count suggests.

Provenance labels. We tag each citation with how the number was obtained. Spot-checking those tags against live pages, we've found cases where the tag is wrong — a score lifted from body copy recorded as if it came from structured data. The publication, the score, and the link are what we display and those check out; treat the internal provenance tag as the less reliable field.

Our own testing — there isn't any. No lab, no review units, no hands-on time with any product on this site. If you want teardowns and measurements, follow the source links; those reviewers did the work. What we add is the aggregation: finding the reviews, normalising the scores, and putting the disagreement in one place.

Whether three independent reviews is even the right bar is a fair question — a single rigorous instrumented test can be worth more than five hands-on impressions, and our method treats them identically. We're not claiming to have solved that. We're claiming to show you the inputs.

Where the pros and cons come from

For each product we read several professional review articles, browse the most upvoted Reddit threads in the relevant communities, and pull transcripts from the most-watched review videos on YouTube. The pros and cons you see are synthesised from that pool of source material — they reflect what reviewers and owners are actually saying.

How we choose the shortlist

You give us a starting product. We search the open web for established comparables in the same category, score the candidates by how often they appear in “best of” lists and head-to-head comparisons, and pick the top 4–5.

How we compare them

After every product has been researched independently, we run a final head-to-head pass that compares each one against all the others in the shortlist. That's where the “How it compares” section on each product page comes from. It's written by referencing the others by name, not in isolation.

Affiliate disclosure

Some links on this site are affiliate links. When you click through and buy something on Amazon, we may earn a small commission at no extra cost to you. This never changes how products are ranked or what we say about them — the ratings come from independent reviewers we don't pay or partner with.