AI Size and Fit Recommendation: Cut Fashion Ecommerce Returns in Half

Why fashion returns are a bigger problem than most stores admit

Online apparel returns typically land between 25 and 30 percent, well above most other ecommerce categories. Ask customers why they sent something back, and wrong size comes up first almost every time: not quality, not dislike, just a garment that does not fit. Two distinct problems hide inside that one number: shoppers picking the wrong size for themselves, and brands whose sizing charts do not match their actual cut. A size M at one label runs slim; at another it runs boxy. The chart says nothing about which.

Returns cost more than the refund itself. Reverse logistics runs both directions: shipping out, then shipping back. A chunk of returned items arrive with damaged packaging or visible wear from being tried on, so they get marked down or written off entirely rather than resold at full price. Stores that actually run the numbers usually find returns eating several points off gross margin, with size-related returns making up more than half of that total.

That is why size recommendation has become one of the fastest-adopted AI features in fashion ecommerce. It is not a novelty feature. It has a return-on-investment case you can calculate before you buy the tool, which is more than can be said for a lot of AI features chasing a demo moment.

How AI size recommendation actually works

Most people assume AI sizing means asking customers to enter their bust, waist, and hip measurements. In practice, the tools that work well rarely lean on that kind of self-reported input, because most shoppers will not take the time to measure themselves, and the numbers they submit are often wrong anyway.

What actually drives accuracy is a combination of three data types. First is purchase history: what a shopper has bought before, in which brands, which sizes they kept, which they returned, and why. Behavioral data like this is far more reliable than self-reported measurements. Second is a brand fit database: providers build a library of how brands actually cut their garments, since a size M at Brand A can run three to five centimeters larger in the chest than a size M at Brand B, and that mapping lets the system translate between brands. Third is body input, but usually kept minimal: height, weight, and maybe one or two descriptive cues like “runs narrow through the shoulders.” That input mainly refines the first two signals rather than replacing them.

Combine all three and the output is not a single hard number but a confidence-weighted recommendation: something like “based on your purchase history across three brands we track, size M is a strong match, size L if you prefer a looser fit.” That kind of explained recommendation converts noticeably better than a bare size number with no reasoning attached.

Comparing the four major tools

Kiwi Sizing suits stores just getting started with size recommendation. It installs like a Shopify app and can go live within days, which makes it a reasonable first test before committing to something heavier. True Fit’s advantage is its data network built up over years of cross-brand purchase history. If your customers have shopped other brands that True Fit already tracks, its recommendations start out more accurate than a newer tool can match, at least until that tool has collected its own history.

Fit Analytics, now under Snap (Snapchat’s parent company), has invested heavily in close-fitting categories like intimates and activewear, with a quiz flow that is short but pulls a surprising amount of signal from a handful of questions. Bold Metrics takes a different route entirely, building 3D body models to predict fit with more precision. That approach fits brands with high average order value where the cost of a bad-fit return justifies the extra setup, but it is probably more investment than a small or mid-size store needs to take on.

The real impact on return rates and conversion

Industry-reported figures put size-related return reduction in the 20 to 40 percent range after adopting AI size recommendation, with the exact number depending heavily on category and starting baseline. A category with tightly managed sizing to begin with will see a smaller improvement; a category with messy, inconsistent brand cuts has more room to gain.

The conversion side matters just as much and gets overlooked more often. Merchants worry that adding a size recommendation step introduces friction and slows checkout. In practice, the opposite tends to happen: a shopper stuck deciding between two sizes is more likely to abandon the page entirely than a shopper who gets a clear, confident recommendation. So these tools pay off twice: fewer returns after the sale, and more hesitant browsers converted into completed orders before they leave to comparison-shop elsewhere.

One detail that is easy to miss: presentation matters as much as accuracy. A bare size number performs worse than a recommendation paired with a short explanation: referencing similar customers’ feedback, or noting that a brand runs small or large. Shoppers trust an explained recommendation more, and that trust shows up directly in how often they follow it.

Virtual try-on and the cross-border sizing trap

Virtual try-on has gotten a lot of hype, but most current implementations (uploading a photo to generate an avatar wearing the garment) are not yet accurate enough to replace data-driven size recommendation. Treat it as a visual aid that helps customers judge style and silhouette, not as a substitute for the purchase-history-and-fit-data approach described above. The two can work together, but virtual try-on alone will not move your return numbers much.

Cross-border sellers face a second, harder problem: US, EU, and Asian sizing systems are not built on the same logic, and converting between them is not a simple number shift. A size M in the US can differ from a size M in an Asian market by a full size step in actual measurements. Many sellers respond by posting a conversion chart and calling it solved, but a static chart cannot account for the fact that the same brand’s fit varies between categories: a T-shirt and an outerwear piece from the same label rarely run the same way.

A more reliable approach is letting the AI sizing tool generate region-aware recommendations based on where the customer is shopping from and their purchase history, rather than relying on a fixed conversion table. If your customer base has meaningful volume in both Asian and Western markets, it is worth confirming upfront that your chosen tool actually has fit data covering both regions, rather than finding out after launch that one region’s database is essentially empty.