On this page
Every new SKU launches with a fit-data blank slate: no orders, no returns, no size-exchange pattern to learn from. That gap is exactly when customers make their first impression of a brand, and a wrong-size guess at launch can turn a promising drop into a return-rate disaster before the marketing spend even has time to pay off.
Why New SKUs Are the Hardest Sizing Problem
A mature product with a year of sales has hundreds or thousands of data points connecting body measurements, purchased size, and outcome (kept vs. returned). A brand-new SKU has zero. This is the classic "cold start" problem, well documented in recommender-systems research: systems that rely purely on collaborative, history-based signals fail hardest exactly when there is no history to draw on.
For apparel and footwear, the stakes are higher than for a music or movie recommendation. A bad genre suggestion costs a few minutes of a customer's attention. A bad size suggestion costs a return shipment, a refund, a restocking cost, and often the customer's trust in the brand's entire size range going forward.
The Cost of Guessing on Day One
Launch windows are when traffic and conversion intent are at their peak, which means launch-week sizing errors get amplified rather than absorbed. A misjudged size chart on a bestseller-in-waiting can generate a spike of returns right when a brand has the least operational slack to handle them — new SKU, no historical return workflow, and often limited stock to re-ship the correct size.
Building a Size Recommendation for New Product Launch SKUs Without History
The fix is not to wait for data to accumulate. It's to substitute other signals for direct sales history until real data arrives.
1. Start From Garment Specs, Not Guesses
Every new SKU has a technical spec sheet: chest width, waist, inseam, ease, fabric stretch percentage. These specs are a legitimate substitute for sales history because they describe the same physical reality that sales history would eventually reveal — just faster. A size-recommendation engine that ingests flat measurements, garment construction, and stretch/fabric behavior can generate a fit prediction on day one, before a single unit sells. For a deeper look at what an engine actually needs to ingest, see what data a size recommendation engine needs.
2. Transfer Learning From Similar SKUs
Very few "new" products are truly novel. A new colorway of an existing style, a new fabric weight in the same cut, or a new style in a familiar category can all borrow fit intelligence from their nearest neighbors. Mapping a new SKU to its closest existing product by cut, fabric, and construction lets a recommendation engine transfer confidence scores and adjustment factors instead of starting from zero.
| Cold-start strategy | Data source | Confidence at launch |
|---|---|---|
| Pure guesswork / generic size chart | None | Low |
| Garment spec ingestion | Tech pack, flat measurements | Medium |
| Transfer from similar SKUs | Existing catalog fit history | Medium-high |
| Blended model (specs + transfer + early returns) | Specs + catalog + first-week orders | High |
3. Blend in Early Signals Fast
Once the first orders start coming in, the system should not wait for a full return cycle to update. Order-level signals — size selected relative to a customer's known measurements, early exchange requests, first support tickets mentioning fit — can be fed back into the model within days, not months. This closes the gap between "pure cold start" and "steady-state, history-rich" recommendation quality far faster than waiting passively for return data to accumulate.
Operational Steps Before Launch Day
Audit the Spec Sheet Before It Reaches the Store
Fit accuracy at launch starts upstream of the storefront, in the tech pack. If specs are incomplete or inconsistent with prior styles in the same size run, the recommendation engine inherits that error on day one. Tightening this process is part of a broader discipline covered in how to reduce fit errors before production — the earlier fit issues are caught, the cheaper they are to fix.
Plan the Launch With Returns in Mind
A launch plan should treat size recommendation as a first-class risk, alongside inventory and marketing. That means setting a return-rate target for the SKU, staffing customer service for size-related questions in week one, and having a fast feedback loop from returns/exchanges back into the recommendation model. For a fuller checklist, see how to plan a product launch that minimizes returns.
Set Expectations With the Merchandising Team
Merchandisers often want a new SKU's size chart locked weeks before launch. Build in a review step where the recommendation engine's spec-based prediction is checked against the merchandiser's intended fit — this catches inconsistencies (like an inseam accidentally copied from the wrong base style) before they reach shoppers.
Measuring Cold-Start Accuracy
Because there's no return history yet, brands need a proxy metric for launch-week sizing quality. Two useful proxies:
- Exchange requests specifically citing "too small" or "too big" in the first 14 days.
- Size-guide engagement rate versus the eventual conversion rate for that SKU.
Both are leading indicators that show up well before a full return cycle completes, and both can be tracked from day one of the launch.
According to industry data on apparel returns, sizing and fit remain the single largest driver of returns in fashion e-commerce, and the problem is worse for new items with no track record — a trend widely discussed in retail research from sources like McKinsey's retail research and cold-start recommender literature, which consistently shows that systems relying solely on historical interaction data underperform sharply for brand-new items until auxiliary signals are introduced.
Frequently Asked Questions
How accurate can size recommendations be for a brand-new SKU with zero sales?
Accuracy at launch depends heavily on the quality of the garment spec sheet and how closely the new SKU resembles existing catalog items. A well-specified SKU that maps closely to a proven style can reach accuracy close to that of an established product, while a truly novel cut with sparse specs will start lower and improve quickly as the first orders and exchanges come in.
Do we need a full season of sales data before size recommendations work?
No. Waiting for a full sales cycle means absorbing the worst returns exactly when the product is newest and most visible. Spec-based prediction and transfer learning from similar SKUs give a usable starting point on day one, and the model should be designed to absorb early order and exchange signals within the first weeks rather than waiting for a complete return cycle.
What's the single biggest lever for improving new-SKU sizing accuracy?
Clean, complete garment specifications submitted before the SKU goes live. Most cold-start sizing failures trace back to incomplete or inconsistent tech packs rather than a limitation of the recommendation logic itself, so tightening that upstream process usually delivers the biggest accuracy gain per hour invested.