How we source data
Verdicts are grounded in three layers of data:
- Specs. Structured attributes (price, size, performance numbers, release dates, etc.) pulled from manufacturer pages, official datasheets, and curated reference sources.
- Verified reviews. Aggregated public reviews from sources like Reddit, G2, Capterra, Trustpilot, and Product Hunt — normalized into ratings and recurring pros/cons.
- Live web context. Recent search-engine results and news snippets used to surface anything that's changed since the page was last refreshed.
The role of the AI model vs. our human editor
We use Claude (Anthropic's large language model) to read the structured data and reviews above and produce a first draft of the verdict, the “Choose X if” cards, and the supporting summary. The model only sees data we've already collected — it does not browse the open web at write time.
A human editor then reviews drafts that fall below an internal quality threshold or that cover sensitive categories (health, finance, legal, safety). Editors check for accuracy, remove unsupported claims, and adjust tone. Pages that have been human-reviewed are flagged in our internal system.
How feedback influences future verdicts
The 👍 / 👎 widget under each verdict is the fastest way to tell us when something is off. We use those votes — and the optional written reasons — in three ways:
- Pages with persistent negative feedback get prioritized for human review.
- Common categories of complaint (“misses recent release,” “ignores battery life,” etc.) feed back into the prompt and data pipeline so the next regeneration covers them.
- We sample reader feedback when we evaluate new model versions — a verdict that wins on human reads but loses on reader feedback gets rolled back.
Disagree with a specific verdict?
Use the “Was this verdict helpful?” buttons directly under the verdict on any comparison page. Tap “No,” and a short text box will appear where you can tell us what we got wrong — recent release, missing context, factual error, biased framing, anything. Reasons go straight into our editorial queue.
Prefer email? Reach us at our contact page.