How recommendation scoring works.

A tool asks a handful of questions, weights the answers, and compares them against curated result profiles. Every shortlist explains why it won, where it compromises, and which alternative would take the lead if priorities shifted.

Profiles carry pros, cons, best-for, and evidence signals (setup, support, integrations). The engine is deterministic — no hidden ranking boost for higher commission. Important facts carry a per-row verification date, are checked against official sources, tracked over time, and surfaced on results and comparisons. Commercial relationships do not influence rankings.

1. Hard filters first. Budget caps, platform support, model capacity, open-weight requirements, and capture preferences make a product ineligible before scoring. Results say what was excluded and why — results disclose which requirement families excluded products; capacity-fit exclusions are described in the model-capacity guidance.
2. Weighted scoring second. Surviving products are scored on normalized answer dimensions with deterministic tie-breaking (listing order, then slug). The same answers always produce the same shortlist.
3. Price as a bounded post-check. For pricing-pilot categories, an estimated cost can only reorder close calls — it never overrides capability fit, and over-budget products are removed outright when in-budget alternatives exist.

Every model comparison needs verified pricing on both sides, a meaningful capability or price difference, mainstream demand for at least one lab, and no missing-data disqualifiers. Pairs below the threshold remain reachable and linked, but are marked noindex and excluded from the sitemap. The same policy is visible on each page as a data-confidence receipt: benchmark sources, per-model citations, and the date each catalog record was last checked.