Improved product search relevancy using machine learning models

Products from different categories with 100s of metadata fields
Customers try to search on terms that could exist in any of the fields
Land the customers on the most relevant set of products that are of interest to them
Enterprise search engines are meant for full text searching and can show irrelevant matches
Searching over 100s of fields and scoring them for relevancy is expensive and slow

Conducted comprehensive review of 5M+ past search phrases, defined “success” using search-to-cart & identified gaps
Built a search phrase parsing microservice using natural language processing models, customized for industrial products, for detecting user intent
Leverage user intent to direct the search engine to prioritize fields to search, dramatically improving relevancy
Leverage customer interactions in-session to provide real-time complementary and competitive product recommendations
Launched NLP and recommendations microservices using AWS SageMaker for testing and Kubernetes containers for autoscaling
Proven results with A/B testing and gradually routed traffic to B

Combining AI with traditional search enabled separation of intent and “full-text” search
Significant reduction in search abandonment rate and improved customer satisfaction
$120M incremental revenue, measured through A/B tests
8% increase in market share
13% improvement in search-to-cart metric
ML & search teams have distinct roles in driving search improvements
Customer interactions used in real-time to drive more personalized recommendations
Improved turn around for search enhancements