The Secret to Creating Powerful Videos

The YouTube algorithm is often viewed as a mysterious black box that determines which videos become famous and which fade into obscurity. But according to Todd Beaupré, who manages YouTube's recommendation systems, the algorithm is less about pushing certain videos and more about connecting viewers with content they like. In a fireside chat at VidSummit 2024 last week, Beaupré shed light on how YouTube's recommendation engine works, offering valuable insight for creators and viewers alike.

The viewer-centered approach

Contrary to popular belief, the YouTube algorithm is not activated when a creator uploads a video. Beaupré explains, “What actually triggers the algorithm on the homepage, in search, or in next suggestions is when an individual viewer visits YouTube.” This viewer-centric approach aims to find the best content for each user in that moment.

“We'll use what we know about you,” Beaupré continues. “We'll understand your past watch history, your subscriptions, your likes, and we'll try to select some videos.”

ForbesTom Bilyeu calls on developers to adapt to the changing AI landscapeFrom Daniela Schröder

Three pillars of recommendation

Beaupré outlined three main methods YouTube uses to get video recommendations:

1. Goal-based recommendations: “The first one is more like what you could call goal-based. If you think about it, if you were the algorithm and you were in charge of getting videos for a user, how would you go about finding videos to recommend?”

2. Collaborative filtering: Beaupré describes this as “automated word of mouth.” He explains, “Word of mouth is when someone tells someone else about something they like, right? Like, Oh, I just saw this movie. Have you seen it yet? And collaborative filtering automates that.”

3. Deep learning algorithms: “The third part is much more sophisticated and uses the most advanced machine learning algorithms that we use in deep learning,” says Beaupré. “The best way to describe it is that we can take everything we know about a viewer and boil it down to the essentials, so to speak. Almost like finding out DNA, so what kind of latent interests are there?”

ForbesHow the AI ​​revolution is reshaping the creator economyFrom Daniela Schröder

Performance vs. Personalization

While creators often focus on metrics like click-through rate (CTR) and average view time, Beaupré emphasizes a more nuanced approach: “There are two main components. One component that I think creators particularly focus on in their analytics is performance… But what makes the system work really well is recognizing that different viewers have different preferences.”

He adds: “Every time someone loads a homepage, the system tries to predict how likely that viewer is to click on each video. And for each viewer, a different click-through rate and a different average watch time are predicted.”

The multilingual future

YouTube is increasingly focusing on multi-language support. Beaupré advises creators looking to expand internationally: “We've found that there's a tipping point at about 80% of watch time. So if you look at your catalog, sort it by popularity and pick the top 80%, then it seems like it really allows viewers to engage with it.”

Debunking common myths

1. The “YouTube Paradox”: Beaupré explains, “This phenomenon is what I sometimes call the YouTube Paradox or the Metrics Paradox. It's where you think, 'You know, this is my best video, it only has a 3% click-through rate, and my latest video has a 10% click-through rate.' But YouTube doesn't show you that much of it.”

2. Traffic from unrelated videos: “Suggested videos are not just based on related videos. They are not called related videos,” Beaupré clarifies. “We see that people who watch the video sometimes want to see a similar video, but we think that too, and we see in the data that people sometimes want to switch channels.”

3. The impact of external views: Regarding concerns about low-quality external views, Beaupré reassures: “We also know that different contexts, such as the home page, the next pages or search, give different results. So on the home page in general, for example, we will find out what happens if we show the video on the home page to a certain type of viewer.”

ForbesFrom rapper to pioneer of the creator economy: Iggy Azalea's bold new projectFrom Daniela Schröder

The role of AI and large language models

YouTube uses AI technologies similar to those used in advanced chatbots. Beaupré explains: “Large language models are the kind of machine learning innovations you see in a lot of chatbots like Gemini, Chat GPT, etc. And we found that we can adapt these models for recommendations as well.”

He compares this progress to the difference between a novice cook and an experienced chef: “This new technology allows the chef to really learn how to cook instead of just memorizing recipes.”

Advice for developers

Beaupré's key advice for creatives is to focus on the audience: “Focus on your audience and understand the value you are creating for them, because that's what we're ultimately going to be listening to.”

He also points to the growing importance of longer content, especially on TV screens: “We're seeing a big increase in longer videos, especially on TV. So you should think about how your content looks on TV, but also how you design your content.”

Beaupré suggests that creators think in terms of shows or series: “Think about packaging your content similarly to shows and formats that are recognizable to people and making it more accessible.”

The future of YouTube

Even as YouTube evolves, the goal remains the same. As Beaupré puts it, “Ultimately, it comes down to the viewer wanting to see the video.” The platform is refining its ability to understand content nuances and viewer preferences, always with the goal of creating meaningful connections between creators and their audiences.

Understanding these principles can lead to a more rewarding YouTube experience for both creators and viewers. Beaupré concludes, “We all have an incentive to get videos in front of the right viewers who want to see them.” Ultimately, YouTube's algorithm is less about gaming the system and more about enabling each individual viewer to discover valuable content.