Exactly how a pile of specs becomes a single, comparable 0–100 score — explained end to end.
The 0–100 score is pure arithmetic over validated specifications. The same specs always produce the same score — there is no AI in the verdict, no opinion, and nothing that drifts between visits. AI is used only to gather and draft product overviews; it never decides who wins.
Specifications are researched from manufacturer documentation and other reputable public sources, then strictly validated against each category's schema, with an anomaly auditor catching out-of-range values before they can ever affect a comparison. A value that can't be verified is left blank rather than guessed.
Each comparable spec earns a 0–1 score that blends two signals, weighted 70/30. ABSOLUTE merit measures the value against fixed real-world anchors — a middling, expected value sits at ~0.5 and a best-in-class value at ~1.0 — so the number means something on its own. RELATIVE advantage measures the head-to-head margin against the other products through a bounded curve: a tie sits at 0.5, small gaps move it a little, and large gaps approach (but never quite reach) the extremes. When a feature has no calibrated anchors, it falls back to the head-to-head margin alone.
Features are grouped into sections (e.g. Display, Performance, Battery). A section's score is the weighted average of its features, and the overall score is the weighted average of the sections using category-specific section weights — so what matters most in that category counts most. Non-comparable, informational specs (like a CPU socket or the OS name) are shown but never scored.
Equivalent-but-different tech is treated equally: a feature satisfied by a substitute counts as satisfied (Face ID counts like a fingerprint reader). Cross-platform values that aren't meaningfully comparable (an iOS version vs an Android version) are treated as a tie instead of scoring one platform's number against the other's. Where ecosystems differ in real-world terms, values are scaled to compare on equal footing, and magnitude specs are judged as ratios (e.g. weight per display inch) so size classes compare fairly. Correlated specs are dampened so a shared signal isn't counted twice.
A missing value is a mild penalty, scored on the same scale as published values and anchored to the field's weakest published spec — enough that a sparsely-documented product can't out-rank a fully-specced one, without nuking a product over a single genuinely-unpublished detail. Features that don't apply to a product (ANC battery life on wired headphones) are skipped entirely, with no penalty.
Raw scores are mapped onto the display band per category, calibrated so the best product in each category lands near the top — but no product is ever auto-100, leaving headroom for a future, genuinely-better product to edge above it. The result is a score that means roughly the same thing whether you're looking at laptops or TVs, whose raw ceilings naturally differ.