1. A "Reference Brand Mapping" engine supplanted earlier "manual measurement prompts" in the recommendation pipeline. High return rates of first-time buyers are frequently caused by their 'measurement anxiety' and/or errors caused by human measurement with a measuring tape. The back-end sizing ontology normalised the fit profile against specific brand garment specs for the user by allowing them to select the brand and size they already own, so they could utilise the fitting algorithms. By architecturally transitioning to treating sizing as a relational data problem, rather than a physical sizing problem, we created a more reliable sizing experience when a user is new to a brand's cut. 2. The metric that was the strongest indicator for us that there was an actual improvement and not just selection bias is the "Second-Purchase Velocity" delta between the first delivery and the second order. We have observed that the elapsed time to make a second purchase for first-time buyers who used the recommendation engine is 18% quicker than for those who manual selected their size, while controlling for initial cart value. If the lower return rate was due only to the capability of a shopper being a "decisive" shopper, we would not be able to observe this statistically significant acceleration in repeat purchases across all body type clusters. In developing a recommendation engine, there are technical aspects and user experience concerns regarding how to make it trustworthy to a stranger. When developing apparel technology, the key to success will be to continue to reduce the cognitive load on a buyer when making that initial, high-stakes transaction. Once an organisation recognizes the friction inherent in the measuring process, it will better understand how to design a system that meets customers where they currently are.
Feedback on post-purchase fit, which was directly related to the recommendation model, led to the biggest decrease, namely, a 48-hour fit check, which included three binary questions as well as a free-text note. The feedback was returned to the size engine prior to the subsequent browsing. The difference was important since it addressed silent incompatibilities that body-shape clustering overlooked like fabric stretch tolerance and preferred drape. New customers often settle with an approximately fitting one, and come back later after the second time of use. Intercepted feedback at an early stage detected that pattern. The persuasive message was a clean holdout. The fit check with no update of the downstream model was given to ten percent of first-time orders. The remaining ten percent was the initiator of the complete loop. The returns related to size decreased by 17 percent in the treatment condition compared to 3 percent survey-only holdout returns after thirty days. The average order value remained constant, and that eliminated self-selection as a factor in determining the size of a bargain. The treatment cohort experienced an increase in repeat purchase within sixty days of 6 points implying comfort as opposed to discount chasing. The effect of body-shape clustering was an aid in acquisition. The loop feedback modified behavior once reality set in. The effect was not by chance, as the metric separation across holdouts substantiated that.