How good is this data?

Coverage

Platform counts overlap (multi-platform titles count once per platform). Standalone games and UGC experiences are scored on different rubric profiles and are never summed into a single figure.

Who produced each score

Every rating carries a review tier, shown as a badge at the point of decision and served in the API. The current mix:

Automated scores are produced by a pipeline applying the published rubric, with periodic spot-check auditing; they are not individually verified by a human. That is stated on every rating, not buried here. The methodology (v1.3) documents the rubric, weights, and limitations, and is versioned — every score stores the methodology version that produced it.

Traceability

Every published score is reproducible: the stored review row keeps all 50 per-dimension inputs, composite scores carry the formula version that computed them, adversarially debated titles keep the full debate transcript on file, and since July 2026 every score change is journaled — old value, new value, which dimensions moved, and what triggered the rescore. A score here is never just a number; it is a number with a paper trail.

Update cadence

Ratings refresh through three mechanisms: newly ingested titles are scored on arrival; a metadata watcher flags previously scored titles whose source data changed (needs_rescore); and an age-based sweep re-queues ratings older than 180 days. In the last 7 days, 0 game ratings and 0 experience ratings were created or refreshed. Formal cadence commitment: FILL: re-review SLA — e.g. top titles within N days of a flagged change

Known limitations, disclosed

• Nearly all scores are automated-tier (see the mix above). Spot-check audits and the published rubric constrain drift, but individual scores reflect a model’s application of the rubric, not a human expert’s judgement. The FAQ says this plainly, and so does the badge on every rating.
• UGC experience ratings are only as good as their input signals. When there is not enough verified input data (confidence below 0.5), the score is suppressed everywhere — site and API — rather than shown with a caveat. Currently 61 experiences are held back this way.
• Simulated-gambling content became its own age-floor dimension (R4.6) in methodology v1.2 (July 2026); titles scored before that carried heuristic floors until rescored. The changelog on the methodology page records every such rubric change.
• Store-identifier mappings extracted from third-party store links start as verified: false and are flagged as such in the API until human-checked.