how recommendation algorithms actually work to personalize a Netflix, Spotify, or TikTok feed

How Recommendation Algorithms Actually Work

By AISEN · Published August 30 2026 – Updated November 6, 2026

Quick answer

A recommendation algorithm predicts what you are likely to watch, play or keep scrolling, using your own behavior and the behavior of people who resemble you. Netflix’s researchers reported in 2015 that recommendations influenced about 80% of the hours people streamed, and described the system’s main business job as keeping subscribers. The story of Netflix’s $1 million prize shows that the most accurate model is not always the one a company puts into production.

Key takeaways

  • Two foundational approaches are collaborative filtering (people like you liked this) and content-based filtering (this is similar to what you liked). Real platforms usually combine them with other machine-learning models, ranking methods and editorial rules.
  • Likes are only one signal. Platforms also use behavior such as watching to the end, skipping, saving and sharing, and none of them publishes exact weights.
  • In 2015, Netflix estimated that the combined effect of personalization and recommendations saves it more than $1 billion a year, mostly through fewer cancellations. That is the company’s own estimate, not an audited figure.
  • A more accurate algorithm is not always a better product. Netflix did not deploy its $1 million prize-winning blend of models, because the extra accuracy was not worth the engineering effort.
  • These systems change. TikTok’s US algorithm is being retrained on American data under a new ownership structure.
  • You can steer your own feed with the controls each platform provides. The “Try it yourself” section shows where to find them.

What is a recommendation algorithm?

It is a system that predicts what one specific person will want next, instead of showing everyone the same list. A search or sort tool organizes what exists. A recommender predicts something that has not happened yet: whether you will finish this show, replay this song or scroll past this video.

Two foundational approaches

Collaborative filtering: “people like you also liked…”

This method compares your behavior with other users’ behavior and does not need to understand the content at all. Here is a toy example (invented numbers, only to show the idea):

Show AShow BShow CShow D
Youlikedlikedlikednot seen
Samlikedlikedlikedliked
Priyadislikeddislikedlikeddisliked

Your history matches Sam’s closely and Priya’s poorly. Sam liked Show D, so Show D is recommended to you. The system never needed to know what Show D is about.

Content-based filtering: “this is similar to what you liked”

This method compares the items themselves: genre, cast, tempo, instruments, keywords. If you finish three slow-burn crime dramas, it looks for other items with those traits, whatever other users did.

Beyond the two

Real platforms rarely rely on just these two ideas. They blend them with other machine-learning models, with context such as device and time of day, with popularity and trending signals, and with editorial rules. Netflix’s 2015 paper describes its recommender as a collection of different algorithms serving different parts of the product, not a single one. It also says its main ranking algorithms work better when personalized signals are blended with a healthy dose of unpersonalized popularity.

How a recommendation system might work, step by step

Platforms do not publish their full designs, and they differ from each other. But published research describes a common pattern: narrow a huge catalog down in stages. The list below is a simplified illustration, not a claim that Netflix, TikTok and Spotify use an identical architecture.

  1. User activity. Your watches, searches, skips, saves, shares and likes are recorded.
  2. Candidate generation. The system narrows a huge catalog to a few hundred items that might be relevant. YouTube’s engineers describe going from millions of videos to hundreds at this stage (Covington et al., 2016).
  3. Ranking. A separate model scores those candidates by how likely each is to suit you. In the same paper, this stage narrows the list to dozens.
  4. Filtering and diversification. Rules and models remove ineligible or repetitive content and add variety. TikTok says your feed generally will not show two videos in a row with the same sound or creator, and Netflix’s paper describes page-building that weighs both relevance and diversity.
  5. The feed you see. Your reactions become new activity, and the loop starts again.

The Netflix story: what the numbers actually say

In 2015, two Netflix executives, Carlos Gomez-Uribe and Neil Hunt, published a paper describing the company’s recommender system. It is one of the few times a major platform has shared real numbers. These are Netflix’s own estimates from 2015, not audited or current figures:

  1. Recommendations influence about 80% of the hours streamed. The other 20% comes from search. The paper also says two of every three hours streamed are discovered on the homepage.
  2. The business case is retention. The authors estimate that the combined effect of personalization and recommendations saves Netflix more than $1 billion a year, mainly through lower cancellations.
  3. Attention is short. Netflix’s consumer research suggests a typical member loses interest after about 60 to 90 seconds of choosing, having looked at 10 to 20 titles. This figure comes from the same 2015 paper.

The paper also says Netflix judges algorithm changes mainly by whether they improve member retention and medium-term engagement. In other words, the system is built less to find your perfect movie and more to help you find something worth watching before you give up and cancel.

Why Netflix didn’t deploy its $1 million winning algorithm

In 2006, Netflix offered $1 million to any team that could improve the accuracy of its Cinematch rating predictions by 10%. In 2009, the team BellKor’s Pragmatic Chaos won. The popular summary says Netflix never used the winner. The fuller story has three parts:

  • Netflix did use algorithms from the competition. Its 2015 paper says the contest produced algorithms that it still uses in production to predict ratings.
  • It did not deploy the final winning blend of models. In a 2012 blog post, Netflix engineers said the additional accuracy did not seem to justify the engineering effort needed to run it in production.
  • The business had changed. The contest was built around star ratings from the DVD era. By streaming, Netflix had much richer data about how people watch, and the paper says there are better ways to help people choose than predicting star ratings alone.

The lesson is bigger than Netflix: the most accurate model on paper is not automatically the right one for a real product.

TikTok, Spotify and Instagram: what the companies say they use

TikTok

TikTok’s newsroom describes three kinds of input for its For You feed: your interactions (such as likes, shares, comments, follows and content you create), information about the videos (such as captions, sounds and hashtags), and device and account settings like language and country, which TikTok says get lower weight. TikTok’s recommendation transparency page also lists videos you watched, finished, liked or skipped among the recent interactions it uses. TikTok says it weights these factors by their value to the user, but it does not publish the weights. Following an account is one signal among several, not the only one. TikTok also says videos that have just been uploaded or are under review may not yet be eligible for recommendation.

TikTok in 2026: the algorithm is being retrained

On January 22, 2026, a new company, TikTok USDS Joint Venture LLC, took over TikTok’s US operations. Oracle, Silver Lake and MGX each hold 15%, and ByteDance keeps 19.9%. According to Reuters, via RNZ, the venture will retrain and update the recommendation algorithm on US user data, with the algorithm secured in Oracle’s US cloud. Reports also say ByteDance licenses the algorithm to the venture and continues to run TikTok’s commercial activities in the US. What this means for what users actually see is, in the words of Engadget, still unclear.

Spotify: editors and algorithms together

Spotify says its recommendations combine editorial curation, where editors use local knowledge to program playlists, with personalised recommendations from its algorithms. It says actions such as searching, listening, skipping and saving to Your Library influence how it interprets your taste, and that recommendations should not only optimise for the next click.

Artists and labels can also mark songs as a priority through Discovery Mode. Spotify says this adds a signal to the algorithms behind personalised listening sessions such as Radio and Autoplay. It costs nothing upfront, but Spotify applies a 30% commission to the recording royalties from streams of those songs in Discovery Mode contexts. Streams outside those contexts stay commission-free (Spotify for Artists).

Instagram

Instagram’s head, Adam Mosseri, has said the top signals for ranking are watch time, likes and sends (shares by direct message), as reported by Social Samosa. Meta does not publish exact weights.

So is it just likes, and which signal counts most?

No, it is not just likes. Every company above lists several signals, including likes, shares or sends, follows, and viewing behavior such as watching, finishing, skipping and saving. None of them publishes numeric weights, so no one outside the companies can say exactly how much each counts. Claims that one action is worth a fixed multiple of another, such as saves or sends counting several times more than likes, circulate in marketing blogs, but we could not find them in company documents, so treat them as guesses.

What each company says its recommendations are for

PlatformWhat the company saysWhere it says it
NetflixHelp members find something worth watching quickly, which prevents abandonment and supports retention2015 paper
TikTokShow each user content likely to interest them, while mixing in variety so they can find new creators and categoriesTikTok newsroom
SpotifyPersonalise listening, and not just optimise for the next clickSpotify

These are the companies’ own descriptions. They do not tell you how much each goal counts inside the company.

Why recommendations go wrong

  • Cold start. A new user or new item has no history, so the system has little to go on. Netflix says it asks new members to pick titles in a sign-up survey for this reason.
  • Feedback loops. If you only get shown what resembles your past behavior, your future behavior resembles your past even more. Netflix’s own paper calls this feedback loop an open problem.
  • Opacity. Platforms rarely explain why a specific item appeared.
  • Optimizing the wrong thing. As the Netflix Prize shows, a system can be very accurate and still not serve the product’s goal.

Try it yourself: steer your own feed

The platforms publish controls that change what their systems learn about you. This section follows each platform’s own help pages (see Sources). Menu names change often, so look for the closest match in your app.

PlatformWhat you can doWhat the platform says it uses
NetflixRate titles with thumbs up or down; remove titles from your viewing activity in your accountNetflix says it looks at the genres you watch, your viewing history, your ratings, and the ratings of members with similar tastes
YouTubeChoose “Not interested” or “Don’t recommend channel” on a video; remove videos from your watch history; pause or delete your historyYouTube says your watch history, search history and this feedback can be used to tune recommendations
TikTokLong-press a video and choose “Not interested”; in Settings, open Content Preferences for “Manage topics”, keyword filters and “Refresh your For You feed”TikTok says these tools help you shape what appears in your For You feed
SpotifyUse Private Session for listening you do not want to count; exclude playlists from your taste profileSpotify says actions like searching, listening, skipping and saving influence how it interprets your taste

TikTok’s refresh option resets the feed, so read its confirmation prompt before you tap it.

A one-week experiment: can you change your own recommendations?

Setup takes about five minutes. The experiment itself takes a week.

  1. Pick one platform and one topic you want to see less of, or more of.
  2. Screenshot the first ten items on your home or For You feed, and count how many are about that topic.
  3. For a week, use the controls above every time that topic appears. To see less, mark it “Not interested” or exclude it. To see more, watch or listen to the end.
  4. Take the same screenshot after seven days and count again.

Platforms do not publish how much each action counts, so your result is your own. One account over one week shows what happened to you, not a rule for everyone. If you try it, we would like to hear what changed.

Frequently asked questions

Does watching something once change my recommendations? Platforms do not say how much a single action counts. TikTok says every new interaction helps its system learn, and Spotify says actions like skipping and saving influence how it reads your taste, so a pattern of actions is likely to matter more than one.

Can I turn off personalized recommendations? You can limit what a platform learns from you, for example by pausing watch history or using Private Session on Spotify. Whether you can switch personalization off entirely depends on the platform and your region, so check its settings.

Does everyone see the same content? No. Even two accounts that follow the same people can see different items in a different order.

Do these apps track what I do outside the app? It depends on the platform and its privacy policy. The platform’s own documentation is the only reliable answer.

The bottom line

Platforms differ, but the published research and company explanations share a pattern: learn from what you do, narrow a huge catalog to a few good candidates, rank them, and keep adjusting as you react. Netflix’s own numbers show how much rides on it, and its $1 million prize shows that the goal is a product that works, not the most accurate model.

Sources

Also Read: How to go viral on social media