How Does the YouTube Algorithm Work?

YouTube has said that recommendations drive a significant amount of overall viewing, more than subscriptions or search. So "the algorithm" decides much of what billions of people watch, and whether a creator's video takes off.

YouTube has explained a lot of it publicly: a 2021 post by its VP of Engineering, Cristos Goodrow, and a well-known 2016 research paper on its deep learning system. This guide combines both.

🎯 Quick Answer (30-Second Read)

  • Two stages: first narrow millions of videos to a few hundred candidates, then rank those for you.
  • It optimises for satisfaction, not clicks. Clicks alone were misleading (2011), so YouTube added watch time (2012), then surveys to measure "valued watch time."
  • Signals: what you watch, how long, likes, dislikes, shares, "Not interested," and survey responses. YouTube says it uses 80 billion+ signals.
  • No single formula: Home, Up Next, Shorts and Search each rank differently.
  • Borderline content is demoted in recommendations.

The Two-Stage System

YouTube's 2016 paper, "Deep Neural Networks for YouTube Recommendations," describes the design used by most large recommenders today:

flowchart LR ALL[("Millions of
videos")] --> CG["Stage 1: Candidate generation
your history → a few hundred
videos you might like"] OTHER["Other sources
subscriptions, trending,
related videos"] --> CG CG --> RK["Stage 2: Ranking
score each candidate with
many features → predicted
satisfaction"] RK --> MIX["Final list
+ diversity, freshness,
policy rules"] MIX --> HOME["📱 Your Home feed"] HOME -.->|"what you watch, skip,
like or rate"| CG
Candidate generation
Fast and broad. Uses your watch and search history to pick a few hundred plausible videos out of millions.
Ranking
Slower and precise. Scores each candidate using rich information about the video, the channel and you, then orders them.

Why two stages? Scoring every video carefully for every user would be far too slow. Narrowing first makes the expensive scoring affordable.

What YouTube Optimises For: A Short History

timeline title What YouTube's recommendations aimed for 2011 : Clicks : Misleading, people clicked but didn't watch 2012 : Watch time : Rewarded videos people actually watched Later : Satisfaction surveys : Valued watch time 2019 : Borderline content demoted

Valued watch time means time spent on videos you'd rate highly. YouTube asks some viewers to rate videos from one to five stars, and only four- and five-star videos count as valued. Since most people never answer surveys, a machine learning model predicts how you'd rate a video based on everyone's answers.

The Signals

Signal What it tells YouTube
Clicks Interest in the title and thumbnail
Watch time Whether the video delivered
Survey responses Whether you valued it
Likes, dislikes, shares Satisfaction, weighted differently per person
"Not interested" / "Don't recommend channel" Strong negative feedback
Your history What you've watched and searched

Borderline Content

Some videos don't break the rules but come close, such as misleading health claims. YouTube uses human evaluators to help classify videos as authoritative or borderline, and demotes borderline videos in recommendations. After it started doing this in 2019, YouTube reported a 70% drop in watch time of such recommended content from non-subscribed channels in the US. Its stated goal is keeping borderline views from recommendations below 0.5% of all views.

What This Means for Creators

Key takeaway: the system tries to predict whether a viewer will be glad they watched. Tricks that earn clicks but disappoint viewers work against you, because watch time, dislikes and survey predictions catch up.
  • Make the title and thumbnail honest and compelling: clicks still start everything
  • Deliver on the promise early, so viewers keep watching
  • Make videos people would rate highly, not just finish
  • Build series and playlists so one good video leads to the next
  • Look at returning viewers and audience retention in YouTube Studio

Frequently Asked Questions

Does the YouTube algorithm favour long videos?

Not directly. YouTube optimises for viewer satisfaction, using watch time, surveys and other signals. A long video people abandon early performs worse than a short one they finish and rate highly. Length helps only when viewers genuinely value the extra time.

How many signals does YouTube use?

YouTube's VP of Engineering wrote in 2021 that the recommendation system learns from more than 80 billion signals, including clicks, watch time, survey responses, shares, likes and dislikes, each weighted differently depending on the viewer.

How do I reset my YouTube recommendations?

You can pause or clear your watch history, remove individual videos from history, and use "Not interested" or "Don't recommend channel" on videos in your feed. Clearing history resets much of what the system uses to personalise Home.

Conclusion

YouTube's algorithm is a two-stage recommender: generate a few hundred candidates from millions, then rank them by predicted satisfaction. Over time, YouTube moved from clicks to watch time to "valued watch time" measured with surveys. For viewers and creators alike, the system rewards videos people are glad they watched.

Sources

Last verified: 14 October 2026.

Related reads: How the Instagram Algorithm Works · Vector Databases Explained · Why Free Apps Cost Billions to Run

I explain the tech behind everyday products. Follow on LinkedIn or X, or invite me to speak.