Methodology

How a price earns its way onto this site.

We don't publish a price because a model said so. We publish it because it survived a chain of checks that an automated system can't talk its way past — deterministic code, a live re-check of the real source, an independent second source, a cross-model judge, and a human's approval. Here's the whole chain.

Ingest

On a schedule, we pull current prices and discounts from public store data and price-tracking sources, normalize them into one shared format, and detect what changed since last time.

Extract with citations

Each candidate price is structured into a record where every field carries the exact source quote and URL it came from. A field with no citation is rejected automatically.

Validate (code, not opinion)

Deterministic checks confirm the data is well-formed and plausible: price is positive, discount is 0–100%, no impossible jumps. Code can't be talked out of a rule.

Re-check the source

We re-fetch the original listing and confirm the quoted figure actually appears there. This kills fabricated numbers on the spot.

Cross-confirm

A second independent source must agree on the price within tolerance. One source is a claim; two agreeing sources is a fact.

Independent judgment

A separate reviewing model — deliberately from a different model family than the one that read the page — rules each field supported or unsupported, with a confidence score.

Human approval

A named editor reviews the day's queue, sees the source quote, the agreeing sources, and the confidence, and approves or rejects each deal. Only then does it go live.

What happens when a check fails

Nothing fails silently. A malformed record is rejected. A price whose quote can't be found at the source, or whose second source disagrees, or that the judge marks unsupported, is routed to human review with the exact reason attached — and the previously-published price stays live, untouched, until a human resolves it.

The safety net

We run known-answer "canary" deals through the live pipeline on every cycle. If the pipeline ever starts rubber-stamping — a source changes shape, a model regresses — the canaries come out wrong and the whole pipeline halts and alerts us before any bad data can spread. Every published figure also writes an immutable audit record, so when something is wrong we can trace exactly where it came from and fix it.

The system works because it isn't trusted. Each step hands its output to something that can prove or disprove it — code that can't be bluffed, a rival model that doesn't share its blind spots, a live re-check of the real source, and a human who signs off.

Read the full standards in our editorial policy.