Inside the YouTube Algorithm: Related Graphs, Satisfaction Surveys and the Viewers Who Are Part of the Code

Fridman asks how YouTube decides what to show, both when someone searches and when it recommends what to watch next. Goodrow separates the two. Search uses the best technology Google has to match queries to videos, starting with obvious signals like whether the word appears in the title, and extending to syntactic or semantic matches and observations such as whether a video is watched a lot after a given query.

For recommendations, the first serious approach was collaborative filtering. YouTube observes which videos are watched close together by the same person and builds a graph where videos watched together by many people sit close to one another. This related graph groups videos by language, sports and music without anyone programming those categories, which Goodrow calls the lowest-hanging fruit in managing millions of videos.

He illustrates with a researcher from Turkey who believed YouTube must think she only speaks English because her academic searches returned English suggestions. When she searched for a baklava recipe in Turkish, the watch-next suggestions were in Turkish too, which she found astonishing. The related graph handles that automatically, since it reflects how people actually behave.

Fridman, who says he is fascinated by what a search history reveals about a person, asks whether YouTube has tried showing users a map of their interests. Goodrow says users can browse their watch history and the company has experimented with showing clusters. Internally the recommender sees a user as the history of videos watched, like a DNA strand or a vector in the space of all videos, and nearby vectors help produce diverse recommendations.

Fridman asks how a high-quality video with only a few views can reach more people and how quality is even defined. Goodrow says it depends on the genre. For news, science or medicine, quality relates to authority, credibility and expertise, such as a real journalism department with fact-checkers, while for entertainment such as pranks or Minecraft videos it is more about what people enjoy and feel satisfied with the next day.

He describes how the measure evolved. Views were the first approximation, but a video can attract many clicks and then be abandoned. Watch time improved on that, on the premise that time spent relates to value, but it is imperfect, as he illustrates by recalling clicking through TV channels late at night and ending up watching a film he later wished he had skipped.

That led to surveys. YouTube asks users afterward how satisfied they were and feeds that into its machine learning, so the system predicts not just what someone will click or watch for a while but what they will rate four or five stars tomorrow. Other signals include likes and dislikes, comments, shares and subscriptions.

Subscriptions are ambiguous, Goodrow says, because people use the button differently. Some viewers who watch many videos from a few channels do not subscribe, believing it costs money, while others subscribe widely to say thanks without watching. Fridman admits he subscribes to give a high five. Goodrow says the company cannot tell users they are using it wrong and has to embrace all the ways people use these buttons.

Asked about analyzing the actual content of videos, Goodrow says the ability is still somewhat crude. YouTube can tell music from sports and probably that people are playing soccer, but not whether it is Manchester United or his daughter's team. The signals mostly help refine clusters and sometimes add words like soccer to a video that lacks them in the title or description.

He therefore asks creators to help with literal titles and descriptions. His example is a World of Warcraft competition live stream titled with a match number and team names rather than the game's name, so searching for the game would not find it. Fridman notes that literal naming can conflict with wit and humor, and argues a mysterious title with a telling thumbnail can create a delightful experience.

Goodrow replies that human searchers also benefit: if every result says World of Warcraft except yours, a person may assume the search made a mistake. He concedes that those who do click an indirect title may be delighted, while others will skip it. Because collaborative filtering starts with the same people watching videos together, and they usually reach those videos through search, discoverability affects the related graph too.

Goodrow says the systems depend on a faithful amount of behavior, so people and machines trying to associate unrelated videos for profit mean the team must always be resilient to gaming. Fridman raises the clickbait titles and thumbnails creators use, citing Veritasium's Derek Muller, who has said he makes high-quality videos but uses attention-grabbing packaging to get clicks.

Goodrow compares it to his bookshelf. His math textbooks from Springer all look identical with yellow covers, whereas his popular-science books have provocative covers and titles. He does not consider the latter clickbait because they are good books, and he says choosing a compelling cover is a legitimate decision, though there is a line you should not cross.

If a thumbnail is especially racy or the title has too many exclamation points, users are sometimes offended, and YouTube suppresses such videos. Fridman praises the feature that lets users say they do not want to see a video again, and Goodrow agrees it is a strong signal. Fridman adds that products like Tesla make owners feel they are helping the system learn, something YouTube could remind users of when they click like or dislike.

Goodrow says experiences are very personalized for logged-in users, because viewers often want one of two things after a video: more on the same topic or something else. What counts as something else differs by person, for example the latest upload from a subscribed channel or from a channel they watch a lot. This applies to both watch-next and the home page.

As for success, he says the simplest indicator is that you come back on another day, since people do not return to things they find no value in. The ideal, though, is that every video someone watches is the best they have ever seen, which is why YouTube surveys them with one-to-five-star ratings. He admits achieving five stars every time is unrealistic and could only be approached asymptotically.

Fridman describes a Ray Dalio video on the economic machine that struck him as the best he had watched and says his utopia is a platform where every video helps him grow and enjoy life, with none he regrets. Goodrow calls the task ambitious.

Fridman asks whether the system resembles an autonomous vehicle stack, patches of heuristics tuned by experts, or a giant machine learning system. Goodrow says the team does not think of a single algorithm but of many systems serving different surfaces. He adds that what outsiders call the algorithm is code and machine learning combined with the behavior of all the people who come to YouTube, and without them it would not work.

Such systems evolve, he says, starting with a simple heuristic such as recommending the most popular videos, then collecting data and chipping away at the heuristics in favor of learned behavior. Diversity is one example: the first measure was to allow no more than three videos in a row from the same channel, and over time the team has made that more fine-grained and responsive to individuals.

To test changes, Goodrow says nearly every one goes through an A/B experiment that runs from a week to two months and tracks hundreds of variables with confidence intervals. Positive signs include fewer dismissals and more five-star ratings, and the team also runs user studies. Before launch they also want to know the impact not just on viewers but on different channels.