Search That Understands Intent
a mid-market home goods retailer
conversion lift on search traffic
fewer zero-result searches
higher average order value
Site search matched keywords, not intent. A shopper typing "cozy reading corner lamp" got zero results while the catalog held dozens of fits, and misspellings or natural phrasing failed outright. Search users converted well when results were right, which made every failed query expensive, and merchandisers burned hours writing synonym rules that never kept up with how customers actually talk.
How we approached it
Mined a year of search logs to quantify zero-result queries, misrankings, and the revenue attached to failed searches.
Built hybrid retrieval combining semantic embeddings with keyword matching, so natural-language queries and exact model numbers both resolve correctly.
Added LLM-based query understanding to interpret intent, expand vague phrasing, and apply filters like room, style, and budget automatically.
Trained a re-ranking layer on click and purchase behavior, with an offline evaluation harness so every change is measured before it ships.
“Customers type the way they talk now, and search just gets it. The queries we used to lose entirely became some of our best-converting traffic.”
No. The system runs hybrid retrieval, so exact matches like SKUs and model numbers still resolve through keywords while natural-language queries flow through semantic matching.
An offline evaluation harness scores every ranking change against a labeled set of real customer queries, and only changes that improve those benchmarks proceed to an A/B test.
Most were retired. Query understanding handles phrasing variation automatically, and merchandisers now spend their time curating seasonal boosts instead of maintaining rule lists.
Want results like these?
Book a free consultation and we'll map the highest-leverage place to start for your team.