Scope and selection
AI Ranking summarizes public-data snapshots, not controlled benchmarks. Measurement windows are unknown. Listings may be stale or incomplete; availability and latency are not service-level guarantees.
Records are deduplicated in source order, with at most 500 retained before scoring. Main and topic listings use static pages of at most 40 entries. Rebuilding the same snapshot produces the same order.
Source order and uptime
Source position has weight 10 and maps linearly from 60 for the first position to 40 for the last; a single retained record receives a neutral 50. Uptime has weight 25: 100 × clamp((u − 80) / 20), where u is percentage uptime and clamp limits values to 0–1.
Missing or invalid uptime, including values outside 0–100%, is excluded rather than clipped into a valid observation. A valid uptime below 80% receives zero for this component.
Latency
Latency has weight 20. Clamp valid t to 1,000–15,000 milliseconds, then use 100 × (1 − log(t / 1000) / log(15)). Lower reported latency receives more points; the logarithmic scale is not a new measurement.
Missing, nonnumeric, zero, and negative latency values are excluded. A positive value outside the scoring interval is capped at its boundary, not treated as missing.
Reviews
Reviews have weight 15. For n reviews and average rating r, the Bayesian rating is (n × r + 35) / (n + 10); convert it with (rating − 1) / 4 × 100. This adds ten prior reviews averaging 3.5.
A valid average must be within 1–5 and have a positive review count. Missing or invalid ratings and nonpositive counts do not enter the score; smoothing does not authenticate reviews.
Catalogue and tenure
Model count has weight 10: log1p(min(n, 40)) / log(41) × 100; only positive counts contribute. Tenure has weight 5: clamp(ageDays / 730) × 100, measured against the snapshot date.
Missing or invalid establishment dates, including dates later than the snapshot, are excluded. Catalogue breadth and age are contextual signals, not proof of capability or reliability.
Commercial features
Payment methods have weight 5: min(count / 3, 1) × 100; an empty list is omitted. Refund and invoice availability each have weight 5, with true scoring 100, false 0, and unknown omitted. These booleans do not assess the quality of the terms.
Missing data and coverage
Missing signals are omitted from both the weighted sum and its denominator. The weighted mean is the available weighted sum divided by available weight. Coverage = availableWeight / 100.
Final score = 50 + (weightedMean − 50) × (0.35 + 0.65 × coverage). Lower coverage pulls the result toward 50, not toward a failing grade. Unknown information is not the same as an explicitly unsupported feature.
Ordering and topic boundaries
Displayed order follows the source dataset rank. The data score is shown for comparison and does not change that order. Displayed scores are rounded to one decimal.
Topic membership comes from keyword matches, not verified capabilities, official affiliation, or successful integration tests. Confirm current support and contractual conditions directly before use.