How the YouTube Algorithm Really Works in 2026
Peer-reviewed papers from YouTube engineers describe the recommender as a staged, multi-objective prediction system. Here is what they document, and what remains folklore.
Creators often describe YouTube's recommendations as a single "algorithm" with secret rules. The published research tells a more useful story: recommendation is a pipeline of machine-learned models trained to predict user behaviour, with objectives chosen by engineers and refined by user feedback. YouTube's engineers have described the architecture in peer-reviewed papers, and while the production system has certainly changed since, those papers remain the best primary sources for how such a system is built and why it behaves as it does. This article draws on them, separates what is documented from what is folklore, and flags what nobody outside the company can verify.
A two-stage design: candidates, then ranking
The platform hosts an enormous catalogue, so scoring every video for every viewer on every visit is infeasible. Covington, Adams and Sargin (2016) described a two-stage solution that has become a template for large recommenders. First, a candidate generation network narrows the corpus to a few hundred videos using the user's history, search queries and context such as device. It is framed as an extreme multiclass classification problem, predicting which video a user will watch next, and it is trained on features like embedded watch history and demographic signals. Second, a ranking network scores that short list with a much richer set of features about the video and the user, and the ordered result is what appears on the screen.
Two design points from that paper matter to anyone interpreting creator advice. The ranking model was trained to predict expected watch time, not clicks alone. The authors explained that optimising click probability favours "clickbait" with misleading thumbnails, whereas weighting positive examples by the time watched favours content people actually watch. And the model includes the age of the video as a feature, because users show a strong preference for fresh content and the training data otherwise biases toward older, popular items.
Multiple objectives and selection bias
A later paper by Zhao and colleagues (2019) described how the ranking stage handles several goals at once. Some are engagement objectives (clicks, watch time), others are satisfaction objectives (likes, ratings, survey responses). A multitask architecture predicts each, and the predictions are combined into a final score. The paper also tackles a subtle problem: users click what they are shown, so logged data is biased by the position of items on the page. A "shallow tower" within the model estimates and subtracts this position effect. For researchers, this is a concrete example of why observational data from a platform cannot be read at face value: interface placement leaks into every behavioural signal.
The takeaway is that "the algorithm" is not one number. It is a set of predictions, each of which can be wrong in a different way, combined according to weights the platform can change at any time. Public statements from YouTube have said it uses viewer surveys and signals beyond raw watch time; the exact weights are not published.
Reinforcement learning and long-term effects
Because each recommendation affects what a user sees next, researchers have studied the problem as sequential decision-making. Ie and colleagues (2019) proposed a method for slate-based recommendation, where several videos are displayed together and the user chooses at most one, and described its use in experiments on a video platform. The motivation is that a myopic model can optimise immediate engagement at the expense of the longer-term relationship with the viewer. How much of this work is in production in any form is not something outside observers can determine.
What the evidence says about creator folklore
Several widely repeated claims have no public support in the primary literature, and treating them as laws is risky:
- Specific retention thresholds. No published paper states that a video must hold a given percentage of viewers to be promoted. Retention is an input to a learned model, and the model's output depends on the viewer, the context and the competition.
- Exact title or keyword placement rules. Search ranking relies on text relevance, engagement and personalisation, but figures such as a character limit for keywords are advice, not documented behaviour.
- Short-form mechanics. Claims about replay boosts or a fixed early-engagement window for short videos are creator reports. They are plausible, because replays are a strong engagement signal, but they are not confirmed by any technical publication.
What is supported is the general principle: the system is trained to predict viewing behaviour, so videos that people choose and continue to watch are favoured, and signals of dissatisfaction (quick abandonment, negative survey responses) count against them. That is consistent with creator experience without requiring any secret rule.
Measuring recommendation effects from outside
External researchers cannot see the model, so they audit it. Ribeiro and colleagues (2020) analysed paths through channels and recommendations to examine whether viewers are drawn toward more extreme content, finding evidence of migration between communities in their data while noting the limits of what an audit can say about the causal role of the recommender versus user choice. The study is a good illustration of method: sock-puppet or real-user traces, defined categories of content, and careful language about what is inferred.
Limitations and open questions
- Staleness. The published architectures date from 2016 to 2019. The production system in 2026 is unlikely to match them in detail, so they should be read as design patterns.
- Feedback loops. Models trained on behaviour they helped shape can reinforce existing tastes. Quantifying this effect needs experiments only the platform can run.
- Satisfaction is hard to measure. Survey responses are sparse and subject to response bias, so the "satisfaction" objective is itself a noisy estimate.
- Heterogeneity. Aggregate advice hides differences between genres, languages and audience sizes.
Questions worth testing
- Within one channel, do changes in thumbnail or title shift click-through and retention in opposite directions, and which dominates subsequent impressions?
- How do impressions change after a publication gap, controlling for upload age and topic?
- Do audience-retention curves differ systematically by format (tutorials, explainers, commentary), and are they better compared within format than across?
- For researchers auditing recommenders: how stable are results across logged-in versus logged-out sessions and across regions?
References and further reading
- Covington P., Adams J., Sargin E. Deep Neural Networks for YouTube Recommendations. Proceedings of the 10th ACM Conference on Recommender Systems (RecSys), 2016.
- Zhao Z. et al. Recommending What Video to Watch Next: A Multitask Ranking System. RecSys, 2019.
- Ie E. et al. Reinforcement Learning for Slate-based Recommender Systems: A Tractable Decomposition and Practical Methodology. IJCAI, 2019.
- Ribeiro M. H., Ottoni R., West R., Almeida V., Meira W. Auditing Radicalization Pathways on YouTube. ACM FAT*, 2020.