How Do Recommendation Algorithms Choose What People See? | How It Works
On this page
What is the short answer?
A recommendation algorithm guesses which items a person is most likely to want next and puts the best guesses first. Its evidence is behaviour: what people rated, clicked, watched, liked or skipped.
Large services work in stages. A first step pulls a short list of candidates from a huge catalogue, and a second model scores them. What the score rewards, such as clicks or time spent, decides the order.
Definition: A recommender system predicts how much a person will value each item in a catalogue, such as a video, song or product, and ranks the items by that prediction to fill a feed.
What does a recommender learn from?
Its raw material is feedback, which Google's course on recommendation systems splits in two. Explicit feedback is a direct statement, such as a numerical rating of a film. Implicit feedback is inferred from behaviour: if a user watches a film, the system infers interest.
Implicit signals can mislead. In September 2021, YouTube's Cristos Goodrow, a vice-president of engineering, listed its system's signals: clicks, watch time, survey responses, shares, likes and dislikes. He wrote that YouTube learned in 2011 that a click does not mean the video was watched.
Arranged as a grid of users by items, the signals form a feedback matrix, mostly empty, since each person meets a small part of the catalogue.
How do content-based and collaborative filtering differ?
Content-based filtering uses the features of items. If a student watched several chess tutorials, it suggests other videos tagged as chess. Google's course says it needs no data about other users and can surface niche items, but needs much domain knowledge and only extends existing interests.
Collaborative filtering uses similarities between users and items at the same time. People with matching histories are taken to share tastes, so one person's favourite can pass to a look-alike. Google says this can help users discover new interests.
Collaborative filtering meets the cold-start problem. A brand-new item has no feedback, so the model cannot score it; one fallback Google names is the average learned profile of items in the same category. A new user, with no history, raises the same puzzle.
How does matrix factorisation predict a rating?
In October 2006, Netflix launched the Netflix Prize: $1 million for predicting customers' film ratings 10% more accurately than its own system, Cinematch. On 21 September 2009, it awarded the prize to the team BellKor's Pragmatic Chaos.
In matrix factorisation, each user and each item gets an embedding: a short list of learned numbers. Multiply a user's list and an item's list entry by entry and add the results; Google's course says this dot product should approximate that user's feedback on the item.
In August 2009, Yehuda Koren, Robert Bell and Chris Volinsky wrote in IEEE's Computer magazine that the Netflix Prize had shown such models to beat classic nearest-neighbour techniques. Such models, they added, can use implicit feedback, changes over time and confidence levels.
How does a large service narrow a huge catalogue to one feed?
Google's course describes three stages. Candidate generation cuts a huge corpus to a much smaller subset. Scoring ranks those candidates with another model and picks about ten to show. Re-ranking applies extra constraints. In 2016, Paul Covington, Jay Adams and Emre Sargin of Google described YouTube's recommender in two main steps: a deep neural network for candidate generation, then a separate deep ranking model.
A ranker that only shows its current best guesses gets no feedback on anything else. In Aleksandrs Slivkins's textbook on bandit algorithms, exploration means trying different options to gain information, and exploitation means making the best near-term choice on what is known.
Why does the choice of objective matter?
A ranking model predicts outcomes, and someone must choose which outcome counts: the objective. Goodrow says YouTube added watch time to recommendations in 2012 and saw an immediate 20% drop in views. Later it added surveys: only videos a viewer rates four or five stars count towards what YouTube calls valued watchtime.
Training also makes a feedback loop: a recommender learns mostly from how people reacted to what it showed, and its picks shape the next round of data. In preliminary findings of 6 November 2026, the European Commission said TikTok reads repeated watching or liking of content such as weight-loss videos as interest, serving more of the same. It said viewing time outweighs a user's choice to see less, and wants more weight on explicit preferences. TikTok called the findings "fundamentally inaccurate", the Associated Press reported.
Eli Pariser's 2011 book warned that personalisation would give each person a unique universe of information, a filter bubble. A 2022 Reuters Institute review, citing earlier work, describes an echo chamber as a bounded, enclosed media space that can magnify messages and insulate them from rebuttal.
What rules apply to recommender feeds?
- European Union. The Digital Services Act, Regulation (EU) 2022/2065, entered into force on 16 November 2022. Article 27 requires online platforms, except small and micro firms, to explain the main parameters of their recommender systems and any options to change them. Article 38 requires very large platforms, with 45 million or more monthly EU users, to offer at least one option not based on profiling. Both applied to the first designated services from late August 2023, and Article 27 to other platforms from 17 February 2024.
- China. The Cyberspace Administration of China and three other agencies issued rules on algorithmic recommendation, in force from 1 March 2022. In DigiChina's translation, providers must let users switch off algorithmic recommendation or avoid targeting by personal characteristics. They must not lead minors to online addiction, and services that can shape public opinion must file their algorithms.
- United States. In Gonzalez v. Google, the father of a US student killed in an attack in Paris, France, argued that YouTube's algorithms recommended ISIS-related videos. On 18 May 2023, the Supreme Court sent the case back in light of Twitter v. Taamneh and declined to address how Section 230 of the Communications Decency Act applied.
- United Kingdom. Under the Online Safety Act 2023, Ofcom's children's codes say a service that runs a recommender system and poses a medium or high risk of harmful content must configure its algorithms to filter harmful content out of children's feeds, from 25 July 2025.
What do supporters and critics argue about recommender feeds?
- Filter bubbles and polarisation. Supporters of algorithmic feeds cite July 2023 studies led by Andrew Guess in Science and Brendan Nyhan in Nature, from the 2020 US election. Chronological Facebook and Instagram feeds cut time on the platforms but did not significantly change polarisation over three months; cutting like-minded sources by about a third had no measurable effect on eight attitude measures. Critics led by Przemek Grabowicz argued in a 2024 Science eLetter that the chronological study gave a misleading picture of how much untrustworthy news the usual feed shows, because it ran during 63 temporary "break glass" changes Meta made to its feed; its authors stood by their findings.
- Young people's well-being. Critics cite the US Surgeon General's advisory of 23 May 2023, which said social media can pose a profound risk of harm to young people, though it may benefit some. Supporters cite the US National Academies' report of December 2023: there is not enough evidence that social media causes changes in adolescent health at the population level, and recommendation algorithms can give young people important health information or expose them to unscientific treatments.
- Transparency and user choice. Supporters cite the European Commission, which said in April 2023 that users would learn why content is recommended and could opt out of profiling-based recommendations. Critics such as Maximilian Gahntz and Claire Pershan of the Mozilla Foundation argued in February 2023 that "main parameters" is vague and a profiling-or-not choice too coarse.
How does this work with real numbers?
Both examples use illustrative numbers.
How does collaborative filtering predict a missing rating?
Ana has not rated the science quiz, so the system compares her with Ben and Cleo on the four videos all three rated.
Table: Illustrative ratings, 1 to 5 stars
| Video | Ana | Ben | Cleo |
|---|---|---|---|
| Space documentary | 5 | 4 | 1 |
| Cooking show | 1 | 2 | 5 |
| Football highlights | 2 | 1 | 4 |
| Chess tutorial | 4 | 5 | 2 |
| Science quiz | ? | 5 | 2 |
Cosine similarity compares the direction of two lists of ratings: their dot product divided by the product of their lengths. For Ana and Ben, 5 × 4 + 1 × 2 + 2 × 1 + 4 × 5 = 44. Each list here has length √46 ≈ 6.78, so the similarity is 44 ÷ 46 ≈ 0.96. For Ana and Cleo, the products sum to 26: 26 ÷ 46 ≈ 0.57.
The prediction is a similarity-weighted average of their quiz ratings: (0.96 × 5 + 0.57 × 2) ÷ (0.96 + 0.57) = 5.94 ÷ 1.53 ≈ 3.9 stars, nearer Ben's 5.
How does the objective change the order of a feed?
Each candidate's score is (click weight × predicted click chance) + (watch weight × predicted chance of watching to the end).
Table: Illustrative scores
| Video | Click chance | Watch chance | Weights 1 and 0.5 | Weights 1 and 3 |
|---|---|---|---|---|
| A | 0.30 | 0.10 | 0.35 | 0.60 |
| B | 0.20 | 0.15 | 0.275 | 0.65 |
| C | 0.10 | 0.25 | 0.225 | 0.85 |
With weights 1 and 0.5, A scores 0.30 + 0.5 × 0.10 = 0.35 and leads. Raise the watch weight to 3, and C scores 0.10 + 3 × 0.25 = 0.85, so the order flips to C, B, A. The predictions stayed the same; only the objective changed.
Where can the student go deeper in World 101?
This page follows a single feed, step by step, from raw signals to a ranked list. The World 101 subject Recommendation Algorithms takes the wider view of systems that rank and suggest information, and of how they shape what people around the world read, watch and buy online.
Three related subjects go further. The Math Behind Machine Learning covers the vectors, dot products, probability and step-by-step optimisation behind embeddings and ranking models. Social Media & Society looks at how platforms change friendship, identity, family life and public debate. Data Privacy Law explains the rules on collecting personal data and profiling people, which sit behind the DSA's option not based on profiling.
Sources
- Google for Developers, Recommendation Systems course, updated 25 August 2025.
- Y. Koren, R. Bell and C. Volinsky, Computer 42(8), August 2009 (abstract).
- UC San Diego, Netflix Prize release, 21 September 2009; S. E. Page, University of Michigan, 2018 (prize launched October 2006).
- P. Covington, J. Adams and E. Sargin, RecSys 2016, Google Research.
- A. Slivkins, arXiv:1904.07272, 2019, revised 2024.
- C. Goodrow, YouTube blog, 15 September 2021 (clicks, 2011; watch time, 2012).
- Penguin Random House, E. Pariser, The Filter Bubble, May 2011; Reuters Institute review, 19 January 2022.
- European Commission releases, 16 November 2022, 25 April 2023, 16 February 2024 and 6 November 2026; Regulation (EU) 2022/2065; Associated Press, 6 November 2026.
- M. Gahntz and C. Pershan, Mozilla Foundation, 27 February 2023.
- DigiChina, Stanford, translation, 10 January 2022 (provisions dated 31 December 2021).
- Gonzalez v. Google LLC, 18 May 2023, and case preview, Cornell LII.
- Ofcom, 24 April 2025; GOV.UK, 24 July 2025 (Online Safety Act 2023).
- A. Guess and others, Science, and B. Nyhan and others, Nature, July 2023 (2020 US election); EurekAlert, 26 September 2024; Indiana University, 27 September 2024.
- AHRQ, 31 May 2023 (Surgeon General's advisory); National Academies, 13 December 2023.
Common errors and misconceptions
- Misconception: A recommender knows what a person likes. Correct: It predicts from signals such as clicks and watch time; YouTube says a click does not show a video was watched.
- Misconception: Collaborative filtering needs to know what each item is about. Correct: It works from patterns in feedback; content-based filtering is the one that uses item features.
- Misconception: The first item in a feed is the best item. Correct: Its place reflects a score built for a chosen objective; change the weights and the items reorder.
- Misconception: Researchers agree that feeds polarise their users. Correct: Experiments around the 2020 US election found that changing feeds did not significantly change polarisation over three months; critics question one study's findings on untrustworthy news.
- Misconception: EU law bans personalised feeds. Correct: The DSA requires very large platforms to offer at least one option not based on profiling; personalised feeds remain allowed.
Key takeaways
- A recommender ranks items by predicted value to a person, learning from explicit feedback such as ratings and implicit feedback such as watch time.
- Content-based filtering matches item features to past likes; collaborative filtering draws on similar people and struggles with new items, the cold-start problem.
- Matrix factorisation gives each user and item a short list of learned numbers whose dot product predicts a rating; Koren and colleagues wrote that the Netflix Prize showed it beating older methods.
- Large systems pick a short list of candidates, then score and rank them, balancing known favourites against exploring new items.
- The objective, such as clicks or watch time, decides the order of a feed, and feedback loops can keep serving more of the same.
- The European Union, China and the United Kingdom have set rules on recommender feeds; in 2023 the US Supreme Court left the Section 230 question open.
Test yourself
What is the difference between explicit and implicit feedback?
Explicit feedback is a direct statement, such as a star rating; implicit feedback is inferred from behaviour, such as watching or skipping.
Why does collaborative filtering struggle with a brand-new video?
It has no feedback yet, so the model cannot score it; one fallback is the average profile of items in its category.
In the worked example, why is Ana's predicted rating nearer Ben's 5 than Cleo's 2?
Ana's similarity to Ben is about 0.96, against 0.57 for Cleo, so Ben's rating carries more weight.
What changed the order of the three videos in the ranking example?
Only the watch weight, raised from 0.5 to 3; the predicted chances stayed the same, but the order flipped to C, B, A.
What does Article 38 of the Digital Services Act require?
Very large online platforms must offer at least one option for their recommender systems that is not based on profiling.
What did the US Supreme Court decide in Gonzalez v. Google?
It sent the case back in light of Twitter v. Taamneh and declined to address how Section 230 applied.
Write about it
Turn a question from this chapter into a paper.
The Writing Studio has a structure guide, a rubric and citation tools for 15 formats, from essays to research papers. A free account keeps up to five drafts. Learnacy Essential adds feedback from an AI writing coach and submission to The OYI Review.
Publication depends on review and is not guaranteed.
