Guide Search Users with AI
My work at Shopify is not yet released and currently confidential. Until I can speak freely about it, this post outlines its underlying UX thinking. This post describes how we can use AI to guide the search user as they move between broad and narrow intents.
User Intent
When a buyer comes to search, they bring more than just their wallets and an idea. People are complicated — they have a query they might type into the search engine, but that query can be characterized in many different ways:
Categorical: seeking all products of a kind, like “trail running shoes”
Compatibility: looking for products that work with something they already have, like “pasta attachment stand mixer”
Attribute: searching for products that meet specific criteria, like “gaming pc 16GB ram”
The query, however, is just one way to express higher-level intent: the thing that they want to get done. So their intent might be to "buy a gift for mom." Or to "replace a part that's broken." That's not what they're going to necessarily type into the search engine today, but that is what they have in mind.
We provide better results when we know the intent behind the query.
And of course people are unique individuals whose context plays a large part in that purchase. If we know that they are in a particular country, or that they've searched with us before, we can use this to tailor the experience, and increase the chances of conversion. In this way, even two individuals with the same intent will end up on different journeys. And all of this can happen before they see a product or even land on our app.
The role of the search engine
If you were a sales associate in a department store and a customer needed help finding something, you would go beyond pointing to a set of items on the shelf. You’d try to understand their intent and ask follow up questions to get to more specific offerings. It’s more of a conversation.
By classifying the user query, we can achieve something like that. We infer what we can and offer the pieces from the catalog we think have the best chance. When a user makes a new search or clicks on a result, that’s their side of the conversation and we get a little more signal to work with.
Search whole page experience
So given all this data from the buyer, we can make our way into result land. Here’s an oversimplification if you ever saw one:
Two search result pages can show the same products in very different ways.
Search has a set of products that we rank and retrieve. We also have search features in the form of filters, and suggested searches. My aim with this work was to expand our search feature set and provide additional tools for our users to get from where they're going to a checkout. (This is the part I can’t get into in-depth.)
And finally, we have the presentation piece, which is the method by which we display the results to users. These results have metadata like images and tiles and ratings reviews and can be arranged in certain ways.
Our aim should be a search whole page framework. This framework combines the buyer's intent with search features, search query understanding and search personalization. So the search page is composed of these search features that respond to the buyer's context, their intent, and even their place within the funnel.
Search results can render differently to best match the user’s intent, and thus, their needs
This means any combination of buyer data can be dynamically matched with these things in result land. To put it another way: for every query we can dynamically rank and select the search results and the search features that will matter most for that moment. We combine this with the result set and compose the page around it. This just-in-time composition means that we can trigger one set of search features for users who are searching broadly (and might need help exploring) and another set for users who are close to purchase (and need help comparing and deciding).
And these search features operate in a couple of ways.
One general way is to associate products with each other. Another might be to suggest pathways a user can take, from query to query, that will lead them to a successful outcome. And another might be to provide feedback to the user that we understand what you’re searching for and we’re trying to bring you along that journey. We’re trying to help you get what you need quickly, efficiently, and in a delightful way.