The Challenge
Fashion is personal. Tagging shouldn’t feel robotic. But for a major UK retailer, product metadata was a mess. Dresses were inconsistently labelled. Styles were misclassified. Size tags varied between vendors. All this made online recommendations feel off. The outcome: returns piled up. On top of that, the retailer couldn’t tell which styles a customer truly liked. Browsing patterns, add-to-cart history, and actual purchases told different stories. Yet all recommendations treated these signals equally. The retailer needed more than a recommender. They needed a taste engine.
A Day in the Life: Before Our Solution
A merchandiser would manually tag each new item: neckline, sleeve length, silhouette, occasion. Some were accurate. Others? Not so much.For example, a potential buyer would filter for “baggy barrel leg denim”. but results showed boot-cut jeans. Shoppers would click, get frustrated, and bounce. Those who ordered often returned items due to fit or mismatch. Meanwhile, the CRM analyst was trying to segment shoppers based on past purchases. But one customer’s eclectic history didn’t fit neatly into predefined personas. Customers felt recommendations were generic. And, engagement was dropping. The team had data. Tons of it. But no system to use it with taste, nuance, or style.
Pain Points:
- Manual product tagging was error-prone and inconsistent
- Shoppers received irrelevant size or style recommendations
- Return rates rose due to poor fit and expectation mismatch
- Algorithms failed to personalize for unique shopper behavior
- Internal teams spent hours fixing tags and filtering bad recommendations