The e-tailer had built its range into a genuine strength: 85,000 SKUs covering nearly any home-goods need. The site's discovery tools never kept pace. Search demanded exact keywords, so a shopper had to know a product's name before the site could find it, and the alternative was a category tree running seven levels deep. Shoppers who arrived with a feeling rather than a term, softer light, a warmer room, had no path to the products that answered it.

The cost showed up in the session data: 68% of visits ended without a single product page view. A question like 'something to soften a dark hallway' returned no useful results, so the shopper behind it simply left. The team could see demand arriving and could see it leaving, but the search stack in between had no way to connect a vague need to a concrete product.

Vague queries were not an edge case, they were the everyday reality of shopping for a home. People buy home goods to change how a room feels, and few of them can translate that feeling into the vocabulary of a product database. Every one of those failed searches was a shopper the merchandising team never got to serve, and no amount of manual curation could cover the ways 85,000 products might answer an open-ended need.

We built a generative shopping copilot that meets shoppers where they actually are: mid-thought, describing a need. A customer types what they want in plain language and receives three curated suggestions, each with plain-English reasoning that explains why the product answers the brief. The conversation feels like talking to a well-informed shop assistant rather than operating a database.

Under the surface, discipline matters more than flair. The copilot retrieves candidates from a vector index built over the full catalogue, so meaning, not keyword overlap, decides what surfaces. Every answer is then grounded in live product and stock data before it reaches the shopper, which prevents the assistant from inventing items or recommending anything that cannot actually be bought. When a conversation turns to order issues, the copilot hands off to human support rather than guessing.

We treated the existing keyword search as a partner rather than a rival. Shoppers who know exactly what they want still get the direct route, while the copilot absorbs the vague, exploratory queries that used to end in zero results. That division of labour let the copilot concentrate on the conversations where it adds the most value.