In the Q4 2025 pilot, shoppers using fit recommendations saw 41% fewer size-related returns. The result supported some of our expectations, but other observations changed what we prioritize when building 3D body models from phone photos for apparel fit.
This report covers the Q4 2025 pilot cohort, measured over 90 days. We will not identify the participating brands. We can share the measurements, the patterns across the cohort, and what the results indicated about where phone-photo body models help with sizing and where they have limits.
The 41% result, in context
Across the cohort, shoppers who used the fit recommendation had a 41% lower size-related return rate than shoppers who did not, for the same products over the same period. This is the figure we publish, and it prompts the most questions.
The key qualifier is "among shoppers who used the recommendation." The 41% does not compare total store returns before and after deployment. It compares shoppers who received and followed a specific fit recommendation with shoppers who chose their size through the brand's standard chart during the same period. The comparison is recommendation users versus non-users, not total store volume. That matters because adoption varies. The effect on total returns depends on how many shoppers use the recommendation, where it appears on the PDP, and how the widget is presented.
Category mix explains part of the signal
Size-return reduction varied by category. Fitted wovens, especially structured trousers and tailored jackets, produced the strongest signal. Casual knits produced a weaker one. The pattern fits the garment construction: structured woven pieces have defined chest seam placement, shoulder width, and waist circumference that must suit the shopper's proportions. A stretchy knit with substantial ease fits more body shapes in one size, leaving the recommendation less to resolve.
This affects how brands should plan fit-recommendation deployment. A range dominated by casual knits with generous ease may produce a weaker ROI than a fitted-woven range where construction precision matters. Fit recommendations still have value for knitwear, but the return-rate difference may be smaller. Brands should account for that when defining a pilot using 3D body models from phone photos.
When a recommendation picks the wrong size
One unexpected pattern was how often some shoppers followed a recommendation and still returned the garment. In several cases, production measurements did not match the stored spec. The published technical spec listed a size medium chest at 40 inches finished, while supplier production runs sometimes differed by 1-2 inches. Our system matched the spec rather than the garment in the box, so the recommendation relied on measurements that were not production-accurate.
We added production-sample verification to pilot onboarding. Before configuring recommendations, we check sample measurements against the brand's spec. This takes work because each SKU sample must be pulled and measured, but recommendation accuracy depends on accurate garment data. A recommendation based on a wrong spec may be no better than a size chart, or confidently wrong, which can reduce trust in the sizing guidance.
Who chose the recommendation surprised us
We expected first-time visitors without purchase history to adopt the recommendation more often, and they did. We also saw meaningful use among returning customers who already knew their brand size. Some had bought and returned an item for reasons other than fit, then entered a new category. They used the recommendation to check whether their known size carried over to construction with different fit demands.
Fit recommendations can support both new-shopper conversion and category expansion within an existing brand relationship. Someone comfortable buying knit tops may still hesitate over fitted trousers. A phone-photo body model can reduce that uncertainty, especially when structured garments offer less room for an existing mental size anchor to guide the purchase.
Why shoppers left the measurement flow
Some shoppers began the body-measurement flow but did not finish. The most common exit came at photo submission. Some were uncomfortable sending full-length photos, even when told the images were processed locally and not stored on our servers. We understand that hesitation and do not dismiss it.
This remains a concern. We are addressing it with clearer camera guidance, including real-time positioning feedback to reduce re-shoots, and direct explanations of photo processing. In the pilot, shoppers who completed the flow used it again: re-engagement among people who had submitted measurements was substantially higher than first-time submission. The main hurdle is the first submission.
What body data showed about grading
The pilot's fit data also revealed product intelligence we had not fully valued. Comparing body measurements from a brand's shoppers with garment specs showed how that population's measurements were distributed against the garment's size grading.
For one pilot brand, many shoppers had a waist-to-hip ratio between the size small and size medium hip grades while matching the size small waist grade. The current grading therefore undersized their hips. The recommendation correctly indicated that neither available size would fit well. The brand used the aggregate result in a grading review with its pattern maker. This was outside the pilot's original purpose, yet proved as valuable as the return-rate data.
