How to serve search results (and host a tasty dinner party)
Imagine you’re holding a dinner party.
You’ve bought all your favorite things to eat and drink; succulent rib eye steaks, some indulgent red wine, and there’s a delicious cheesecake in the fridge, ready to go for dessert.
You just know your guests are going to love it. You’re going to be the perfect host.
But when they turn up, there’s a problem.
The vast majority are vegetarian. Some are teetotalers, and most are lactose intolerant. Your dinner is ruined. Your food is largely uneaten. Your reputation, tattered.
You were so sure they were going to love it. What went wrong?
Obviously, you’ve failed to take your guests needs, context, and preferences into account.
And that’s a bit like manually ordering your site’s search results.
You’re running a website, you know your customer, and you think you know what they’re looking for. So you start tinkering with your search software and manually ordering the results for popular queries, putting the links you think are most helpful right on the top.
“Surely this is a helpful thing to do”, you think, “It’ll all be so much easier to find now. Customers are going to love it.”
But you thought wrong. Your handpicked search results aren’t actually that useful. Customers are upset, sales fall off a cliff, and you don’t even know about all the leads that left without a trace when they couldn’t find anything relevant. Instead, you’re left helping customers locate the content, products, and answers that you know exist on your website already.
Potential customers sometimes ask us whether Sajari allows you to set the order of results manually, one-by-one, but that’s something we’re not going to do. We do allow custom rules and configuration of what data is considered in calculating relevancy scores, but we’ll always stop short of letting users “hand craft” their own results.
Most search software does allow you to do this — they even advertise it as a feature. We don’t believe that’s the best way to do things, and once those potential customers see the results from our search software, they agree with us. We can trust our technology to do the best job. Here’s why:
It’s impossible to know who all your users are, let alone what’s on their mind when they start typing a query.
They might use different keywords, or not know how to describe what they want at all. They might misspell what they’re trying to find or use legacy brand words from an advertising campaign you ran years ago. Add context — which state they’re searching from, the device they’re using, or what time of the year it is — and it gets even trickier.
It’s impossible to know who all your users are, let alone what’s on their mind when they start typing a query
Having one person or team developing “one-size-fits-all” search results just doesn’t make much sense. Like the dinner party, most people are going to end up a bit underwhelmed. It’s not a very good customer experience.
What if though, extending the metaphor, you held a follow-up dinner party?
Now you know more about what your guests want, you could put together a menu that worked better for everyone. It mightn’t be everyone’s favorite foods, but at least you know it fits their dietary requirements. If you continued holding follow-up dinner parties, you could even keep improving your menu as you learned what your guests like and dislike. It’s a much better solution than just serving your favorite foods for dinner.
This is a key feature of most search engines you use today. Instead of a human manually setting search results, the search engine serves its best guess of relevant results and then observes what people do with the results. If a result is clicked i.e. users think its relevant, then it moves up in the rankings for a given query. If it doesn’t get clicked, it moves down over time.
This is a simplified explanation of what’s going on behind the scenes, but it’s the same general principle; by studying human interactions, machine learning models can learn which results are most helpful for each query, and other queries similar to it.
But what if you hadn’t eaten with the guests at your dinner party before?
You wouldn’t know exactly what they want to eat, but if you knew they shared some characteristics with your previous guests, you could probably make a pretty good guess at what they’ll enjoy.
If your guests were to RSVP to your party with their dietary requirements and some demographic data, you could statistically calculate the best menu to put together, based on your knowledge of your previous guests and the menus you’ve previously served. The more dinner parties you have, the more data you could collect, and the more granular you could get with your menu creation.
Male vegetarian, over forty, Italian heritage, for a late September get-together? Make sure you cook that eggplant parmigiana.
Female pescatarian, mid-twenties, no dairy, for a summer BBQ? Sounds like you’re grilling some salmon.
This is a little bit like how we serve search results with the Sajari engine.
We don’t allow manual ordering, and we don’t stop at improving result ranking based on how generic users interact with results.
Instead, our engine takes in as many data points as you tell it to consider, both in your content and on the query side, learns from other results over time, and then serves up the best possible results for each individual user. We don’t even store the personal information on your users; everything is calculated statistically from the data we have access to, for every query, to serve the best search set of search results, every time.
Oh yeah, and it all happens in less than a hundredth of a second.
It’s less work for you, and a better experience for your customers. It could even mean more sales and revenue in your pocket.
The only limit to how good your search results can get is the quality of your data — and we can help you with that too, via our Site Search Health Report.
Why would you want to do search any other way?
Originally published at www.sajari.com.