Semantic Search in a Fashion E-commerce Site
A fashion marketplace finds that 28% of searches return zero results or irrelevant results. Users type the way they speak: 'dress for a beach wedding', 'something comfortable for flying', 'trainers like the ones in the photo'. The current search engine only finds exact word matches in the product title.
Where the money comes from
What to do now
- 1Measure revenue lost from searches with no result.
- 2Launch hybrid search (words + meaning) on 20% of traffic.
- 3Reorder the top results by margin and stock.
For your committee
One page with the decision, the impact, the risks, the steps and the KPI to track.
The print dialog opens: choose "Save as PDF".
Watch out: more volume is not always more profit
Showing first what is most similar even if it is out of stock in your warehouse or has minimal margin. Combine relevance with margin, own stock and return rate in the ranking of results.
KPI for the committee
Margin per search and rate of searches with no result.
Detail for your data team
Proposed technique: Embeddings + Vector Search
Why: The problem is one of meaning, not words. Embeddings place 'dress for a beach wedding' close to 'long white linen dress' even though they share no words. Vector search finds the neighbours in milliseconds across hundreds of thousands of products.
Cost of getting it wrong: Every search with no result is a customer who goes to the competition. A search engine that shows out-of-stock products causes frustration and returns.
Data needed: Catalogue: 300,000 products with title, description, attributes and photos.; Search logs: 50M queries with clicks, add-to-carts and purchases.; Searches with no result and immediate abandonments after searching.