Product scoring methodology - SplitAtlas
Principle: scores are computed from normalized structured product data. No hand-tuned per-product numbers and no sponsorship influence. Published pages show the methodology link, eligible category scores, and a data-completeness indicator.
Category weights
Cooling efficiency 15% . Heating efficiency 15% . Cold-weather performance 15% . Noise 10% . Warranty 10% . Installation accessibility 10% . Feature set 10% . Price-to-performance 10% . Documentation and parts availability 5%.
Normalization rules per category
All normalizations are min-max within capacity class and installation type peer groups (a 9k DIY unit is scored against 9k-class units, not 36k), onto 0-10, using the 5th-95th percentile of the peer group to resist outliers.
- Cooling efficiency: SEER2 primary, EER2 secondary (70/30 blend).
- Heating efficiency: HSPF2 primary; COP at 47F secondary when published.
- Cold-weather performance: composite of capacity retention (rated heating at 5F divided by rated at 47F), minimum operating temperature, and published low-ambient rating existence. Missing all three means the category is not scored, never imputed.
- Noise: indoor minimum dB(A) (60%), outdoor dB(A) (40%); lower is better.
- Warranty: compressor years plus parts years, registered terms; dealer-only warranty restrictions noted and penalized for DIY-class units.
- Installation accessibility: DIY status, precharged line set, included line-set length, 120V availability, weight class.
- Feature set: count of verified supported features, weighted (inverter, Wi-Fi, low-ambient kit, dehumidify, etc.).
- Price-to-performance: median current price record divided by capacity and efficiency blend, inverted. Stale prices (older than 60 days) are excluded; category is not scored if no fresh price is available.
- Documentation and parts: manuals on file, spec sheet on file, error-code coverage in the model database, parts-channel availability flag.
Peer-group minimum
Min-max percentile normalization degenerates when a peer group is small; with 1-2 members everything scores neutral. Overall and category scores are recalculated for trend history, but they display only when the system's peer group contains at least 4 rated systems. Below that, pages show "not yet rated - we score against comparable systems, and this class needs more verified models first." Non-public test systems are excluded from scoring.
Confidence gating
- Each category has required fields. A category with missing required fields is shown as "insufficient data" and excluded from the overall roll-up (weights renormalize over non-null categories).
- Overall score displays only if at least 7 of 9 categories are non-null and overall field completeness is at least 80%. Otherwise the page shows category scores that exist plus a "not yet rated - data incomplete" state. A system must never look mediocre because data is missing; it looks unrated.
- Score history is retained internally so methodology changes can be audited.
Governance
- Weight or normalization changes require a new methodology version, a changelog entry on the public methodology page, and owner sign-off.
- Sponsored status is invisible to scoring by construction; sponsorship fields are not included in scoring inputs.
- Editorial "best for X" picks may consider scores but are labeled editorial judgment and never presented as computed output.