How YouTube recommendations work: what is confirmed vs what is inferred
Short answer: YouTube confirms that its recommendations are personalized, use viewer history and satisfaction signals, and differ by surface (Home, Suggested, Search, Shorts). Google research papers describe a two-stage design (candidate generation, then ranking) with multiple objectives. Beyond that, most "algorithm secrets" are inference or creator anecdote. This article labels each claim with its evidence tier so you can tell which is which.
The evidence tiers we use
| Tier | Source type | How we treat it |
|---|---|---|
| 1 | YouTube Help, official blog, official statements | Confirmed as of the page date |
| 2 | Google/YouTube research papers | Describes a published system, not necessarily today's production system |
| 3 | Credible employee interviews | Useful context; weaker than Tier 1 |
| 4 | Creator anecdotes, SaaS dashboards | Not accepted as proof of mechanism |
Tier 1: what YouTube states
- Recommendations use signals such as watch and search history, subscriptions, likes and dislikes, "Not interested", "Don't recommend channel" and satisfaction surveys (YouTube Help, support.google.com/youtube/answer/6342839).
- YouTube has said it does not treat all watch time as equal and uses surveys to understand satisfaction (YouTube Blog, "On YouTube's recommendation system", 2021, blog.youtube).
- YouTube's Help guidance on click-through rate warns against chasing CTR through misleading packaging.
- As of December 2025, Studio supports title-and-thumbnail A/B testing; the winner is selected by watch-time share (check YouTube Help for current availability).
Tier 2: what Google's papers describe
- Two stages. Candidate generation narrows a huge corpus; a ranking model scores the candidates, optimizing expected watch time (Covington, Adams & Sargin, RecSys 2016, doi:10.1145/2959100.2959190).
- Multiple objectives. The watch-next ranking system predicts engagement and satisfaction objectives together and uses a shallow tower to reduce position bias (Zhao et al., RecSys 2019, doi:10.1145/3298689.3346997).
- Long-term value. A reinforcement-learning approach to slate recommendation (SlateQ) models long-term user value rather than only immediate clicks (Ie et al., IJCAI 2019, doi:10.24963/ijcai.2019/360).
Caveat: these papers are 2016 to 2019. YouTube has not said they describe the live system, and we do not assert that.
Reasonable inference (not confirmed)
- A video that viewers finish and rate well is more likely to keep being recommended. This follows from the satisfaction objectives above, but YouTube does not publish a formula.
- Channels with a clear single audience are easier to match. This is an inference from how matching works.
- Studio metrics (CTR, average view duration) are downstream shadows of the real objectives, not the objectives themselves.
Folklore to treat as unproven
- "Post at 3pm on Tuesday to beat the algorithm": no Tier 1 or 2 source.
- "The first hour decides everything": no Tier 1 or 2 source that we can verify.
- A single magic number for retention or CTR that unlocks reach: YouTube states CTR varies by source; thresholds you see online are creator or vendor data.
- Specific percent uplifts for tactics quoted without a primary source.
If we cannot find a Tier 1 or 2 source, we say "I cannot confirm this."
What to do with this
- Design each video for one identifiable viewer.
- Read CTR together with retention; do not optimize either alone.
- Use A/B testing for packaging questions where the platform measures the answer.
- Treat any tactic with no primary source as a hypothesis, test it once, and keep the result only if it replicates on your channel.
Our experience: On Decoroo, we treated consistency of audience and topic as the working principle over several years rather than chasing algorithm tricks. That is a description of our method, not proof of the mechanism. See the Decoroo case study and our approach.
Evidence limits
- Tier 2 papers may not describe production today.
- Our "reasonable inference" items are labeled because they are not confirmed.
- YouTube's features change; verify current behavior on YouTube Help.
Sources
- YouTube Help, "Manage your recommendations and search results." support.google.com/youtube/answer/6342839
- Goodrow, C., "On YouTube's recommendation system," YouTube Blog, 2021. blog.youtube
- Covington et al., RecSys 2016. doi:10.1145/2959100.2959190
- Zhao et al., RecSys 2019. doi:10.1145/3298689.3346997
- Ie et al., "SlateQ," IJCAI 2019. doi:10.24963/ijcai.2019/360
FAQ
Is there one YouTube algorithm? The papers describe different models for different surfaces; YouTube says it runs different systems for Home, Suggested, Search and Shorts.
Does watch time still matter? Yes, but YouTube says it also weighs satisfaction signals.