Explainer
How recommendation algorithms work
Recommendation systems usually work in stages: they pull a shortlist of candidates from a huge catalogue, then rank and re-rank the list for extra goals such as freshness and diversity.[1][2] Platforms say they rely mostly on what you do, such as what you watch, like, share or finish, plus information about the content itself.[3][4][5] In the EU, the Digital Services Act requires platforms to explain the main parameters of their recommender systems, and the largest must offer an option not based on profiling.[6][7]
This explainer describes the general design taught in Google’s machine-learning course and what YouTube, TikTok and Netflix say about their own systems. Companies do not publish their full ranking formulas, so platform descriptions are self-reported. Choose a reading level in each section.
The basic pipeline
A video app may have millions of videos but can only show you a few. So its recommendation system works like a funnel. First it grabs a shortlist of likely candidates, then it puts them in order and tidies up the final list.[1][2]
Under the hood, the system turns each item and each request into a list of numbers, a kind of map position. Things that are similar, like videos usually watched by the same people, end up close together on the map.[8]
In Google’s course, candidate generation is the first stage: given a query, the system generates a set of relevant candidates.[1] Both main candidate-generation methods represent items and queries as vectors in an embedding space, where items watched by the same users end up close together.[8] The final stage re-ranks candidates against extra criteria or constraints, for example filters or score transformations based on video age or click-bait detection.[2]
Two ways to find candidates
Content-based filtering looks at what you already like and finds similar things. If you watch lots of cooking videos, it suggests more cooking videos.[9]
Collaborative filtering looks at people like you. If people with tastes similar to yours loved a film you have not seen, it suggests that film. This can surprise you with things you would not have searched for.[10]
Content-based filtering scores items by their features against a user’s previous actions or explicit feedback.[9] Collaborative filtering uses user and item similarities simultaneously, which enables serendipitous recommendations (user A gets what similar user B liked), and learns its embeddings automatically rather than from hand-built features.[10]
What platforms say they measure
TikTok says its For You feed looks at what you like, share, comment on and follow, details of each video such as captions, sounds and hashtags, and settings such as your language.[3] Watching a long video all the way to the end counts for more than weaker signals.[3]
Netflix says it uses your viewing history and ratings, what similar members watch, and facts about each title such as genre and release year. It also considers the time of day, your language and your device. It says it does not use your age or gender.[4]
YouTube says it uses what you watch, like and dislike, your subscriptions and answers to satisfaction surveys, and it also weighs a channel’s reputation and quality.[5]
TikTok weights strong indicators, such as completing a longer video, above weak ones, and lists interactions, video metadata and device or account settings as inputs.[3] It says follower count and past hits are not direct ranking factors.[11] Netflix lists interactions, similar members’ preferences, title metadata, time of day, language, device and viewing duration, and says demographics are excluded.[4] YouTube combines viewing behaviour and explicit feedback, including satisfaction surveys, with channel-level reputation and quality.[5]
Filter bubbles and diversity
If a system only ever shows you more of what you already watch, your feed can get narrow. This is sometimes called a filter bubble.[12] TikTok says it deliberately mixes in different types of content to break up repetitive patterns.[12] Google’s course lists freshness, diversity and fairness as goals designers can build into the last stage of the pipeline.[2]
TikTok acknowledges that optimising for personalisation and relevance risks an increasingly homogeneous stream, and says it intersperses diverse content.[12][12] Google’s course discusses freshness, diversity and fairness under re-ranking.[2] Similar ranking ideas drive search engines and ad auctions.
Rules in the EU
The EU’s Digital Services Act treats recommendation systems as something users should understand. Platforms must explain in plain language the main factors their recommendations use, and any settings you can change.[6] The biggest platforms and search engines must also give you at least one feed option that is not based on profiling you.[7]
The DSA defines recommender systems broadly, covering any fully or partially automated suggestion or prioritisation of information.[13] Article 27 requires all online platforms using them to disclose the main parameters and any user options in their terms and conditions.[6] Article 38 adds, for very large online platforms and very large online search engines, a duty to provide at least one option per recommender system not based on profiling.[7]
Read next
How search engines work
How search engines crawl the web, build an index and rank results, what robots.txt does, how AI answers fit in, and the US antitrust ruling on Google.
Questions readers ask
What is a recommendation algorithm?
It is an automated system that suggests or prioritises content for each user. The EU Digital Services Act defines a recommender system as a fully or partially automated system a platform uses to suggest or prioritise specific information to its users.[13]
What is the difference between content-based and collaborative filtering?
Content-based filtering recommends items similar to ones you already liked, using the items' features. Collaborative filtering uses similarities between users and items together, so it can suggest something because people with similar tastes liked it.[9][10]
Does follower count decide what goes viral on TikTok?
TikTok says neither follower count nor an account's past high-performing videos are direct factors in its recommendation system.[11]
Does Netflix use my age or gender for recommendations?
Netflix says its recommendation system does not use demographic information such as age or gender. It relies on viewing history, ratings, the tastes of similar members and information about titles.[4]
Can I turn off personalised recommendations?
In the EU, very large platforms must offer at least one option for each recommender system that is not based on profiling. YouTube also lets users pause, edit or delete their watch and search history.[7][5]
What is a filter bubble?
It describes a feed that keeps showing more of the same. TikTok acknowledges that optimising for personalisation risks an increasingly homogeneous stream of videos and says it mixes in diverse content to break repetitive patterns.[12]
Sources
· 13 claims from 5 publishers
Each numbered claim is a statement we checked against the sources listed with it. Tap a number in the text to see its sources in place.
- [1]
In Google's machine-learning course, candidate generation is the first stage of a recommendation system, in which the system picks a set of relevant candidates for a query; two common approaches are content-based filtering and collaborative filtering. confirmedas of 2026-11-07
- Candidate generation overview · Google for Developers · Primary source · Candidate generation overview (retrieved 2026-11-07)
- [2]
In the final stage a system can re-rank candidates to account for extra criteria, for example filtering or transforming scores based on video age or click-bait detection, and to improve freshness, diversity and fairness. confirmedas of 2026-11-07
- Re-ranking · Google for Developers · Primary source · Re-ranking (retrieved 2026-11-07)
- Re-ranking · Google for Developers · Primary source · Re-ranking (retrieved 2026-11-07)
- Re-ranking · Google for Developers · Primary source · Re-ranking (retrieved 2026-11-07)
- [3]
TikTok says its For You feed ranks videos using user interactions such as likes, shares, follows and comments, video information such as captions, sounds and hashtags, and device and account settings such as language, country and device type (which it says get lower weight), weighting strong signals such as watching a longer video to the end more heavily than weak ones. confirmedas of 2020-06-18
- How TikTok recommends videos #ForYou · TikTok Newsroom · Primary source · 2020-06-18 · What factors contribute to For You? (retrieved 2026-11-07)
- How TikTok recommends videos #ForYou · TikTok Newsroom · Primary source · 2020-06-18 · What factors contribute to For You? (retrieved 2026-11-07)
- How TikTok recommends videos #ForYou · TikTok Newsroom · Primary source · 2020-06-18 · What factors contribute to For You? (retrieved 2026-11-07)
- How TikTok recommends videos #ForYou · TikTok Newsroom · Primary source · 2020-06-18 · What factors contribute to For You? (retrieved 2026-11-07)
- [4]
Netflix says it estimates how likely a member is to enjoy a title from their interactions such as viewing history and ratings, from members with similar tastes and from title information such as genre and release year, along with time of day, language, device and viewing duration, and that it does not use demographic data such as age or gender. confirmedas of 2026-11-07
- How Netflix's recommendations system works · Netflix Help Center · Primary source · How Netflix's recommendations system works (retrieved 2026-11-07)
- How Netflix's recommendations system works · Netflix Help Center · Primary source · How Netflix's recommendations system works (retrieved 2026-11-07)
- How Netflix's recommendations system works · Netflix Help Center · Primary source · How Netflix's recommendations system works (retrieved 2026-11-07)
- [5]
YouTube says its recommendations draw on a user's viewing behaviour, likes, dislikes, subscriptions and feedback including satisfaction surveys, also weigh a channel's reputation and quality, and that users can pause, edit or delete their search and watch history. confirmedas of 2026-11-07
- How YouTube recommendations work · YouTube · Primary source · How YouTube recommendations work (retrieved 2026-11-07)
- How YouTube recommendations work · YouTube · Primary source · How YouTube recommendations work (retrieved 2026-11-07)
- How YouTube recommendations work · YouTube · Primary source · How YouTube recommendations work (retrieved 2026-11-07)
- [6]
Under Article 27 of the Digital Services Act, online platforms that use recommender systems must set out in plain language in their terms and conditions the main parameters those systems use and any options users have to change them. confirmedas of 2022-10-27
- Regulation (EU) 2022/2065 (Digital Services Act) · EUR-Lex (Publications Office of the EU) · Primary source · 2022-10-27 · Article 27(1) (retrieved 2026-11-07)
- [7]
Article 38 of the Digital Services Act requires very large online platforms and search engines to offer at least one option for each recommender system that is not based on profiling. confirmedas of 2022-10-27
- Regulation (EU) 2022/2065 (Digital Services Act) · EUR-Lex (Publications Office of the EU) · Primary source · 2022-10-27 · Article 38 (retrieved 2026-11-07)
- [8]
Both filtering approaches use embedding spaces that represent items and queries as vectors, where similar items, such as YouTube videos usually watched by the same user, end up close together. confirmedas of 2026-11-07
- Candidate generation overview · Google for Developers · Primary source · Candidate generation overview (retrieved 2026-11-07)
- Candidate generation overview · Google for Developers · Primary source · Embedding space (retrieved 2026-11-07)
- [9]
Content-based filtering recommends items similar to what a user already likes, using the features of the items and the user's previous actions or explicit feedback. confirmedas of 2026-11-07
- Content-based filtering · Google for Developers · Primary source · Content-based filtering (retrieved 2026-11-07)
- [10]
Collaborative filtering uses similarities between users and items at the same time, so it can recommend an item to one user based on the interests of a similar user, and it can learn its embeddings automatically rather than relying on hand-built features. confirmedas of 2026-11-07
- Collaborative filtering · Google for Developers · Primary source · Collaborative filtering (retrieved 2026-11-07)
- Collaborative filtering · Google for Developers · Primary source · Collaborative filtering (retrieved 2026-11-07)
- [11]
TikTok says neither an account's follower count nor its past high-performing videos are direct factors in its recommendation system. confirmedas of 2020-06-18
- How TikTok recommends videos #ForYou · TikTok Newsroom · Primary source · 2020-06-18 · What factors contribute to For You? (retrieved 2026-11-07)
- [12]
TikTok acknowledges that optimising for personalisation risks an increasingly homogeneous stream of videos, sometimes called a filter bubble, and says it intersperses diverse content to interrupt repetitive patterns. confirmedas of 2020-06-18
- How TikTok recommends videos #ForYou · TikTok Newsroom · Primary source · 2020-06-18 · Interrupting repetitive patterns (retrieved 2026-11-07)
- How TikTok recommends videos #ForYou · TikTok Newsroom · Primary source · 2020-06-18 · Interrupting repetitive patterns (retrieved 2026-11-07)
- [13]
The EU Digital Services Act defines a recommender system as a fully or partially automated system an online platform uses to suggest or prioritise specific information to its users. confirmedas of 2022-10-27
- Regulation (EU) 2022/2065 (Digital Services Act) · EUR-Lex (Publications Office of the EU) · Primary source · 2022-10-27 · Article 3(s) (retrieved 2026-11-07)
Revision history (1)
- Page created.
Created Nov 7, 2026. Last reviewed by an editor on Nov 8, 2026. Next scheduled review: May 7, 2027.
Cite this page
"How recommendation algorithms work." ContentLora, updated Nov 7, 2026. https://contentlora.com/explain/how-recommendation-algorithms-work
Spotted an error? Suggest a correction or emailcorrections@contentlora.com.