Open TikTok and a video is already playing. Open YouTube and dozens of thumbnails are waiting. Netflix has placed one series prominently at the top of the screen while quietly moving another several rows down. Spotify has decided what belongs on your Home page before you have typed the name of a single artist.
None of these screens is neutral. Before you make a choice, thousands of possible choices have already been reduced, ranked and arranged on your behalf. Something has been placed first, something else tenth, and most of what technically exists on the platform has not been shown to you at all. The modern internet increasingly works this way: instead of asking us what we want and waiting for an answer, it predicts what we are most likely to want next.
We usually call the thing making those decisions “the algorithm.” The phrase is convenient, but misleading. There is rarely one mysterious algorithm sitting behind a platform and deciding what everyone sees. Recommendation systems combine multiple models, rules, signals, business objectives, safety policies and sometimes human editorial decisions. Their job is not simply to find good content. It is to rank an overwhelming supply of content for a specific person, in a specific context, at a specific moment.
The question is no longer whether algorithms influence what we watch, hear and read. The more interesting question is what they are actually looking for when they decide what deserves to appear first.
Before Recommendations, the Internet Had an Order
The early web was not free from ranking. Search engines ordered results, editors selected homepage stories, music stores placed albums on particular shelves and television programmers decided what aired at 8 p.m. Somebody always controlled distribution. The difference is that those decisions were largely shared. Millions of people opened the same newspaper front page or watched the same television schedule.
Personalized recommendation systems changed that model. The front page could now be different for every person. Instead of asking an editor to predict what an audience might want, platforms could observe what each individual actually did and continuously rearrange the experience around those signals.
This solved a genuine problem. Netflix contains far more titles than anyone could realistically browse. YouTube receives more video than any editorial team could manually organize. Spotify cannot present millions of songs on a single screen. TikTok would be almost unusable if videos simply appeared in random order. Ranking is not an optional feature in environments with effectively infinite choice. It is the mechanism that makes them navigable.
But ranking also creates power. If only a tiny fraction of available content can appear in the first few positions, deciding what fills those positions can influence what becomes popular, what creators are discovered, which songs spread and even what subjects people encounter often enough to begin caring about them.
Visibility is no longer simply a consequence of popularity. Increasingly, visibility can help create it.
You Are Constantly Telling the System What You Want
Recommendation systems need signals. Some are obvious because we deliberately provide them: following an account, subscribing to a channel, liking a song, saving a film or pressing “Not interested.” Others are less explicit. How long did you watch? Did you finish the video? Did you immediately swipe away? Did you replay it? What did you search for afterwards?
TikTok has publicly explained that its For You system considers user interactions such as likes, shares, follows and comments alongside information about the video itself, including captions, sounds and hashtags. Not all signals are treated equally. TikTok has given completing a longer video as an example of a stronger indication of interest than simply sharing the same country as its creator. The system continually uses these interactions to rank videos according to how likely they are to interest a particular viewer.
YouTube describes a similarly enormous feedback system. Its recommendations use watch and search history, subscriptions, likes, dislikes, “Not interested” responses and satisfaction surveys, among other signals. According to YouTube, its recommendation systems learn from more than 80 billion pieces of information it calls signals. Crucially, different parts of YouTube do not rank content in exactly the same way. The homepage relies heavily on viewing history, while the video currently being watched becomes an important signal for what appears “Up Next.”
This creates a strange relationship between user and machine. We are not merely consuming recommendations; our behaviour while consuming them becomes training data for the next set. Watch one video about running shoes and nothing dramatic may happen. Watch several, search for a marathon plan and finish a review comparing five pairs, and the system receives a much clearer picture.
Every scroll is also feedback.
What You Do May Matter More Than What You Say
People are not always reliable narrators of their own preferences. Ask someone what kind of films they like and they might say serious dramas, then spend Saturday evening watching three romantic comedies. A recommendation system does not necessarily need to resolve the contradiction. It can simply observe the behaviour.
Netflix explains that its recommendations consider viewing history, previous ratings, people with similar tastes and information about titles themselves. But it also uses contextual information such as time of day, preferred language, device and how long a member watched a particular title. More recent behaviour can eventually outweigh preferences expressed earlier.
This means personalization is often less about constructing one permanent profile of “who you are” and more about predicting what version of you has opened the app right now. The person who watches documentaries on a laptop during weekday evenings may behave differently from the same person opening Netflix on a television late on Saturday night. YouTube likewise says time of day and device can influence recommendations because viewing habits may differ across contexts.
The distinction matters. Recommendation systems are frequently discussed as though they are trying to uncover our deepest, fixed preferences. In practice, many are solving a more immediate problem: given everything they know about previous behaviour and current context, what is this person most likely to choose now?
That makes them prediction machines rather than mind readers. Sometimes the prediction is excellent. Sometimes one curious click causes the system to spend a week trying to turn you into someone you never intended to become.
Netflix Does Not Just Choose the Movie. It Chooses the Shelf
Recommendation becomes even more powerful when it influences the architecture of the interface itself.
Netflix does not simply maintain a single ranked list of films and series. Its own explanation describes several layers of personalization on the homepage. The system can decide which rows appear, which titles belong inside those rows and the order in which the titles are displayed. Stronger recommendations generally receive more prominent positions.
That distinction is easy to overlook. If a platform only personalized the content inside a fixed interface, users would still share roughly the same map. But once the map itself becomes personalized, two people can open the same service and encounter substantially different versions of it.
Spotify operates with a mixture of approaches as well. The company says some recommendations are created through human editorial curation, while others are personalized by algorithms based on individual listening patterns. Its recommendation systems influence areas including Home, playlists and search, meaning discovery can be shaped before a listener actively chooses what to play.
This is why the phrase “the algorithm showed me this” is both right and incomplete. The algorithm may have selected the song, but it may also have determined which playlist you noticed, where that playlist appeared and what was positioned around it. Recommendation does not begin when you press play. It begins with the arrangement of the choices.
The System Is Not Simply Asking, “Will You Click?”
For years, the darkest version of the recommendation-system story has sounded something like this: platforms only care about keeping people online for as long as possible, so they recommend whatever produces the most engagement.
There is truth behind the concern, but modern systems are generally more complicated than a single metric such as clicks or watch time. YouTube, for example, says its recommendation system evaluates signals across three broad questions: whether viewers choose a video, whether they remain engaged once they begin watching, and whether they appear satisfied afterwards. Satisfaction surveys and explicit negative feedback can therefore matter alongside raw viewing behaviour.
This is partly practical. A system optimized entirely for immediate clicks could become very good at recommending sensational thumbnails that leave viewers disappointed. That might increase short-term engagement while making the platform less useful over time. Recommendation systems therefore face a difficult optimization problem: predict not merely what someone will choose, but ideally what they will be happy to have chosen.
Those goals do not always align. The video most likely to attract your attention may not be the video you later consider worthwhile. The song you replay obsessively this week may be one you never want to hear again next month. A recommendation engine has to translate vague human concepts such as interest, relevance and satisfaction into measurable behaviour.
No metric perfectly captures them.
Recommendation Systems Need to Show You Things They Know You Might Not Like
If an algorithm only gave you more of what had already worked, personalization would eventually become extremely repetitive. Watch cooking videos and every video would become cooking. Listen to one artist and every playlist would slowly converge on that artist. The system would become confident, predictable and boring.
Recommendation systems therefore have to balance exploitation and exploration. Exploitation means using what the system already knows about you. Exploration means deliberately testing something less certain to learn whether your interests might extend further.
TikTok has openly described this problem. It says its For You feed occasionally introduces content that may not obviously match a viewer’s expressed interests and works to avoid excessively repetitive patterns. The goal is partly to help people discover new categories and creators, but these unexpected recommendations also give the system new information. If you stop and watch, it has learned something. If you swipe immediately, it has learned something else.
This is why a strange video appearing in your feed is not necessarily evidence that the recommendation system has failed. It may effectively be asking a question.
Do you like this too?
That constant experimentation helps explain why personalized feeds can develop surprisingly quickly. The system does not need to wait until you explicitly describe your interests. It can repeatedly place small bets, observe the result and update its predictions.
Over time, you are not just browsing the feed. The feed is interviewing you.
Popularity Still Matters, Even in a Personalized World
Personalization can create the impression that everything in a feed was chosen exclusively because of our individual behaviour. That is not quite true.
Recommendation systems have to start somewhere. A brand-new user has little or no behavioural history, a problem usually known as the “cold start.” TikTok says new users can select initial interests, but if they do not, the platform may begin with a more generalized set of popular content before learning from their early interactions. Netflix similarly says new profiles can select titles they like, or begin with a diverse and popular selection until enough behaviour exists to personalize recommendations more meaningfully.
Even after personalization begins, broader popularity and performance can still matter. YouTube says videos compete against other videos a particular viewer might watch and that audience interest in a topic affects how much demand exists around it. A brilliant video about an extremely narrow subject does not compete in exactly the same environment as a major football story with a vast potential audience.
This creates feedback loops that can be difficult to separate. Content receives exposure because it performs well, then may perform even better because it received exposure. A song enters recommendations, attracts more streams and becomes even more attractive to systems looking for signals of interest. Popularity can generate visibility, while visibility generates more data that can reinforce popularity.
The ranking system does not merely observe culture from the outside. Once it becomes large enough, it participates in creating it.
There Are Humans Behind “The Algorithm” Too
Automated prediction is only part of the story. Platforms decide what their recommendation systems are allowed to recommend, what objectives they should optimize and what categories require different treatment. Safety policies can remove certain content from recommendation even when people might engage with it heavily. Human editors still build playlists, classify content and make programming decisions. Engineers choose which signals enter models and product teams decide where recommendations appear.
TikTok, for example, says certain content may be ineligible for recommendation even if it is permitted to remain on the platform, and that its systems deliberately introduce diversity to avoid overly homogeneous feeds. Spotify explicitly describes recommendation as a combination of people and technology, with human editorial curation operating alongside algorithmic personalization.
There are also commercial structures surrounding every recommendation engine. A platform may contain advertising, promoted content, licensed catalogues, subscriptions, contractual obligations and product priorities. These do not mean every recommendation is secretly purchased. They do mean that what appears on a screen emerges from a system designed by a company with its own definition of what a successful experience should look like.
There is no entirely objective order in which millions of songs, films, videos or posts are naturally supposed to appear. Someone has to define the ranking problem before a machine can solve it.
The Feed Learns You, but You Learn the Feed Too
Recommendation is often described as a one-way process: the platform studies the user and changes accordingly. Something else happens at the same time. Users learn what the platform rewards and begin changing their behaviour in response.
Creators learn which thumbnails attract clicks, how quickly a video needs to capture attention and which formats travel well through a recommendation feed. Musicians think about how songs begin in a streaming environment where skipping takes one tap. Publishers adjust headlines for social distribution. Restaurants design dishes that photograph well. Entire creative industries start adapting to the logic of discovery systems.
Users adapt too. We save videos to “teach” TikTok what we like, hide songs from Spotify, press “Not interested” on YouTube and create separate streaming profiles so one person’s viewing habits do not corrupt another’s recommendations. The audience becomes conscious of being modeled and occasionally tries to manage the model.
This creates a loop that would have been difficult to imagine in traditional media. Platforms predict behaviour, people react to the predictions, creators react to the distribution system and the resulting behaviour becomes new data for the next prediction.
The algorithm changes culture, and culture trains the algorithm back.
So Who Actually Decides What You See First?
The uncomfortable answer is that there is no single decision-maker.
You do, through everything you watch, skip, search, like and ignore. Other users do, because their behaviour helps platforms identify patterns and predict what people with similar tastes may enjoy. Creators do, by producing content designed to compete for attention. Human editors and product teams do, by choosing categories, rules and interfaces. Companies do, by deciding what their recommendation systems should optimize. Machine-learning models then process those signals at a scale no individual human could manage.
What appears first is the product of all of them.
That makes personalized recommendations neither entirely manipulative nor entirely passive. They are enormously useful systems built to solve a real problem: there is far more information available than anyone could possibly consume. But solving that problem requires ranking, and ranking means giving some things visibility before others.
A generation ago, the question was which television channel you would choose, which newspaper you would buy or which record you would take from the shelf. Today, the first selection increasingly happens before you arrive.
You still choose what to click.
But someone, or something, has already chosen what you get to choose from.
