Future of TV Briefing: YouTube gives a peek at how its recommendation algorithms work
This Future of TV Briefing covers the latest in streaming and TV for Digiday+ members and is distributed over email every Wednesday at 10 a.m. ET. More from the series →
This week’s Future of TV Briefing recaps a VidCon session during which YouTube executives and creators unpacked the video platform’s recommendation engine.
The ‘You’ in YouTube
A session titled “YouTube Decoded” featuring YouTube employees is usually a dead giveaway that the inner-workings of the Google-owned video platform’s algorithm will not, in fact, be decoded. Turns out, that was a misperception by me.
OK, YouTube executives did not project the platform’s source code on screen during the session, which was hosted during the annual Comic-Con for the creator economy, VidCon, in Anaheim, Calif. last week. But they did shine some light for creators and publishers on how YouTube decides which videos to put in front of audiences and address some myths and misperceptions surrounding its algorithm – or, I should say, algorithms.
“There’s actually kind of dozens of different algorithms that are working together to try to find videos in different ways and then bring them all together into one ranked list for [an individual viewer,]” said Todd Beaupré, senior director of growth and discovery at YouTube.
Two important things there. First, yes, there are multiple YouTube algorithms and corresponding recommendation systems, which we’ll get into in a second. But second – and more importantly – YouTube’s algorithm isn’t oriented around the video, let alone its creator, but the viewer.
“We’re not posting a video, and then YouTube is trying to find viewers for our video. No, it starts when a viewer goes on YouTube, and now they’re trying to find the videos for that viewer. So because of that, when I’m creating my videos, I think about a specific person [who would be the target audience for that video],” said Jenny Hoyos, a YouTube creator with more than 9 million subscribers on the platform.
Now back to algorithms — plural. Those algorithms take into account what videos a person is most likely to watch – duh – but also to what extent that person tends to be more interested in videos from a regular set of channels versus open to videos from channels they haven’t come across before. Time of day also factors in, so does the viewing behavior of people who watched the same video as a given person, such as what other videos those people watched that this person might also be interested in.
YouTube also takes video formats into account. Like if a person is more likely to watch a YouTube Short when on their phone or tune into a longer video when watching on a TV screen.
In fact, the platform has teams that specialize in specific formats.
“There’s a Short-specific recommendations team. There’s a live-specific recommendations team. There’s teams that are starting to work on more shopping content. And what we recognize in working in these different formats is that the signals that audiences give us to help us understand the value that they’re getting are different for the different formats,” Beaupré said.
Beaupre also dispelled some myths and misperceptions regarding YouTube’s algorithms. For example, some creators have come to believe that any engagement with their videos — including dislikes and negative comments — will have a positive effect when it comes to YouTube choosing to push a video out to more people. Nope.
“That’s not true. We’ve seen that dislikes are a signal that tell us that somebody is not having a great experience, and when we experiment with a recommendation system to show less content that a user is likely to dislike, then they have a better experience long-term on the platform,” Beaupré said.
That’s not to say that dislikes will lead to a video not being recommended to anyone. A video disliked by one person may be liked by another, and the latter like makes the video more likely to be recommended to people who share viewership habits with the second person.
“We try to identify the people who are likely to dislike it and different from the people who are likely to like it. Most videos have something like a 90% [like-to-dislike] rate, which means 9 out of 10 people that rated it liked it,” Beaupré said.
Which underscores the primary takeaway from the session and the most important thing to know about how YouTube’s algorithms work: YouTube makes its recommendation decisions at the level of the individual viewer, not the video.
Is that a pretty basic takeaway? Sure. But know what’s another term for basic? Fundamental.
What we’ve heard
“By traditional TV, do you mean like cable?”
— Lucy, a 17-year-old VidCon attendee, when asked if she watches any traditional TV
Let’s talk CTV in NYC
Are you tired of talking about CTV advertising? No? Me neither. Which is why I’m psyched to bring together a bunch of my favorite buy-side executives to talk about the state of the CTV ad market next month in New York City for the Digiday CTV Advertising Strategies event on July 15 (I’m less stoked about leaving southern California in mid-July for hot and humid NYC, but that’s another thing).
Anyway, come hang out and hear how top executives from brands and agencies — including Amica Insurance, Danone, Horizon Media and UM Worldwide — are navigating the opportunities and challenges in today’s CTV ad market and what they say it will take for streaming to overtake the TV once and for all. Plus we’ll be hosting a behind-closed-doors, no-holds-barred town hall session, which is always my favorite part of Digiday events. Click here to attend.
Numbers to know
200 billion: Average number of views that YouTube Shorts receive on YouTube each day (YouTube counts a view as soon as a video appears on screen).
9%: Percentage share of TV ads that included closed captions in 2024.
500+: Number of episodes that Tubi will syndicate collectively from six YouTube creators.
-10%: Percentage by which Google cut the budget for its Android TV and Google TV team.
6: Number of minutes’ worth of ads that Max serves per hour of programming, up from 4 minutes in February.
What we’ve covered
How TikTok’s ‘The Secret Lives of Mormon Wives’ landed on Hulu:
- Select Management Group’s Danielle Pistotnik joined the Digiday Podcast to go behind the scenes of the show’s development and eventual sale to Hulu.
- When Pistotnik initially pitched the creator-led series to traditional entertainment companies, no willing buyers emerged.
Listen to the latest Digiday Podcast episode here.
Why there’s a CPM slump in a growing CTV market:
- Programmatic ad prices for CTV inventory is down by 10% to 30% this year compared to 2024 rates.
- Streaming ad prices had come down last year as the amount of available streaming ad inventory swelled.
Read more about CTV ad prices here.
Spotify’s video podcast program draws praise from creators — and skepticism from networks:
- Spotify’s Partner Program doesn’t allow for dynamic ads to be served to the platform’s premium subscribers.
- As a result, some podcasters are opting not to join the program.
Read more about Spotify here.
Creators turn to agentic AI to manage fan engagement:
- Creators are using AI agents to respond to comments and DMs from fans.
- Some creators are drawing a line when it comes to using AI agents to engage with fans in connection to sponsored posts.
Read more about creators here.
Snap’s AI play targets the advertisers tired of Meta and Google:
- Last month Snapchat launched an AI-powered suite of tools called Smart Campaign Solution.
- The tools are akin to Google’s Performance Max and Meta’s Advantage+.
Read more about Snap here.
Omnicom wraps up its Cannes Lions presence with a YouTube livestream partnership:
- Through the deal, Omnicom’s clients can target ads against YouTube livestreams.
- Previously YouTube packaged its live inventory as part of broader media buys.
Read more about Omnicom’s YouTube deal here.
What we’re reading
Netflix’s move into traditional TV:
The streaming service will distribute traditional TV channels from France’s TF1 starting next year, according to Variety.
Disney’s lawsuit against generative AI company Midjourney — which just released an AI video generation tool — over copyright violation is part of the entertainment industry’s efforts to put guardrails around intellectual property and draw lines around fair use, according to Bloomberg.
YouTube creators weren’t aware that Google was using videos on the platform to train its AI models, and YouTube has not given creators a way to opt out of their videos being used for AI training, according to CNBC.
YouTube’s AI-generated Shorts:
YouTube will pull Google’s video-generating AI tool Veo 3 into its short-form platform this summer, according to The Hollywood Reporter.
For the third time this year, President Donald Trump has extended the deadline for ByteDance to spin off TikTok or have the platform be banned in the U.S., according to The Wall Street Journal.
