Trying to sift through YouTube on your own would be a monumental task. With more than 100 hours of video uploaded every minute, the amount of new content that’s published in a single week is longer than your entire lifetime.
That’s where discovery comes into play. To help viewers find videos that are relevant and interesting to them, YouTube (and practically every video, set-top box, and TV company) is investing heavily in video personalization. The idea is twofold: A personalized video experience creates a more engaged viewer, who will, in turn, sit through more ads and help the platform’s bottom line.
On Sunday, the Emmy Awards will honor the best TV has to offer, but the National Academy of Television Arts & Sciences hasn’t forgotten about Silicon Valley, recognizing YouTube, Amazon, Adobe, Netflix, and TiVo for their work in video personalization. While these companies landed Emmy statues–to be presented January at the International Consumer Electronics Show–for their advancements in video technology, overall the industry recognizes there’s no clear winner in the race to build the smartest video recommendation engine.
Interviews with executives at pay TV, streaming, and hardware companies show the industry recognizes the importance of video personalization to capture more eyeballs, but still views it largely as a work in progress. Satellite company Dish called its system “an ongoing journey,” while TiVo says it’ll continue experimenting with ways to incorporate social data. And despite pouring his heart and soul into YouTube’s recommendation engine, engineering director Cristos Goodrow says it, too, has “far to go.”
The most glaring problem of video recommendation systems is that they’re too simplistic, deducing viewers’ tastes from clips viewed here and there. The suggestions they provide are obvious, but the goal is to give viewers nuanced recommendations they would have trouble discovering on their own. “I don’t perceive the system yet understands what my true interests are,” Goodrow laments.
Take, for example, videos of the Olympics. Every four years, interest is drummed up because it’s timely, but most people aren’t interested in pole vaulting during the off years. “If I happen to watch a lot of videos about a certain topic because it’s timely, that tends to confuse the recommendation system to think I’m going to watch a lot of that stuff in the future,” Goodrow points out.
Key to luring more viewers is an engine that almost intuitively knows what people want to watch–even before they know it.
“We want to understand when people form an attachment to a show,” Goodrow says. “The next episode is available to watch, and we can start recommending it right away.”
YouTube used to measure its success by counting video clicks, but the focus now is on increasing total view time. The site now streams 6 billion hours a month–an hour for every person on Earth–up from 4 billion within the last year. Google doesn’t break out advertising revenue for YouTube, but the video site consistently tops video rankings in the U.S., with 167 million watching 17.4 billion videos in August, the most recent data available from comScore.
Goodrow approaches video personalization like the mathematician he is. “I could say the whole system is based on statistics, or to make it simpler, it’s really based on accounting.” Much of the personalization algorithm is based on the habits of people who previously viewed a given video. Keeping track of where they jump to, the system is able to derive what future viewers are likely interested in.
In addition to personalizing the experience based on other viewers with similar watching habits, another core component of recommendation is regionalization. It’s no surprise people in India like to watch Bollywood movies or that soccer is far more popular in Europe than in North America. Yet part of online video’s appeal is that it transcends geography, so it was important for YouTube not to specify rules that say people in the United States aren’t interested in Bollywood, soccer, or even Psy’s blockbuster music video, Gangnam Style.
“Who would’ve thought a Korean pop song would be a worldwide sensation? If we had a system that said people in the U.S. never want to watch anything that’s Korean or Kpop, then we never would’ve shown people that Psy video,” Goodrow said.
So did YouTube’s personalization lead to Psy’s rise in fame? “It’s hard to say,” Goodrow hedges. “I would say the recommendation system enabled it in the same way that we noticed it was becoming very popular. That gave us more confidence to put it on the home page or show it to more people. Maybe it accelerated its growth, but it’s hard to say whether it played a significant difference in the end.”
Recognizing the limitations of its algorithm, YouTube has a back-up plan: subscriptions. In YouTube’s quest to transform television as we know it, it has been pushing channels–an idea from traditional TV applied to the Web. In 2011, the streaming site began doling out a reported $100 million in advertising advances to select content creators to produce studio-quality videos for their channels. To keep viewers coming back week after week, new episodes were released on a regular schedule as they are on television–quaint, given the on-demand world that streaming services have created. Goodrow considers it a success when a viewer subscribes to a channel. “If we can do that, it takes some of the burden off the recommendation system, of trying to observe a pattern in your watching,” he said about the algorithm, which has to carefully straddle the fence between being helpful and being overbearing.
“A subscription allows the viewer to say, ‘Okay, you don’t have to guess anymore. I do want to see more.'”
Of course, the personalization continues even after one subscribes. That expressed interest in a content creator, show, or channel will inform the recommendation engine, which continues to analyze patterns and suggest new videos.
Another focus for YouTube as it moves forward is standardizing its experience across platforms. Online video reports have found different sized screens to be more conducive for viewing long- or short-form video. Generally, people prefer to watch videos measuring 10 minutes or longer on larger screens, such as televisions and tablets, and shorter videos measuring less than three minutes on smartphones and desktops. These viewing correlations have given rise to the notion that videos of different lengths should be recommended based on viewers’ screen sizes. However, YouTube’s own experiments have led it to a different conclusion: Goodrow says viewers prefer to have consistent experiences across all platforms.
“It’s remarkable to me how long people watch video on devices that are pretty small, like phones,” he said, citing his daughter, who watches full episodes of iCarly on her phone. “I guess I wouldn’t do that, but somebody is.”
In fact, YouTube is looking to make its apps across phones, tablets, and set-top boxes more uniform. “In the past, in order to move fast, we had developed the different apps somewhat separately,” he said. But that’s created a “crummy experience” for the viewer who might be frustrated that a desktop feature isn’t available on the phone. “What we’re doing this year is making sure that as we develop new features or as viewers make connections to these things, they’re consistent throughout the platform.”
If there’s a trend in video, chances are YouTube has tested it: live, long-form, studio-quality, social, pay-per-view, subscription. . .
“Maybe I don’t know the difference between wacky and non-wacky,” Goodrow said. “We’re always experimenting with a bunch of things.” For example, YouTube integration with Google Hangouts has let viewers try group viewing–another concept borrowed from the traditional living room that is now being facilitated by the Web.
“Doing more exploration in ways that aren’t overly burdensome to viewers is where we need to go. We haven’t found balance for how to do that yet.”