Unbxd, Site Search and Ecommerce
I’ve taken a decision to begin writing at an exciting time at Unbxd. We are growing really fast and have a great opportunity to solve some of the hardest problems software can solve. And solve them well.
Today, Unbxd has grown its product portfolio to solve product discovery — every onsite interaction that visitors engage in to find products. We power all of them — search, browse, recommend on multiple devices. Our engine determines the most relevant set of products to the right visitor at the right time on the right device. We also allow retailers the ability to control the UX of how they want to show products and maybe even influence the assortment, filters and even banners with merchandising.
However, when we started off — it was all about Site Search. Everybody asked us — why Site Search? What’s so special about Site search that appears so hard or important to solve? We’ve raised 4 rounds of funding — I’ve answered this multiple times, thought I’ll start out answering that question.
First of all, everybody has a search box. Every single eCommerce site. There’s good reason for that. Some sites see 10% of all visitors using search while others have 90% using search. However, almost all sites have nearly 40% of total revenue originating to user sessions with site search. 40% is a huge number — it can make or break the business model of an eCommerce company.
I searched for something specific on eCommerce sites — ‘bluetooth noise cancelation earphones’. I’d either get ‘no results found’ or gibberish — either earphones that were not wireless or I’d get wireless earphones without noise cancelation. I always assumed the site didn’t carry what I was looking for.
So, I go back and Google it. Well, the product appeared in search results with the very site I tried searching earlier. How can Google do a better job searching through the site than the site when they have 90% more metadata about products and product information?
Turns out, site search is not just about looking for words. Millions of types of searches querying long and short descriptive product information meant that ‘looking for words’ approach to site search wasn’t enough. To build and understand intent, you needed many many complex search techniques and apply them differently on the diverse product data.
What does that even mean? You just need to scale what you’d otherwise do — build rules for every single search query manually. The best way to scale is to programmatically solve different types of search queries automatically. You identify the different types — for instance, queries that have features (bluetooth). Queries that have specification (wireless). Maybe club different queries together as similar types — ‘bluetooth noise cancelation earphones’ and ‘wireless noise cancelation earphones’.
The only way to do it is to use data to automatically learn each type and solve them differently for better relevance. This is now fashionably called semantic search using machine learning. Instead of just creating fixed rules for just a few queries, you now have a machine that solves millions of queries — new and old. Basically, using more compute and more data you replace manual effort automagically with relevant search results.
It sounds good in theory. How real is it?
When we got our first few customers, we saw that it just changed the game for online retailers. They had more conversions — more people buying stuff with the same traffic. As much as a 50% jump in 2 weeks. We got a company to grow 25% top-line revenue (50% of total online revenue was search) in 2 weeks by delivering a 50% lift in search revenue. We immediately knew this was game changing for an online retailer. Thus began our journey: validated by a real problem with a real solution.