Here is a question: how does a video site know what to show you? The answer is a recommendation system. A research article defines recommender systems as software tools and techniques providing users with suggestions for items a user may wish to utilise. A recommendation system picks videos for you from a huge number of videos, and it suggests videos similar to ones a person watched. A 2016 Google research paper said YouTube represents one of the largest scale and most sophisticated industrial recommendation systems in existence. As of August 2026, Common Sense Media says YouTube's algorithm plays a big role in what kids and teens see.
The 2016 Google paper is split according to the classic two-stage information retrieval dichotomy. It first details a deep candidate generation model, and then it describes a separate deep ranking model. Candidate generation is the first stage of recommendation: given a query, the system generates a set of relevant candidates. In one common recommendation design, the system starts from a potentially huge collection and generates a much smaller subset of candidates. The model that generates candidates needs to evaluate queries quickly given the enormous size of the corpus. In a scoring stage, the system can use a more precise model relying on additional queries. In a re-ranking stage, the system can remove items the user explicitly disliked or boost the score of fresher content.
Systems also differ in what they compare. Content-based recommending uses similarity between items to recommend items similar to what the user likes. Google's course notes give an example: if user A watches two cute cat videos, the system can recommend cute animal videos to that user. In collaborative filtering, if user A is similar to user B and user B likes video 1, the system can recommend video 1 to user A. Underneath both methods sits behavior. In a Columbia Knight Institute explainer, engagement means any score defined only in terms of the moment-to-moment actions of the user. The explainer says the behavioral record is the fuel of the recommendation engine.
In a September 2021 post, YouTube said its system sorts through billions of videos to recommend content tailored to a viewer's specific interests. YouTube said its recommendations do not connect viewers to content through their social network. It said its system learns every day from over 80 billion pieces of information it calls signals. The importance of each signal depends on the viewer, YouTube said, and its system does not follow a set formula but develops dynamically as a viewer's habits change.
In its September 2021 post, YouTube said clicking on a video is a strong sign that a viewer will also find it satisfying. YouTube also said it learned in 2011 that clicking on a video does not mean the viewer actually watched it. As of September 2021, the Knight Institute says that before 2012, YouTube optimized for click-through rate instead of watch time, which led to clickbait thumbnails. As of September 2021, YouTube added watch time to its recommendations in 2012. YouTube said that when it first added watch time it saw an immediate 20% drop in views.
As of November 2026, the Knight Institute explainer says YouTube optimizes for expected watch time: how long the algorithm predicts the video will be watched. As of November 2026, it also says that if a user sees a video in their recommendations and does not click on it, the watch time counts as zero. As of November 2026, Common Sense Media says YouTube's business model rewards watch time: the longer users watch, the more YouTube makes in ad sales.
Ratings and reactions are signals too. YouTube said it asks why a viewer gave a video a one- or two-star rating. It said only videos rated four or five stars are counted as valued watch time. YouTube said a dislike is a signal that the viewer probably did not enjoy watching the video. All of this is from the September 2021 post.
Next, look at popularity and quality. If popularity is captured carelessly, popular items may end up dominating the recommendations. In September 2021, YouTube said an earlier system ranked videos by popularity to create one big Trending page. It also said it uses classifiers to identify whether a video is authoritative or borderline, and that any video classified as borderline is demoted in recommendations.
As of August 2026, Common Sense Media says the YouTube algorithm isn't designed with kids in mind and is designed to keep users engaged. Features such as autoplay are designed to keep kids watching, and autoplay and endless scrolling can make it hard to stop watching. As of November 2026, KidsHealth says most apps, games and streaming platforms are designed to keep users involved as long as possible.
Age rules come next. As of November 2026, Common Sense Media says YouTube requires users to be at least 13 years old to create an account. YouTube's terms of service say you must be at least 13 years old to use YouTube, though children of all ages may use YouTube and YouTube Kids if a parent or guardian enables it. In most countries, 13 is the minimum age to manage your own Google Account. For children under the minimum age, parents can help create and manage a Google Account with Family Link. As of August 2026, Common Sense Media says YouTube Kids is intended for the under 13 crowd.
Next comes a law about children's privacy online. Congress enacted the Children's Online Privacy Protection Act (COPPA) in 1998, and COPPA required the Federal Trade Commission to issue and enforce regulations about children's online privacy. As of November 2026, the COPPA Rule was designed to protect children under age 13. Websites covered by COPPA must notify parents and get their approval before collecting a child's personal information.
Here is the version made for younger viewers. As of November 2026, YouTube Kids says its Recommended videos are selected by its algorithm without human review. Recommended videos are also based on the content setting a parent selected. If a child's profile is set to Approved content only, Recommended videos are drawn from content the parent has approved. With Search turned off a child can't search for videos. Controls made for kids are not perfect, and a child may still see something a parent did not want.
A viewer has real controls over recommendations. As of November 2026, on certain pages, such as Home and Watch Next, a viewer can mark a video as Not interested. YouTube says Not interested and Don't recommend channel feedback may be used to tune recommendations. YouTube says that if a viewer deletes watch history, it won't base future video recommendations on that content. YouTube's help page gives one example of turning off watch or search history: researching a subject for a school project that you are not personally interested in.
Other controls are about time, as of November 2026. If Autoplay is on and a viewer wants to stop the next video, they can tap Cancel at the end of a video. On the Time management tab, a YouTube viewer can change the Take a break reminder and Auto-play settings. KidsHealth says to turn off autoplay on streaming platforms. The American Academy of Pediatrics' family media plan advises keeping room for other activities so screens do not crowd out other important experiences.
A debate about personalization has a name. In 2011, author and internet activist Eli Pariser coined the term filter bubble. In personalization, the program caters the content of a search or news feed to reflect the user's preferences, and personalization algorithms compile information to assign a relevancy score to content. Google, Facebook and similar websites claim their algorithms provide users with the content they want. Critics argue that filter bubbles skew a person's views on reality and leave them open to media manipulation.
The evidence is not settled. A policy research article says there is limited empirical evidence as to the extent, or harmfulness, of filter bubbles, and that many earlier studies did not consider personalisation, which is at the core of the filter bubble hypothesis. A systematic review of research on recommender systems reports evidence of filter bubbles in recommendation systems. The same review says several biases contribute to the existence of filter bubbles, and that incorporating diversity into recommendations can potentially help alleviate the filter bubble issue. As of April 2019, a UK government white paper said companies would be required to ensure that algorithms selecting content do not skew towards extreme and unreliable material in the pursuit of sustained user engagement.
Words to know
- algorithm: step-by-step rule
- signals: pieces of information
- autoplay: automatic next video
- personalization: fitting content
- engagement: user actions
