Skip to main content

The internet is a constant recommendations machine — but it needs you to make it work

Why can’t the internet tell you what to read, watch, and eat? Because you’re complicated

Share this story

There’s more content than ever out there — and it’s harder than ever to know what to watch.
There’s more content than ever out there — and it’s harder than ever to know what to watch.
Illustration by Grayson Blackmon / The Verge

Spending hours on endless side-scrolling rows of Netflix movies or hunting through the forever-long lists of identically rated restaurants on Yelp — this can’t be the way it’s supposed to work. Part of the whole promise of the internet is that platforms and services would take the web’s infinite supply of everything — the stuff to watch, read, look at, play with, buy, eat, invest in, comment on, listen to, or have feelings about — combine it with a deep understanding of who you are and what you like, and feed back to you an endless supply of all your favorite stuff. 

When it works, it can feel magical, like the TikTok algorithm that seems to know you better than you know yourself. But that’s pretty rare. More often, you’re chased around the internet by Amazon ads for products you already bought, or you’re stuck flipping through hundreds of 3.5-star Yelp listings or a hundred same-sounding true-crime podcasts on Spotify just to find something you like. Or you just end up watching The Office. Again.

Good recommendations seem like a simple enough problem, right? The companies and platforms working on these personalization machines say it’s a harder problem than it seems. Mostly because humans, you see, are tricky to figure out. But they also say there’s a way to do better. And a way you can help.

When the team at the content recommendations app Likewise first started building its platform, it thought the best way to do recommendations was to build a social network. “What happens in real life,” says Likewise CEO Ian Morris, “is you go out to lunch or dinner, and the first thing after the ‘how are you doing, how are the kids’ is you’re talking about things you’ve read or that great new show you watched or a podcast you really need to start listening to. That’s life!” Online, he felt, those human connections and recommendations had been replaced by bad algorithms that optimized for engagement and growth over actual quality content. He thought Likewise could be a resource for finding movies, shows, books, and podcasts, all in one place.

Morris is still convinced that was the right approach. It didn’t take off as fast as he’d hoped, though — building a social network from scratch is seriously hard work — and so Likewise started to think about how to make the platform more useful even for those who didn’t have a big group of Likewise-using friends. It hired an editorial team to scour the internet for the best and most interesting new stuff and simultaneously began building a machine-learning system that could make automated recommendations.

Likewise collects all the stuff you want to watch and all the stuff it thinks you should watch.
Likewise collects all the stuff you want to watch and all the stuff it thinks you should watch.
Image: Likewise

Now, when you first start using the Likewise app, it requires you to tell it about things you like. If you want movie recommendations, first you have to pick a couple of genres — comedy, drama, western — and then choose some of your favorites from a curated set of titles. You can’t access the rest of the app until you’ve picked at least 20. “The payoff is huge,” says Salim Hemdani, Likewise’s CTO. “The more you tell us, the better it’s going to be.” He says people never stop at 20 because it’s just fun to pick things you like. And in doing so, you tell Likewise’s algorithm who you actually are.

In the recommendations world, you are who you cluster with

Likewise uses that information to put you into a “cluster,” which refers to a group of people with similar tastes to yours. These clusters are constantly changing based on what else you watch and rate, and they inform everything else Likewise recommends to you. “It gives us an initiation point to say, how many people are like you in the world, and how many clusters can we create?” Hemdani says. The more granular and specific those clusters are, the more accurate they can be. Knowing you like Succession is slightly useful; knowing you like Succession, novels by Michael Crichton, the podcast The Adventure Zone, and anything with Marvel in the title is vastly more useful. 

The simplest and most pervasive recommendation system, on Likewise and elsewhere, is known as collaborative filtering. It works by assuming that if you like something, and someone else likes that thing and also a second thing, you’ll probably like the second thing too. That’s it! It typically involves more data and more people, but that’s the core idea: if you like Severance and other people who liked Severance are really digging The Old Man, you probably will, too.

One of Morris’ theories is that Likewise can provide better recommendations, not just by knowing users better, but simply by having more things to offer them. Netflix, HBO, and Disney will never recommend each other’s catalogs, but Likewise (along with apps like Justwatch and Reelgood) can index them all. “We’re not aware of any recommendation engine out there who’s looking at things like the social graph or looking across books, podcasts, TV shows, movies,” Morris says, “and letting your preferences and other things influence each other across those categories.”

The simplest way to get better recommendations, nearly everyone in this space told me, is to give the apps and platforms more to work with. Multiple executives described the ideal personalization process as a collaborative exercise in which you and the AI work together to paint an accurate picture of what you actually like. Everything you thumbs-up on Netflix helps the app put you into the right clusters; every filter you tick on Yelp makes the restaurant recommendations more useful. Downvotes and dislikes are just as useful. Clicks, likes, and even engagement can mean a lot of things, but an explicit endorsement sends a much stronger signal.

Screenshot of a Pinterest search for “summer hairstyles” showing filter options for protective, coily, curly, wavy, straight, and shave/ bald. Protective is selected, showing various pins of people with braids and twists.
Pinterest has embraced personalization as a collaborative process with users.
Image: Pinterest

Strangely, though, many platforms have gone the other way, opting to infer what you like based on what you click or linger on as you scroll or engage with in some way. It’s based on a desire for a totally frictionless user experience, but from Facebook to YouTube to TikTok, we’ve seen what that can lead to: misinformation, rabbit holes, echo chambers, problems of all kinds. It also requires collecting astonishing amounts of data, grabbing every possible bit of information about you and your habits in case some of it is useful.

Naveen Gavini, the SVP of product at Pinterest, says he understands the impulse toward frictionless-ness. “If you opened up your favorite streaming content platform and you were gonna watch a movie,” he says, “I don’t think you want to first answer a 30-question quiz: Hey, what are all your favorite movies? Okay, how would you rate them? Who are your favorite actors? I don’t think anyone wants to go through that work.” Instead, he says, the key is to find just the right moments to ask questions. “I have a barber that I’ve been going to for 10 years that cuts my hair,” Gavini says by way of example. “And if you think about that experience every time, it’s a personalized experience, and I don’t need to tell him when I walk in how I want my haircut because he knows me. But it started with that first conversation: It was an explicit conversation, like, ‘Hey, so how do you generally like your hair cut?’” Making that same kind of dialog explicit, without overusing it, is a key goal for Pinterest.

Guessing what you like based on your actions is much harder than just... asking you what you like

One side effect of that collaborative process is that it can also offer users more transparency about what they’re being recommended and why. Nearly everyone I spoke to for this story said that’s important both in helping people have good experiences online and in engendering trust in the stuff that’s being recommended. “More and more,” Gavini says, “I think we want to know: What are the decisions? What are the things that are informing some of these algorithms that are actually delivering content to us?”

Trust is everything, really. There’s a hypothetical version of the Yelp app — and the Netflix app, Spotify app, Kindle app, and dozens of others — that is nothing more than a big button. You sit down to watch something, smash the button, and Netflix knows exactly what you’re looking for. Spotify puts on exactly the right song. Yelp orders the exact dish you’re craving. Everything is personalized and automated and delivers the One True Recommendation every time. But would you believe it enough to just hit the button? Akhil Ramesh, the head of consumer product at Yelp, doesn’t think so. “I often joke that if God landed in front of me and said, ‘This is the person you’re going to marry, and you’ll never have to waste a second,’ I wouldn’t believe a second of it,” he says. “I would go do my exploration.”

The One True Recommendation isn’t just impossible — it’s not even really worth pursuing. But that doesn’t mean things can’t get better. As the services we use get better at knowing us — and, just as important, get better at asking us about ourselves — they might be able to narrow the world down to a handful of options instead of an endlessly scrolling list. All you’ll have to do is pick your favorite and go. Because, really, there is no right answer. There’s just the one you picked.