How the YouTube algorithm works: I read the 2016 paper and what YouTube has said since

What I read: a 2016 engineering paper, and YouTube's own explainers since

My main engineering source is "Deep Neural Networks for YouTube Recommendations", by Paul Covington, Jay Adams and Emre Sargin of Google, presented at the RecSys conference in 2016. confirmed · 1 It describes the production system YouTube was running as of 2016, judged by live experiments on real users, with examples centered on the home page. confirmed · 1 It's ten years old and YouTube has changed since, so read it as the foundation of the design, not today's recipe. confirmed · 2, 7

Since then, YouTube has mostly explained itself in official posts and Help pages:

• 2021, YouTube Official Blog: "On YouTube's recommendation system" by Cristos Goodrow, VP of Engineering, on the signals and how YouTube's goals changed. confirmed · 2
• Undated, How YouTube Works: the official recommendations page, which covers channel quality and outside evaluators. confirmed · 3
• Undated, YouTube Help: "Search & discovery tips" (long-form and Shorts versions) and "Good to know about recommendations", which answer creator questions. confirmed · 4, 5, 6
• May 2025, YouTube Official Blog: "Shorts truths", which says Shorts and long videos have separate recommendation systems. confirmed · 7

These are YouTube's own words: specific and useful, but with no model details, and nobody outside can audit them. Me included.

The 2016 paper doesn't claim to have invented the two-step design; its authors call the split "classic". confirmed · 1 The same shape shows up in what other platforms have published since: X's 2023 open-source code (candidate sources, a neural ranker, then filters) and LinkedIn's 2025 feed retrieval paper, which narrows hundreds of millions of posts to a small set for ranking. confirmed · 9, 10

It's a funnel: millions of videos, a few hundred candidates, a few dozen shown

Why this matters: YouTube says recommendations, on the home page and in Up Next, drive more viewing than subscriptions or search. confirmed · 2

1Understanding the video. YouTube has published little here. The 2016 paper doesn't describe analyzing images or sound: its first model learns about videos from viewing patterns (which videos the same kinds of people watch), and its ranking model adds features such as the video's language. confirmed · 1 Today's Help pages say the system tracks the topics viewers watch and, for Shorts, trending songs or sampled audio. confirmed · 4, 5 They don't say how.

2Finding candidates. The first model reads a viewer's history and pulls a few hundred videos out of millions. confirmed · 1 In 2016 its inputs included the videos the person had watched, the words of their recent searches, and details like region, device, age, gender and logged-in status, which help recommendations behave sensibly for new users. confirmed · 1 It gives every viewer and every video a numeric profile, then looks up the videos whose profiles best match the viewer's, within tens of milliseconds. confirmed · 1 YouTube still puts it plainly: it compares your viewing habits with those of similar viewers. confirmed · 2

An example from the paper: if you'd just searched for Taylor Swift, a model predicting your next watch would copy that search page onto your home page, which the authors say "performs very poorly", so they hid the order of searches from it. confirmed · 1 The model also learned from all YouTube watches, including videos embedded on other sites, so videos people found elsewhere could spread to similar viewers. confirmed · 1

Freshness got its own fix, a feature called "example age". confirmed · 1 A model trained on several weeks of past viewing tends to predict a video's average popularity over those weeks, and misses the spike after upload. confirmed · 1 So in training, the team told the model how old each example was, and when recommending, set that age to zero, meaning "now". confirmed · 1 In live tests, this "increased the watch time dramatically on recently uploaded videos". confirmed · 1 The paper adds that users prefer fresh content, "though not at the expense of relevance". confirmed · 1 So it's neither a fixed boost for new uploads nor a countdown after which videos stop being shown. inferred · 1

3Ranking. A second, more detailed model scores each candidate, and videos are shown in order of score. confirmed · 1 Handling hundreds of videos instead of millions, it can use "hundreds of features". confirmed · 1 The paper names the most important as the viewer's past interactions with the video and similar ones, such as how many videos they've watched from this channel and when they last watched this topic. confirmed · 1 A video recommended but not watched gets pushed down on the next page load. confirmed · 1 Packaging counts too: the paper's own example is a viewer likely to watch a video in general but unlikely to click it "due to the choice of thumbnail image". confirmed · 1

In 2016, the ranking goal was generally based on expected watch time per impression, not the chance of a click. confirmed · 1 The paper's reason: "Ranking by click-through rate often promotes deceptive videos that the user does not complete ('clickbait') whereas watch time better captures engagement." confirmed · 1

In a later account, YouTube says it learned in 2011 that a click doesn't mean someone watched, added watch time in 2012, then looked beyond watch time to whether viewers felt it was well spent. confirmed · 2 It lists clicks, watch time, survey responses, sharing, likes and dislikes as signals. confirmed · 2 Surveys ask viewers to rate a video from one to five stars, and only four or five count as "valued watchtime". confirmed · 2 A model trained on the answers predicts everyone's likely response. confirmed · 2

4Final filters and quality. The 2016 paper stops at sorting by score. confirmed · 1 YouTube's later statements add a quality layer: since 2019 it has demoted "borderline" videos, which come close to breaking its Community Guidelines but don't. confirmed · 2 It says this cut U.S. watch time of such videos recommended to non-subscribers by 70%. confirmed · 2 For news and information, human evaluators rate videos against published guidelines, YouTube trains its systems on their judgments, and more authoritative videos are promoted. confirmed · 2 YouTube also weighs "the reputation and the quality of a channel" in deciding how, when and to whom content is surfaced. confirmed · 3 Whether this is a separate last step or part of ranking, none of these sources says.

5Learning from what happens next. In 2016, live A/B tests had the final say, and the ranking goal was "constantly being tuned" on their results. confirmed · 1 Today's Help pages describe the loop: when a video is recommended, do people choose to watch it, ignore it or click "Not interested"; do they stick around; and did they enjoy it, judged by likes and post-watch surveys. confirmed · 4, 5 YouTube names average view duration and average percentage viewed as ranking signals, and for Shorts adds the share of viewers who chose to view. confirmed · 4, 5

Shorts. YouTube says it built separate recommendation systems for Shorts and long videos, because viewers engage with a broader range of content in the Shorts feed. confirmed · 7 It says that feed "may tune up on the recency of content", and personalizes on the Shorts, channels, topics and trending sounds a viewer has enjoyed or engaged with. confirmed · 5, 6 Interest can carry across formats, but YouTube says viewers crossing over is "not common behavior". confirmed · 6

What I'd do: earn the watch and the satisfaction, not just the click

These are my deductions from the sources above, not YouTube's published rules.

• Make the thumbnail and title promise what the video delivers. The 2016 ranking was built so clicks without watching don't pay off, and YouTube says it also asks viewers to rate what they watched. confirmed · 1, 2 A package that over-promises may win the click but will likely lose out in ranking. inferred
• Cut filler instead of chasing a length. YouTube names both average view duration and average percentage viewed, says there's no universal ideal length, and advises against filler. confirmed · 4, 6 That points to a video people stay with to the end, at whatever length the idea needs. inferred
• Build for the return visit. In 2016, the most important ranking signals were a viewer's past interactions with similar videos, such as your channel and topic. confirmed · 1 Today YouTube says long-term performance can suffer if a viewer "consistently stops watching videos from a channel when they are recommended". confirmed · 6 It also says a library of good videos helps new viewers go deeper. confirmed · 6 So a clear topic and a back catalog worth watching next likely matter more than any single upload. inferred
• Post what your audience will choose, not just more. In 2016, an ignored recommendation got pushed down for that viewer next time. confirmed · 1 Today "Not interested" is a named signal, and YouTube says there's no minimum posting cadence. confirmed · 5 Fewer videos people pick likely beat many they skip. inferred
• Bring your own first viewers. In 2016, the system learned from all watches, embeds on other sites included, so discoveries outside recommendations could spread to similar viewers. confirmed · 1 Sending a new video to people who already care about its topic, through a newsletter, a community or an embed on your site, likely gives it useful early viewing data, if they really want it. inferred
• Judge Shorts and long videos separately. They run on separate systems, and few viewers cross formats. confirmed · 6, 7 In Studio, Shorts analytics show how often viewers chose to view versus swiped away, which matches a signal YouTube says it ranks on. confirmed · 5, 8 That makes it a better first check on a Short than raw views. inferred

What nobody outside YouTube knows, and four common beliefs checked

The one detailed engineering description of YouTube's recommender I could verify dates from 2016. I found no YouTube engineering paper on how Shorts are ranked. No source I read gives the weight of each signal, or how the Shorts and long-form systems share information. I watch videos all day, and I still can't see YouTube's weights. Nobody outside can.

• "YouTube tests every new video on a small group first." Half true. No published paper or official page I found describes a test group, a size or a window. But the system in the 2016 paper picks candidates for each viewer on every request and shows the highest-scoring ones, and YouTube says it watches how a video performs when recommended. confirmed · 1, 4 So a new video's first viewers are mostly the people it's predicted to suit, not a random group, and how they watch shapes what happens next; that's my inference. inferred (see the myths page)
• "Posting Shorts hurts your long videos." Contradicted by YouTube, which says Shorts performance doesn't negatively impact long-form recommendations and that it has seen no evidence of harm. confirmed · 5, 7 That's YouTube's own claim, and outsiders can't check it.
• "Longer videos always win because YouTube counts watch time." Partly true, and dated. The 2016 ranking was based on expected watch time per impression. confirmed · 1 On its own, that leans toward videos that keep people watching longer. inferred But YouTube's later statements also count satisfaction surveys, name average percentage viewed alongside view duration, and say there's no ideal length. confirmed · 2, 4, 6
• "One flop hurts your whole channel." Contradicted, with a caveat. YouTube says one video's underperformance doesn't penalize a channel overall. confirmed · 6 But it says long-term performance can suffer if particular viewers consistently stop watching your videos when they're recommended. confirmed · 6

Also, "the algorithm" isn't one thing. YouTube says search, the Shorts feed and the Subscriptions tab may be tuned for different goals; Subscriptions, for instance, lists videos from followed channels, most recent first. confirmed · 6