How YouTube decides what to recommend

A plain-language map of the recommendation system behind Home, “Up next” and Shorts, with every claim labelled by how sure we actually are.

Confirmed
Stated in YouTube's official docs or blog, or described in research papers published by YouTube/Google engineers.
Said by staff
Said by YouTube employees in interviews, Creator Insider videos or talks. Real, but informal and sometimes vague.
Educated guess
Inferred from creator analytics, outside research and how large recommender systems usually work. Plausible, not verified.

1. The big picture

YouTube doesn't pick from all ~billions of videos at once. Published Google papers describe a funnel: cheap models narrow the pool to a few hundred candidates, then a heavier model scores each one for you, right now, and a final pass adjusts the list for policy and variety. Confirmed

INPUTS THE FUNNEL OUTPUT About you watch & search history subscriptions, likes, “not interested”, surveys About the video who watches it, how long, clicks, likes, shares, title / topic / freshness Context device, time, location, what you're watching now Exact weights how much each signal counts: not public All eligible videos billions · minus removed / restricted content ① Candidate generation several fast “sources”: similar viewers, related to what you just watched, your subscriptions, trending, new uploads… → a few hundred candidates ② Ranking a deep model predicts, for each video: will you click? how long will you watch? will you like / share / dismiss it? will you be satisfied? → combined into one score, sorted ③ Final adjustments demote “borderline” content, boost authoritative sources on news/health, add variety so the feed isn't one topic Your screen Home feed Up next Shorts each tuned separately Feedback loop: what you do next becomes new training data confirmed guess / unknown not public

2. Stage by stage

① Candidate generation

Pulls a few hundred videos you might like out of billions, using several retrieval sources and learned “embeddings” that place similar users and videos close together. Confirmed (Covington et al., 2016)

Practical meaning: your video has to be “near” something a viewer already watched, or near people like them, to even enter the race.

② Ranking

A bigger model scores each candidate on several predicted outcomes at once: engagement (clicks, watch time) and satisfaction (likes, survey answers, dismissals). Confirmed (Zhao et al., 2019)

The same paper says it corrects for position bias: a video clicked because it was on top isn't automatically “better.” Confirmed

③ Final adjustments

Since 2019 YouTube reduces recommendations of “borderline” content and raises authoritative sources for topics like news and health. Confirmed

Some diversity / freshness mixing so a feed isn't 20 copies of one thing. Said by staff The exact rules are not public. Guess on details

3. Which signals matter

SignalWhat we knowConfidence
Watch timeIn 2012 YouTube switched from optimising mostly for clicks/views to watch time, to fight clickbait. Still a core ranking input.Confirmed
Satisfaction surveysUsers are asked to rate videos they watched; a model learns to predict satisfaction for everyone else (“valued watch time”).Confirmed
Likes, dislikes, shares, “Not interested”, “Don't recommend channel”Officially listed as signals. Dismissals are a strong negative.Confirmed
Watch & search history, subscriptionsListed as signals; you can pause or delete history and recommendations change.Confirmed
Click-through rate on thumbnail/titleClicks are a signal, but YouTube stresses that clicks without watching don't help. CTR alone isn't the goal.Said by staff
“The algorithm follows the audience”YouTube says it ranks viewer reactions, not videos in isolation; it doesn't “judge” a video's quality directly.Said by staff
Upload frequency / posting timeStaff say there's no direct penalty for uploading less often. Effects people see are likely indirect (audience habits).Said by staff
New videos get a “test audience” firstWidely observed: a new upload is shown to core fans/subscribers, then wider if they respond well. Consistent with how such systems work, but the mechanics are not documented.Educated guess
Relative weights of each signalHow much watch time vs. likes vs. surveys counts, and how that differs on Shorts vs. long-form. Not public and changes over time.Unknown
Shorts rankingSeparate surface with its own tuning. Swipe-away vs. watch-through (and loops) likely matter a lot there, but YouTube hasn't published the details.Educated guess
“Shadow-banning” channels / tags / keywords in descriptionsNo evidence of hidden channel bans in ranking; tags are officially described as minor. Mostly folklore.Guess (likely myth)

4. What this means in practice

For viewers

  • Your clicks and what you finish train your feed. “Not interested” is the strongest steering wheel you have.
  • Pausing watch history makes recommendations generic; deleting items removes their influence.

For creators

  • Optimise for the viewer being glad they clicked: honest packaging + a video that delivers on it.
  • Your real competition is the other ~hundreds of candidates for that specific viewer, not “the algorithm.”
  • Read retention and CTR together; one without the other misleads.

Caveat: the research papers describe the system at a point in time (2016, 2019). Production YouTube changes constantly and runs many models per surface. Treat the funnel as the reliable shape, and the specific numbers and weights as unknown.

Sources