I’m always on the lookout for off-the-beaten-path Twitch streams, and today, a pal recommended a great one: a mechanic who streams himself fixing cars on Twitch. (The audio is surprisingly clear, too.) I watched for a while today as he replaced the transmission on a Kia, which had its previous transmission destroyed after something put a hole in it on the highway. It’s the kind of thing I’m always trying to find on Twitch: a keyhole look at someone else’s life. I love watching people doing things I’ll never be able to do.
Last summer, I wrote a piece about how I was having a hard time discovering things on Twitch; it was hard, I said, because trawling through the sheer number of streamers on Twitch was like trying to surf TV channels without any kind of guide. Since then, it feels suspiciously like things have gotten better. There’s a new way to organize streams in the browse menu, for example, and now you can sort by “recommended for you” instead of just organizing categories from most viewers to least. The left rail has also gotten more useful. It now seems to recommend streamers who are currently live that are similar to the people I already follow and watch regularly.
That, as it turns out, is the result of a lot of hard work. The other day, I spoke to Tom Verrilli, a product manager and the head of Twitch’s viewer experience team. It’s his job to figure out how to connect viewers with streamers; he’s the guy who’s figuring out how to get streamers found, and to give the people watching an easier time finding them. The problems, he says, are threefold: first, Twitch has an infinity of content. Second, all of that content — all of those people — are live. Third, channels are people; they’re not interchangeable units. “Twitch has 1,000 times more streamers than any video platform has kind of movies, TV shows, individual pieces of content,” he says. “So tons of people have experienced the ‘spend 30 minutes trying to find a movie to watch’ problem of streaming services. We have that on steroids.” Because the internet, as he points out, is an “on-demand” platform, it runs mostly on pre-recorded, instantly available content. But if you’re live, the problem is magnified.
Verrilli gave me an example to put things into perspective. “If you are one of those hardworking streamers who stream eight hours a day, seven days a week, you’re only live 33 percent of the time that someone’s up on the platform and they come looking for you,” he says. “A really easy comparison point to this would be like, what would a search engine look like if two of the three times I searched for The Verge you weren’t there? All we could say is, yes, The Verge is a thing that exists, but you can’t read any of the articles on it. Try again later.” That is very different than the way most of us use the internet — like when we’re bored or when we have a couple minutes of downtime, it’s easy to cue up something like a YouTube video and just take a moment to watch it. Not so with stuff that’s live.
“There is this kind of beautiful serendipity that happens when you stumble upon just the right channel and the right community for you at the right moment when you were ready for it,” Verrilli says. “My team’s job is to try and mechanically create serendipity in the order of hundreds of millions.” They started from scratch only a couple of years ago, he says, but now it’s starting to be fruitful. The percentage of videos watched because of their recommendations is up something like 700 percent year over year, he says. “We’re starting to prove out value to both streamers and viewers that we can help them find each other at the right time. But we got a long way to go.”
To make recommendations to viewers like you, Twitch uses a form of machine learning that lets the machine work out for itself what viewers are interested in. Verrilli’s team points the system toward certain “features” of streams, like how chatty the audience is, and the AI determines how important it is to viewers. Take, for example, the feature chat velocity. “How often are people chatting in the channel that you’re watching? We don’t then go through and tell the models how much they should consider chat velocity for any one person,” he says. “But the model understand that some people are watching very chatty channels and like it; some are not. And then classifies channels as chatty or not chatty, and may use that as one of the many inputs to determine if that’s the right recommendation to make when you arrive.” (Another thing they had to help the models understand was the concept of time because recommending a stream that’s nearly done or one that’s never online at the time of a recommendation is not very helpful.)
This can be scary. On most platforms — Twitch included — recommendations drive growth and can be the thing that makes one channel blow up over another. As far as Twitch goes, however, Verrilli assures me that the new recommendation system is more equitable than the old one. “Conceptually, historically, we have had the least equitable form of discovery, which has always been ranked by big to small. And that means that it’s fantastic if you are kind of top tier talent, and increasingly hard for folks who aren’t those,” he says. “Volume disproportionately affects the knowledge.” In the last two years, his team has been working, though, Verrilli says that growth has gone disproportionately to Twitch’s smaller communities. The big ones are still growing, he says, “but recommendations afford us the capacity to kind of make sure there is more equitable outcomes for everyone.”