By some estimates, more than half of the visitors to an online store will use that store’s site search to find products, and hopefully make a purchase. This fact alone makes ecommerce site search important.
I recently spoke with Amir Konigsberg, co-founder and CEO of Twiggle, a search engine that helps ecommerce site search think like shoppers.
Practical Ecommerce: What is a good site search for both customers and merchants?
Amir Konigsberg: The benefits of site search is that it is the most personal touch point a merchant and customer have in the buying process. It’s the place where a customer is telling the merchant what they want, and if that experience is a bad one, the merchant may never see the customer again. Site search is about helping people and reaching the handshake moment between a merchant and a customer.
PEC: What are some of the shortcomings of current ecommerce site search?
Konigsberg: Sometimes keywords appear in product pages without those pages actually being the relevant ones to the keyword. If someone searches for “bike helmet,” what they are looking for is a bike helmet, but a search engine would search for the two words “bike” and “helmet.” It might pull out “bikes” and not “bike helmets” simply be because the word appears there.
The other downside of this is that bike helmets, which the merchant does have in stock, are not showing up when they should be showing up. And that means the shelf costs of these products is going up.
PEC: What about helping shoppers understand how to word a query?
Konigsberg: Today people think about how to type so that the engine will understand them, and they’ve become accustomed to a kind of guided search approach where they often start off with some generic keyword. Another approach is a suggestion mechanism, which we’re all familiar with from Google, when you start typing and then the search engine automatically suggests relevant results.
We try to think what kind of experience a person should have with a search engine, to make it as close as possible to the kind of positive experience they have when they are actually talking to an in-store salesperson.
PEC: Can you relate site search to natural language processing tools?
Konigsberg: The idea of a chatbot is exactly what we’ve just described, where you can actually have a conversation with an automated system that mimics a conversation that you’d like to have with a person. Chatbot technologies are typically narrowly focused on kind of frequently asked questions and frequently given responses.
Twiggle’s technology actually understands the products. Our platform learns each and every feature of products, each and every characteristic, so it knows how to compare products at the feature level, at the characteristic level, and at the price level, at the value level. And when it knows all these things, the type of responses it can give back to even very complex questions, or very human-like queries, is astounding.
PEC: How can an artificial intelligence tool learn like you’re talking about?
Konigsberg: You can use different methods to do this, but you have to first of all be able to know about products. You have to have information and data to work with. Getting good data about products is not always easy. And that information isn’t always consistent. It’s not always what we call “clean data.”
After we’ve cleaned it and normalized the data, we have to find the uniform way to present data so that no matter how someone searches for a product, the right product will appear in the search results irrespective of how that product is actually described on the product page. And that’s super important, because someone might search for “big laptop screen” and a smart search engine has to understand what a big screen is.
PEC: How do you know if the search is working?
Konigsberg: The standard question here is to measure things like conversion rates and click rates and average order value.
It starts when someone searches for something and then it ends when someone actually buys something. If you only measure the end part — the conversion rates and the average order values — it’s not the optimal way to measure that whole process. At the end of the day that process has several physical stages.
The first stage is when someone searches for something, and you look at the quality of the results that they get, which is measured by the click rate on those results. After that, the person then visits a product page if the results were relevant. When they visit a product page, they can either leave or they can continue. You have to dissect the whole process into very small elements.
Amazon, for example, has done amazing things with its search because they have so much user data. So many search queries go through the Amazon system — very popular queries and long-tailed queries. And because there’s so much data distributed, the search can learn.
But that data is not shared outside of Amazon. At Twiggle, however, we give companies the search quality which will take them to the top of the market level, whether or not they have that data in their own system.