Methodology

1. A product is four different things

The catalog keeps these apart because mixing them is how comparison sites give wrong answers:

• Brand → family → model: “Pixel 9 Pro” is a model of a brand.
• Variant: a hardware or regional version of a model — different chipset, memory, radios or box contents. Variants carry their own values; nothing leaks from one to another.
• Offer: a retailer’s price for one exact variant. (Not built yet.)

A model is never merged with another because the names look alike. Shared identifiers (a source id, a GTIN) can link records; identical names alone cannot. If two claims collide, the record waits for a human.

2. Every value has evidence

Each fact stores where it was read (a specific statement in a specific record), when it was first retrieved, when a real re-fetch last confirmed it, which licence applies, and how it was parsed. Numbered references on every page lead to these records.

“Last checked” is honest. It changes only when the source was actually fetched again and returned the same content. Drawing the page does not refresh it. A fact older than 90 days is labelled stale.

3. Five states, not two

• Known. A value with evidence and a real check time. Only known values are shown as facts and used in comparisons.
• Unknown. Nobody we trust has told us. Shown as “Unknown” — never as zero, “no”, or a blank that looks like a value.
• Not applicable. The field does not exist for this kind of product (for example a headphone jack on a keypad phone with no audio out).
• Conflicting. Independent sources give different values. The disagreement is disclosed; no value is shown and nothing is averaged.
• Unverified. A claim that has not passed validation. It is kept for review but never published.

The database itself refuses a “known” value without evidence, and refuses any value on a field that is unknown.

4. Comparison rules

• Models are compared only within one category. There is no universal technical score.
• A difference is reported only where at least two models have a known value. A value known for one model and unknown for another is labelled “known for one” — not a win, not a loss.
• There is no overall winner or score. Nothing here ranks products by combining numbers.
• If a model has several variants you must choose one; the page never silently picks a default chipset for you.
• Frame rates and similar ratios keep their exact form: 59.94 fps is not shown as 60 fps.

5. Rights come first, per action

For each source we record fourteen separate permissions — from downloading to showing facts, hosting images, AI processing and redistribution — each allow, deny or unknown. Unknown means off. You can read them on the source ledger.

No AI touches source data in this release. Values are read by a deterministic parser and checked against the category dictionary. AI helped to write the software; that is a different thing from sending a source’s data to a model, which would need its own explicit permission.

6. Checks that run on every import

• Types and units are checked against the dictionary; units are converted with a tested table and the original is kept in the evidence.
• Physically implausible values (a 500 kg phone) are quarantined and never overwrite a verified value.
• Contract manufacturers are not treated as brands: a record whose only “maker” is one waits for review instead of appearing under the wrong brand.
• Regional differences that a source lists at model level (for example two chipsets) are not flattened into one value.
• Re-importing the same record changes nothing; a changed value supersedes the old one and the history is kept.

7. Corrections

Spotted a mistake or know a missing value? Send a correction with evidence. Corrections go to a review queue; they never edit a page directly.

8. Not built yet

• Prices and retailer offers, with country, condition and staleness rules.
• Compatibility checks (camera–lens, laptop–dock–monitor, phone–carrier bands).
• Coverage against full target lists, and sitemap/indexing once every page passes the publication gate.
• More categories and sources, each added only after its rights are reviewed.

How a new category is added

A category arrives with its field dictionary, identity rules, at least one permitted source, and tests with hard real examples. No separate site is created: the same identity, evidence and rights model serves every category. Twelve are defined today; one has data.

See what is covered