Methodology
How we match, label and rank
Published so you can judge our results — and catch us if we drift.
Evidence labels
These labels are not interchangeable:
- Vendor-described. Captured from the vendor's own materials. Not independently verified.
- Source-verified. Exact facts confirmed in official documentation or pricing pages.
- Tested. A dated test case with inputs, observed outputs and limitations exists.
- Maintained playbook. Tested and periodically rechecked. A second AI model reviewing prose does not count as product testing.
Matching: match-v1 (2026-10-15)
- Hard filters first. Budget, industry, API and integration requirements remove tools that don't meet them. Unknown never passes: an unknown price fails a budget cap, an unknown API fails “needs an API”, an undocumented integration fails a required integration.
- Keyword and tag relevance. Your words are normalized (lowercased, common words removed, light stemming) and expanded with a small published synonym list. Each matched term scores by the strongest field it appears in:
| Field | Weight |
|---|---|
| Product name | 5 |
| Tags | 3 |
| Capabilities | 3 |
| Categories | 2 |
| Mapped use-case titles | 2 |
| One-line description | 1.5 |
| Documented integrations | 1 |
| Overview text | 1 |
Synonym expansions count at 0.5× weight. Results scoring below 2 are dropped rather than padded in.
- Ties break on evidence level (tested → source-verified → vendor-described), then most recent verification, then name.
- Explanations are generated only from the matched catalog fields, and every result is checked against the catalog before it is shown. If nothing matches, we say so.
- Meaning match (when switched on): each published listing's own description is turned into a vector, and so is your search. Tools whose description is close in meaning can join the results even without a shared keyword, labeled “Related by meaning”. They must pass every filter, need a similarity of at least 0.35, are capped at 8 per search, and are merged with keyword results by rank fusion that gives keyword matches the edge (1 vs 0.8). The results page says whether a search used keywords only or keywords plus meaning.
We do not publish numeric quality scores for tools we haven't tested. Match relevance (how well a tool fits your words) is kept separate from editorial quality.
What money can and can't do
- Sponsorship and affiliate status are not inputs to the ranking code. This is covered by automated tests.
- Sponsored placements are labeled “Sponsored”, shown separately from organic results, must fit the category, and expire on schedule.
- Nobody can buy a “tested” label or an undisclosed review.
Freshness
- Pricing is marked stale after 14 days; other facts after 30 days.
- A failed check never erases the last known value — it marks it uncertain.
- “New” means added to this catalog in the last 30 days. A launch date appears only with release evidence.
Third-party ratings and user feedback
- We show scores from review sites like Capterra, G2 and Trustpilot exactly as each site published them, with a link and the date we checked. We never blend them into our own score.
- Scores based on fewer than 10 reviews are left out, because they're too small to mean much.
- “What people like” and “Common complaints” are our own short summaries of themes that come up repeatedly in public reviews and discussions. Each one links to where it came from. We don't copy review text or show reviewers' names.
- Ratings and feedback never change the order of search results.
Collections
Weekly picks state a reason for every pick. Established collections are for one task, based on documented evaluation, and dated — not “best forever”.
Professional use
For lawyers and other regulated professions we never infer that a tool is suitable for confidential client data from marketing language. Review each vendor's terms and your professional obligations.