How a claim gets onto this site

A claim cannot exist in our database without the verbatim span that establishes it. That is not an editorial policy that could be relaxed under deadline; it is a not-null column, and a claim arriving without a quote fails to insert.

What someone observably said and what we think it means are stored as two different things. The quote is the durable layer and is never overwritten. When we improve how we read a source, the new reading is written as new claims against the same original quotes, so the record of what was actually said survives every change we make to our interpretation of it.

3418 claims · 4893 quoted spans · 151 sources · 78 named people

One. Acquire. We fetch the published captions or transcript of a long-form source and segment it. We never publish a transcript, and we never reproduce more of one than the quotes a claim needs.

Two. Read. A language model reads one segment at a time against a versioned instruction set and proposes claims. This is the only place a model touches the product, and its output is treated as derived data rather than as truth. Every claim stores which model produced it and which version of the instructions it read, so a bad batch is identifiable after the fact rather than merged invisibly into the rest.

Three. Check. The proposed quote is checked back against the transcript. A quote that is not verbatim is not a quote, and the claim it was supporting does not load.

Four. Resolve. Product names, brands, categories and reasons are matched against a controlled vocabulary. This step is deliberately unable to guess. Fuzzy matching is allowed to decide whether we have seen a phrase before; it is never allowed to decide that two products are the same product, because two different brands' skillets read as nearly identical strings and one brand inheriting another's claims is the worst thing this database could do quietly. A name we cannot resolve is queued with a count of how often it recurs, and the claim keeps the words that were actually said so it can be resolved later rather than being lost now.

Five. Publish, or refuse to. Pages are generated from the graph. A subject whose evidence does not clear the threshold does not get a thinner page; it gets no page and no URL.

One sentence usually carries several claims — a recommendation, a performance observation, a reason, sometimes how long the speaker has owned the thing — and they are stored separately even though they share one quote. Collapsing them would destroy exactly the detail that makes the graph answerable.

Four things are kept apart that a simpler system would fold into one number, and each separation exists because folding it lost data:

Stance is which way the claim points. Strength is how emphatically it was said. Confidence is how sure we are it was said at all. Disclosure is what commercial interest the speaker had. A paid reviewer is not a less confident claim; a mild recommendation is not a negative one. Once any two of those share a column, one of them dies silently.

Reasons are stored as references to a controlled vocabulary rather than as prose, which is what makes “how many people cite heat control” a countable question instead of a text search.

4256 cited reasons · 1013 negative or mixed claims (29.6% of all claims) · 650 claims carrying a declared interest

The publication gate is the structural defence against thin content, and it is a rule rather than an intention: a page is generated only once its subject holds enough independent evidence. There is nothing to deindex later, because the URL was never minted.

516 pages published · 240 of them offered to search engines · 733 candidate pages held back

Some published pages are deliberately kept out of search anyway. A page about a source exists so a claim can cite it and so a reader can see everything drawn from one video in one place — it should not compete with that video for its own name.

A chef is authoritative about cookware and is not authoritative about headphones. We hold that as a per-person, per-domain value rather than as a general reputation, and it is derived from the breadth of sources a person's expertise shows up across, never from how many claims they make. Ranking by volume ranks whoever talks most, which is a measure of prolificacy wearing the costume of expertise.

Subscriber and follower counts are recorded as reach and are never a ranking input on their own. Nor is disclosure: a paid reviewer can be genuinely expert, and the conflict travels with the claim where a reader can see it, rather than quietly demoting the person.

It is also thin, and worth saying so plainly: 61 expertise scores exist, covering 15 of the 93 people on this site across 34 domains. Everyone else carries no expertise score at all — not a low one, and not a guessed one. Where you see a person on this site with no stated expertise, that is what it means.

We do not compute an overall score. Every attempt to rank products on one axis smuggles in a decision about what matters that belongs to the reader.

We do not suppress negative claims. Objections get the same extraction rigour, the same visual weight and the same page position as endorsements, including on products we would earn a commission on. 29.6% of the claims here are negative or mixed.

We do not publish transcripts. Short quotes and deep links only. The source keeps its own traffic and its own words.

We do not write product prose by hand. If a sentence about a product cannot be derived from the graph, it does not go on the site. This page and the about page are the two exceptions, and neither makes a claim about a product.

One voice holds 46.4% of every claim that names a product

Of the 1772 claims here that name a specific product, 822 come from Prudent Reviews — 46.4%. Across the whole database, including the claims that name only a brand or a category, the same reviewer holds 1116 of 3418, or 32.7%. We lead with the first number because it is the one that describes what you are looking at when you are standing on a product page, and because it is the worse of the two.

It comes from 35 of our 151 sources. Every count on this page is over claims we still assert — not superseded by a later reading, not rejected in verification. Today that is all of them, and stating the rule means the denominator cannot move quietly later.

It is worse on individual pages than it is overall. 7 of our 20 published product pages are more than half Prudent Reviews. The most concentrated is All-Clad stainless pan, where 80 of 102 claims — 78.4% — are that one reviewer's. A page like that is one publication's opinion presented in the apparatus of a consensus, and reading it as a consensus would be a mistake.

It is not a secret relationship: Prudent Reviews is a named source like every other, the claims are individually attributed, and where that reviewer discloses an affiliate interest the claims say so. It is a sampling problem. The fix is more sources rather than fewer claims — deleting evidence to improve a ratio would be the dishonest version of this — and until the ratio moves, the sentence stays on this page.

Exactly one product has more than one chef behind it

The idea this site is built on is that you should be able to ask what people who actually know a subject converge on. We hold 13 people identified as chefs. Exactly one product — Lodge cast iron skillet — has more than one of them saying anything about it at all, and that one has 3. Every other product a chef has mentioned — 32 of them — has exactly one.

That has a sharper consequence than it first sounds: there is no comparison in this corpus with chefs on both sides of it. So “chefs converge on this pan” and “chefs are split between these two” are both sentences our own data cannot support, and you will not find either anywhere on this site.

What the corpus does support is narrower and still useful: what one named person said, what they said it for, whether anyone contradicts them, and what a group of reviewers running comparable tests found. The depth behind any single cohort is thin, and that is a question of how many sources we have read, which is fixable and is being fixed.

Every source is on one platform

All 151 sources are YouTube. That means the corpus inherits whatever YouTube rewards — production values, personality, an audience that returns — and it under-represents people whose expertise never reaches a camera. Podcasts, forums and written long-form are in the design and not yet in the data.

Every claim on this site shows the person, the source, the timestamp and the verbatim quote. Open the source and check the second. If a quote is wrong, a person is mislabelled, or a claim is attributed to someone who did not make it, tell us — that is the failure mode that matters most here, and it is the one we most want reported. The same graph is available as an API, so the counts on these pages can be reproduced rather than taken on trust.

Disclosures we currently record: earns a commission on a sale (455), was sent the product (138), works for the brand or a related business (43), was paid for the coverage (14). The disclosure page explains what each means and what we do with it.