A womenswear label in Leeds sells a linen shirt dress in six sizes. Across one season, the size 12 comes back more than any other size in the run — not because the stitching fails, but because on this pattern a 12 sits closer to a generic 10. Shoppers order their usual size, it arrives wrong, and goes straight back.

Apparel is consistently reported as the highest-return category in UK online retail, and most of it is fit, not damage or dislike. Shopify reported European GMV growth of 34% year over year in Q2 2026, compared with 28% in North America. The company also highlighted AI as an important area of commerce innovation, but did not attribute those regional growth rates specifically to AI-powered fit or returns features.

Key takeaways

• Shopify reported European GMV growth of 34% year over year in Q2 2026, compared with 28% in North America; the company highlighted AI as an important commerce innovation area but did not attribute those regional growth rates specifically to AI-powered fit or returns features.
• US e-commerce reached 17.1% of total US retail sales, per the US Census Bureau; fashion lags partly because fit resists a screen.
• Shopify said that, over the past five years, merchants reaching $1 million in annual GMV had a 92% retention rate.
• Fit-prediction tools learn from a store’s own order and exchange history at style level, not one size chart applied across a brand.
• The cheapest fix for a small label is rarely software: consistent photography, explicit measurements and honest fit notes cut guesswork without a model.

Why do most fashion returns start with a size, not a fault?

Most apparel returns are triggered by fit uncertainty at the point of purchase, not by a faulty garment or a change of heart. A shopper cannot try before buying, so they guess against a size chart built for an average body rather than for how this particular style is cut and graded.

That guess is the entire transaction risk in fashion e-commerce. A generic chart — bust, waist, hip in three columns — answers a question nobody is asking, because two dresses from the same brand can be graded by different factories and fit differently under the same label.

What is AI integration for eCommerce actually doing when it predicts fit?

AI integration for eCommerce, in a fit-prediction context, is the use of a store’s own order, return and exchange data to recommend a size for an individual shopper on a specific product, rather than displaying one static chart. It is pattern-matching on the store’s own history, not a universal model of bodies.

The system does not know anatomy; it knows outcomes. It watches which orders were kept, exchanged up, exchanged down, or returned outright, then correlates that against whatever the shopper told the store — height, usual size elsewhere, sometimes a photo.

How does a returns history turn into a fit prediction?

The model treats every past order as a labelled example: garment, size ordered, and outcome — kept, exchanged, or returned — become training data. Given enough examples per style, it learns that shoppers who kept a 12 tend to share certain measurements, then applies that to a new shopper entering similar numbers.

The exchange matters more than the return. A straight return only says the garment came back; an exchange from a 12 to a 14 says the 12 ran small on that style. Flat garment measurements separate a fit problem from a taste problem — a reason code alone rarely does that.

Why does fit prediction work per style rather than per brand?

Fit prediction runs at style level because grading is not consistent within a brand’s own catalogue. A size 12 trouser and a size 12 wrap dress can come off different base blocks, sometimes cut by different factories, so one brand-wide size profile will be wrong for part of the range.

A tool trained per style needs enough clean, representative order and return outcomes for that product to make a useful recommendation; there is no universal order-count threshold. A brand with three colourways of one bestseller can build that quickly; one launching forty styles a season, in runs of a few dozen units, cannot.

Why can’t most independent labels use this yet?

Independent labels usually lack the volume and clean records that fit prediction needs. Returns get logged inconsistently, exchange reasons are rarely captured beyond a tick box, and a young label’s order history is too small and too season-specific for a per-style model to trust.

A data-hygiene problem hides behind the volume problem. Returns run through email and a spreadsheet do not produce the structured record a prediction tool needs; they produce a folder of half-typed notes — as much an eCommerce automation gap as a data-science one.

What can a brand with only a few hundred orders do instead?

Three unglamorous fixes usually cut returns faster than any model: consistent photography shot against the same model reference so proportions read the same way each time, explicit garment measurements taken flat rather than a generic size chart, and plainly written fit notes that admit when a style runs small.

None of that needs machine learning, only discipline and testing what is already on the page. Run it as a proper Shopify conversion rate optimization exercise — heatmaps of size chart hesitation, A/B tests of a plain fit note against the generic one — and the leak usually surfaces within a few hundred sessions, well before there is enough data to train anything.

Can AI integration for eCommerce fix a returns problem caused by bad factory grading?

No. AI integration for eCommerce can route a shopper away from a size that has historically disappointed people like them, but it cannot correct a garment that is inconsistently graded between production runs. If the product varies, the prediction is learning to compensate for a manufacturing fault, not describing one style honestly.

That matters commercially, not just ethically. Returns cost margin twice — reverse logistics, then the carbon of a second delivery — and a brand that lets prediction paper over inconsistent grading is dodging a quality-control fix that will keep generating returns anyway. The cheaper route: tighten the tech pack, hold factories to it, then judge whether prediction adds anything.

Frequently asked questions

Do shoppers notice if a fit tool sits behind a size recommendation? Most don’t, and don’t need to. It usually looks like an ordinary size-guide widget with a suggested size highlighted.

Is fit prediction the same as a virtual fitting room? No. A fitting room visualises how a garment looks on a body; fit prediction only recommends a size from historical outcomes. Vendors market them similarly, which causes the confusion.

How much order history does a brand need before this works? There’s no fixed threshold, and any figure quoted as one is a vendor estimate, not a proven benchmark. The honest test is a consistent exchange pattern, not a total order count.

Figures are drawn from Shopify Inc.’s Q2 fiscal 2026 earnings disclosures and US Census Bureau e-commerce retail reporting. Editorial brief; not commercial advice.