How SizeGuide decides
The rules behind every chart, conversion, and recommendation we publish. Versioned like our charts, and every product page on this site is accountable to it.
Provenance
Every public chart carries one of exactly three provenance labels. There is no fourth state. A page whose chart cannot earn a label does not publish.
Supplied or verified by the brand itself through a claimed account.
Taken from the brand's own published chart. The source URL and capture date stay on file.
Our synthesis where no published chart exists, built by the rules in this document.
Every chart also shows a last-verified date. Sources are re-checked on a schedule, and drift between our copy and the source flags the chart for review.
Measurement semantics
Body measurements and garment measurements are different things, and we never mix them.
A chart row either describes the body a size is intended for or the garment itself, and that semantic is declared at ingestion, per measurement type. Charts that are ambiguous about which one they mean are rejected at import rather than guessed at, because a chest number that silently switches meaning produces confident, wrong recommendations.
Units and conversion
Charts are stored once, in millimeters, and converted at display time.
Inches, centimeters, and regional size systems all render from the same canonical record, so a correction lands everywhere at once. Cross-brand and cross-region conversion runs through canonical body measurement space, never through pairwise brand-to-brand tables: US 8 to EU 38.5 means "the body that fits US 8 also fits EU 38.5," and the answer carries the provenance of the charts on both sides.
Confidence
Every recommendation ships with a confidence score, and the score reflects how much we actually know.
Input quality
How many measurements the shopper gave, and how they were captured. More and better inputs raise it.
Chart provenance
Official and manufacturer-published charts score higher than reference synthesis.
Margin
How far the shopper sits from the size boundary. Mid-range placements score higher than edges.
Zone agreement
Whether chest, waist, and hip point to the same size. Disagreement lowers it and is disclosed.
Scores are banded for shoppers as high, medium, and low rather than implying false precision. The factor weights are part of the versioned engine, and changes to them change the engine version on every response.
Declining
Below the confidence floor, the engine declines instead of guessing.
Outside a chart's measured range we do not extrapolate: the engine says the chart does not cover you, and does not invent a size beyond the table. Between sizes, we say so, show both options, and let fit preference break the tie. A decline is a feature of the system, not a failure of it: a wrong size costs the shopper a return and costs the store its credibility.
The engine is deterministic: the same inputs against the same chart version always produce the same answer, which is what makes every recommendation replayable and auditable.
Ease rules
Garment-measurement charts are interpreted through published ease assumptions, by category.
| Category | Fit intent | Ease applied |
|---|---|---|
| Fitted tops | close to body | +50 to +100 mm chest |
| Regular tops | standard | +100 to +150 mm chest |
| Outerwear | worn over layers | +150 to +250 mm chest |
| Bottoms | waistband | +10 to +40 mm waist |
Where a brand states its own fit intent, the brand's statement wins over the default, and the recommendation's explanation says which was used.
Versioning
Charts are immutable once published. Changes create a new version; they never rewrite the old one.
Each version carries a content hash and a date, and superseded versions stay on record. Every recommendation ever made logs the engine version and the chart version that produced it, so any answer can be replayed and any dispute can be settled from records rather than memory.
Privacy
Measurements belong to the shopper.
Merchants receive a size, never a body. Measurement data is not sold, not shared with advertisers, and deletable by the shopper in one click. Recommendations work at reduced confidence with no stored data at all, and the confidence score honestly reflects that reduction.
What we refuse to do
The claims we will not make and the shortcuts we will not take, in writing. Read each one as a guarantee: every refusal here runs in favor of your store and your shoppers.
No invented measurements
A value that cannot pass review does not reach a shopper.
No fabricated accuracy claims
We publish no accuracy percentage we cannot show the derivation of.
No promised conversion lift
Lift is measured per store against a control group, and reported whichever way it goes.
No silent machine learning
Outcome data tunes the engine only after real evidence accumulates, and any adjustment is disclosed in the recommendation's explanation.
No dark patterns
No fake urgency, no invented social proof, no hiding the decline.
This document governs
If a page on this site appears to break a rule in this document, the page is wrong. Tell us.