YouTube has published an outline of how it recommends videos. It helps creators establish an audience and connect to more viewers with their videos.
Presented by Rachel Alves, a YouTube product manager, the overview shares insights on how the recommendation systems works.
The presentation is often seen at conferences where Alves joins in. and it helps people appreciate the systems that affect video distribution in the background.
“You don’t need to be an expert in algorithms of analytics to be successful on YouTube,” says Alves.
The aim is for users to stay longer and keep returning. They do it by ensuring the users have a good time on the platform.
“So we really maximize long-term satisfaction, so viewers keep coming back to YouTube,” adds Alves.
“If we go back to 2011, what we optimized for was clicks and views […] but that’s not that great of a metric, because it may indirectly incentivize clickbait-y or sensational titles or thumbnails that get people in to watch a video, but doesn’t make them very satisfied or happy.”
Alves says the most feedback YouTube had initially for its feed was that users’ home feeds got flooded with “sensational or off-putting videos.” They fixed this by switching to watch time as a key metric in 2012.
“How much time somebody spends watching a video or channel is much more indicative of the quality of the content, because if you spend more time watching something, it’s more likely that you’re going to be interested in it,” says Alves.
Alves says watch time also has a downside. You may watch videos for hours, but you may not feel good about how your time was spent.
YouTube has since looked for ways to better define “quality or value” watch time to optimize more towards satisfying users.
Here’s what they did:
- Send millions of user surveys every month to find out and optimize for what people like and enjoy.
- Focus on convincing content from established outlets for news content
- Stop the spread of “borderline” violative content
The highlight is on user satisfaction and to ensure people feel good. But YouTube must also maintain accountability on what gets amplified through its recommendations.
Alves says they keep mum from creators on this information for now. They often have limited feedback on each video to offer useful feedback. But the info can inform better on their algorithms.
“We are looking at adding more satisfaction data, and externalizing it to creators, so it is something we’re working on,” says Alves.
YouTube also uses signals like:
- Clicking/tapping ‘Not interested’ in the video menu
- Likes and dislikes
Alves notes the Homepage and ‘Suggested’ listings use different algorithms. The idea of having one central YouTube algorithm is incorrect.
“The Home page offers up a broad array of videos when you visit youtube.com, and it uses similar signals as ‘Suggested’, but they are designed to do slightly different things,” says Alves.
She notes that creators look for ways to know how they can optimize. But she says they can’t.
“You can’t optimize for a traffic source, you can only optimize for people or viewers.,” explains Alves.
You can see more of the overview here.