How YouTube’s Recommendation System Works In 2025

In a recent video interviewYouTube Connection René Ritchie spoke with Todd Beaupré, YouTube’s senior director of growth and discovery, to discuss the platform’s recommendation system features and what creators can expect this year.

Their discussion revealed how time of day, device type, viewer satisfaction, and the emergence of large-scale language models (LLMS) are reshaping YouTube’s algorithms.

Here’s what you need to know about YouTube’s referral system and how it works.

Personalized recommendations

One of the central themes of the interview is YouTube’s focus on tailoring content to individual viewer preferences.

According to Beaupré:

“A lot of times creators will say hey, uh, the referral system is promoting my video to people or why isn’t it promoting my video yeah, they can ask that, and the way the business works … it’s not so much … it’s not so much about putting it out there as much as it recedes … “

He goes on to explain that YouTube’s home feed prioritizes content based on what each viewer is most likely to enjoy at any given time:

“When you open the homepage, YouTube will say Hey Rene here, we need to give Rene the best content to make Rene happy today.”

Metrics and satisfaction

While click-through rate (CTR) and watch time remain important, YouTube’s system also counts user satisfaction through direct surveys and other feedback signals.

Beaupré Notes:

“We introduced this concept of satisfaction … we’re trying to understand not only about the behavior of viewers and what they’re doing, but how they feel about the time they’re spending.”

He explains that YouTube’s goal is to foster long-term viewer satisfaction:

“… We’re looking at things like likes, dislikes, these survey responses … We have a variety of different signals to tap into this satisfaction … We want to build a relationship with our audience like creators want to do with their fans.”

Evergreen and trendy content

YouTube’s algorithms can identify older videos that become relevant due to trending topics, viral moments or nostalgic interests.

Beaupré states the system’s ability to turn:

“… Maybe there’s a video now that reaches a certain audience, but like six months from now … that makes this video relevant again … if it’s relevant and maybe to a different audience than it was enjoyed the first time.”

Context: time of time, device and viewer

Beaupré found that YouTube’s system can show different types of content depending on whether someone is watching in the morning or at night, on a mobile phone or on a TV:

“The recommendation system uses time of day and device … as some of the signals we learn from to understand if there’s different content that appeals in those different contexts … if you have a tendency to watch news in the morning and comedy at night … We’ll try learn from other viewers like you if they have that pattern. “

Fluctuations in views

Creators often worry if their views fall, but Beaupré suggests that this can be a natural ebb and flow:

“… The first thing is that it’s natural … It’s not particularly reasonable to expect to always be at an all-time high view level … I would encourage you not to worry too much …”

He also recommends comparing metrics over longer periods and using tools like Google Trends:

“… We see that seasonality can play a role … it encourages you to look beyond … 90 days or more to see the full context.”

Multiple audio languages

Many creatives are exploring multilingual audio to expand their audience.

Beaupré points out how YouTube has adapted to support named records:

“… We had to add some new features … aware that this video is actually available in multiple languages ​​… so if you’re a creator interested in expanding your reach through dubs … make sure your titles and descriptions are too .. . [in] translated titles and descriptions … “

It also emphasizes consistency:

“We’ve seen creatives in particular that last at least 80%…view time…have more success than those that have less…”

LLM integration

Looking to the future, large-scale language models (LLMS) enable YouTube to better understand video content and viewer preferences.

Beaupré says:

“… we applied LLM technology to YouTube recommendations to … make them more relevant to viewers … not just remembering that this video tends to do well with this type of viewer … could actually understand the ingredients dishes better and maybe some more elements of the video style … “

Beaupré compares it to an expert chef who can adapt recipes:

“… We want to be more like an expert chef and less like … a memorized recipe.”

A key turning point for creatives

Here are the top takeaways from their 21-minute conversation on the YouTube referral system.

1. The recommendation system is about “pulling” content for each viewer, rather than pushing videos universally.
2. Metrics like CTR and Watch Time Fatter, but satisfaction (likes, dislikes, polled feedback) are also essential.
3. YouTube may resurface older videos if there is renewed interest.
4. Time of day and device impact recommendations.
5. Fluctuations in displays are normal – the season, trending events and external factors can be at play.
6. Home and translated titles can help reach new markets, especially if a high percentage of your content is available in the same language.
7. Large language models empower more nuanced understanding—creators should be attuned to how this affects discovery.

Watch the full interview below.

https://www.youtube.com/watch?v=dhyib72l1hu

YouTube plans to share more updates at Vidcon later this year.

---

Featured image: mamun_sheikh/Shutterstock