Semantic query parsing blueprint
Sep 16, 2020 • 10 min read
Sep 16, 2020 • 10 min read
Replacements, Ltd. is the world's largest retailer of tableware, silver, estate jewelry and watches, including patterns being produced by manufacturers as well as those discontinued. Customers looking to expand their collections, along with those who have misplaced or broken pieces, use the company’s vast website (www.replacements.com) to track down specific items.
Efficient, frictionless and delightful product discovery experience is on the forefront of Replacements' business agenda. We already wrote about visual search for dinnerware patterns which we helped replacements.com to build recently.
In this blog post, we will talk about another essential product discovery feature we just launched - smart autocomplete.
Original implementation of autocomplete on replacements.com was primarily based on out-of-the-box Apache Solr capabilities and had issues with coverage and relevance of suggestions. There were suggestions which led to zero or irrelevant search results, repeating and semantically duplicate suggestions, and suggestions which didn’t look like complete search phrases. Suggestion matching logic could only handle suggestions starting with what user typed and didn’t support spelling correction. These were significant shortcomings compared to autocomplete best practices.
Grid Dynamics engaged with Replacements, Ltd. to design and implement a new autocomplete microservice for site search. The goal was to create a self-learning solution which would learn new suggestions based on customer interactions with the search box. In suggestion ranking decisions, we combined the popularity of suggestions with business KPI’s such as conversion rate, click-through ratios, search result page dwelling time.
We implemented a data pipeline which processed search engagement analytic data collected and applied a multi-step process of refinement, deduplication and validation of query suggestions.
On the suggestion matching side, we implemented an autocomplete service that was able to match parts of the query suggestion and support spelling correction for the user-typed prefix. We then perform the two-stage search, gradually relaxing match requirements until we find a satisfactory number of suggestions.
A/B testing of the new autocomplete service showed an improved relevance of suggestions. The new solution drove 37% more autocomplete suggestion clicks, which improved overall quality of search results and engagement with search functionality. This led to overall revenue uplift of almost 1.7%.
Matt Kleweno, Software Development Manager at Replacements, Ltd., says: “Search is a very important component to our website experience. We are continually looking for ways to improve search and help our customers more easily find the products they are looking for.”
Autocomplete is an important gateway into your product experience. As you can see from this case study, improvements in relevance of suggestions can make the customer shopping journey more delightful and move the needle in conversion rates and revenue.
Happy searching (with smart autocomplete!)