Design Size Recommendation

Fit intelligence on the product detail page

Size and fit is the single largest controllable driver of fashion returns: roughly 35% of apparel returns are size-related, and every returned garment erodes the margin on the sale. A fit-intelligence system attacks that at the product detail page — before add-to-cart — by turning a shopper's body measurements, purchase-and-return history, and the garment's brand-specific size chart into a recommended size label (numeric or alpha) with a confidence score and a human-readable rationale ("based on your shoulders, this brand runs small — we suggest L").

The constraint that shapes everything is that this sits on the hot PDP path: 18M monthly actives browsing fashion, a p95 budget around 45 ms, so the recommendation must be served at read-path latency — it cannot wait on a slow model call while the shopper stares at a spinner. The real systems here are True Fit, Fit Analytics (Fit Finder), and Amazon's and Zalando's in-house size recommendations; all resolve a per-(shopper, style) size on the PDP and fall back to chart-based guidance when they lack a confident personalized answer.

The defining tension, present from the first sentence, is personalized-but-slow vs chart-based-but-generic. A trained model gives a confident, explained, personalized size but needs a profile and an inference budget; a brand size chart gives an answer for anyone (including guests) but ignores how this shopper actually fits. The design resolves it with a confidence-gated two-path system: serve the personalized recommendation when confident, fall back to the chart otherwise — and the whole system is judged by one number, the size-related return rate. The failure to avoid from the outset is leading with model architecture; the answer opens with the business metric (size returns) and the user moment (PDP, pre-add-to-cart), because that is what the recommendation exists to move.