A recommender system is software that predicts which items a person is likely to want and puts those items in front of them, in order. The items can be films, songs, products, short videos, job listings or advertisements. The software exists because catalogues outgrew human attention long ago: a person scrolling a home screen will look at a handful of options before giving up, and the recommender decides which handful.

The EU's Digital Services Act (DSA) gives the term a legal definition: a "fully or partially automated system" an online platform uses "to suggest in its online interface specific information to recipients of the service or prioritise that information", including the order of search results.

How a recommender works

Three families of technique dominate, and most production systems mix them.

Collaborative filtering ignores what the items are and looks only at behaviour. If two people rated, bought or watched many of the same things, each is likely to enjoy what the other liked and has not yet seen. The variant Amazon described in 2003, item-to-item collaborative filtering, inverts the comparison: it precomputes which products are bought together and recommends neighbours of whatever a shopper is viewing, which scales far better than comparing every customer with every other.

Content-based filtering works from item attributes instead: genre, cast, product category, text, audio features. It copes with new, unrated items (the cold-start problem that defeats pure collaborative filtering) but tends to recommend more of the same.

Hybrid and deep learning models combine both and now carry most of the load. Items and users are converted into embeddings, lists of numbers that place similar things close together in a mathematical space, and neural networks learn which combinations predict a click, a purchase or a long watch.

Large platforms run these models in stages. Google researchers Paul Covington, Jay Adams and Emre Sargin set out the template for YouTube in a 2016 paper: a candidate generation network cuts a corpus of millions of videos to hundreds, then a ranking network scores those hundreds and surfaces a few dozen. The ranking objective was expected watch time per impression rather than click-through rate, because ranking on clicks "tends to promote clickbait", according to the paper. X's open-source code shows the same shape at a social network: retrieval systems capped at 1,000 and 1,200 candidates, 17 filters, then a model that predicts each possible action and multiplies each probability by a published weight, from 20.0 for a copy-link share to -234.0 for a report.

The inputs are mostly implicit signals: watch time, dwell, scroll speed, replays, purchases, skips. Explicit signals such as ratings, follows and "not interested" taps are scarcer. Business rules for diversity, freshness, ad insertion and policy then adjust each list; how the output becomes a scrolling stream is covered in the consumer feed explainer.

Origin and evolution

Xerox PARC's Tapestry system, described by David Goldberg and colleagues in Communications of the ACM in December 1992, filtered email by colleagues' reactions and introduced the phrase "collaborative filtering". GroupLens automated the idea for Usenet news in 1994, and a March 1997 special issue of the same journal, edited by Paul Resnick and Hal Varian, popularised "recommender systems". Amazon engineers Greg Linden, Brent Smith and Jeremy York published their item-to-item method in IEEE Internet Computing in January 2003.

The Netflix Prize turned the field into a public competition. Launched on October 2, 2006, it offered $1 million to any team that could beat Netflix's own Cinematch algorithm by 10% on root mean squared error, a measure of how far predicted star ratings fell from real ones. The training set held 100,480,507 ratings from 480,189 subscribers on 17,770 films. BellKor's Pragmatic Chaos won on September 21, 2009 with a 10.06% improvement; a rival team, The Ensemble, matched its score but had submitted 20 minutes later. The contest popularised matrix factorisation, which decomposes the ratings table into hidden "taste" factors. It also exposed privacy risk: in 2007 researchers re-identified subscribers by matching the "anonymised" ratings against IMDb reviews, and Netflix cancelled a sequel on March 12, 2010 after a lawsuit and Federal Trade Commission (FTC) concerns.

Deep learning took over from 2016. Facebook open-sourced its Deep Learning Recommendation Model (DLRM) on May 31, 2019, built around large embedding tables for identifiers such as user and ad IDs. The latest shift is towards generative recommenders, which read a person's history as a sequence and predict the next action much as a language model predicts the next word. Meta researchers' HSTU paper of February 27, 2024 described a deployed model of 1.5 trillion parameters that improved online A/B test metrics by 12.4%. Meta's ad retrieval engine Andromeda, unveiled on December 2, 2024, narrows tens of millions of candidate ads to thousands before ranking.

Why it matters for marketing

For most platforms the recommender is the product. In a December 2015 paper, Netflix executives Carlos Gomez-Uribe and Neil Hunt wrote that recommendations influence "about 80% of hours streamed" and estimated, without giving a method, that personalisation and recommendation saved the company "more than $1B per year", through reduced cancellations.

The same machinery now decides advertising delivery. Automated campaign products that ask advertisers for broad targeting and many creatives, from Meta's Advantage+ to Google's Performance Max, are recommender systems pointed at ads, which is why hyper-personalization and lookalike modelling increasingly sit inside platform models rather than in advertiser-built audiences. Meta reported that ranking improvements lifted time spent on Instagram Reels by 10% in the first quarter of 2026.

Recommenders also govern organic reach and retail visibility. On Amazon, the Rufus assistant narrows roughly 50 search results to about five named products, according to an April 2026 Workflow Labs analysis.

Regulation

The DSA, published on October 27, 2022, set the first binding rules aimed specifically at recommenders. Article 27 requires every online platform to explain, "in plain and intelligible language", the main parameters of its recommender systems in its terms and conditions, and any options users have to change them. Article 38 goes further for very large online platforms and search engines, those with more than 45 million monthly users in the EU: each recommender must offer "at least one option" not based on profiling as defined in the General Data Protection Regulation (GDPR). Those duties applied to the largest services from August 25, 2023;

Enforcement has moved from disclosure to design. On October 2, 2024 the European Commission asked YouTube, Snapchat and TikTok for the parameters of their recommendation algorithms and their effect on systemic risks. On February 6, 2026 it preliminarily found that TikTok's addictive design, including its "highly personalized recommender system", breached the DSA, and suggested TikTok adapt its recommender. Meta received similar preliminary findings over Instagram and Facebook on July 10, 2026. The European Data Protection Board's guidelines of September 11, 2025 add that platforms must present recommender options equally and cannot keep profiling a user who has chosen the non-profiling option.

Criticisms and disputes

The best-known criticism is the filter bubble, a term coined by activist Eli Pariser in his 2011 book: personalisation narrows what people see until it confirms what they already believe. Evidence is mixed. In a study published in Science on July 27, 2023, researchers moved Facebook and Instagram users to reverse-chronological feeds during the 2020 US election. The users spent substantially less time on the platforms and saw more political and untrustworthy content, but over three months polarisation and political attitudes did not change significantly.

The deeper dispute is about engagement optimisation. Objectives such as watch time and predicted interactions reward whatever holds attention, which regulators now link to compulsive use, especially among minors. In the US, Judge Edward Davila ruled on August 5, 2026 that engagement-based ranking for minors' feeds does not convey an expressive message protected by the First Amendment, in a preliminary decision. MIT economist Daron Acemoglu has argued that removing algorithmic feeds would break TikTok's business.

Transparency remains partial. Article 27 disclosures tend to list signal categories rather than weights, and researchers in Germany have gone to court over access to recommender data. Popularity bias, in which already-popular items win ever more exposure, is a long-standing technical objection.

Not the same as

Search ranking. Search responds to an explicit query; a recommender usually acts without one.

Ad targeting. Targeting defines who is eligible to see an ad. A recommender decides which eligible item each person sees and in what order, and in automated campaigns it increasingly makes the targeting decision itself.

Personalisation. The broader practice of tailoring any experience, from email subject lines to prices. A recommender is one personalisation mechanism; ad personalization is another.

Algorithmic feed. The ranked stream users scroll. The feed is the output; the recommender is the system producing it.

Retrieval-augmented generation. RAG fetches documents to ground a chatbot's answer. It retrieves facts for one answer, not items ranked for a person.

Recent developments

On October 13, 2026, Meta added a "Your algorithm" panel to Facebook where users type topics they want more or less of, with large language models adjusting ranking within the same session. X published its recommendation code in January 2026, describing a Grok-derived transformer that predicts 19 action types. On October 19, Elon Musk said X had added a ranking model predicting viewing time and a retrieval model trained on posts people read all the way through, according to Tom's Guide; X has not said how much weight the new models carry. The same day, Meta engineers described GenRecs Tokenization, which replaces per-ad IDs with shared semantic tokens, reporting a 0.35% conversion lift from internal data. The Commission's EU KIDS Act proposal of September 17, 2026 would require platforms to switch off engagement-driven recommendations for minors by default. It is not yet law.

Timeline

  • December 1992 - Xerox PARC describes Tapestry and introduces "collaborative filtering".
  • 1994 - GroupLens applies automated collaborative filtering to Usenet news.
  • March 1997 - Communications of the ACM special issue popularises "recommender systems".
  • January 2003 - Amazon engineers publish item-to-item collaborative filtering.
  • October 2, 2006 - Netflix launches the $1 million Netflix Prize.
  • 2007 - Narayanan and Shmatikov re-identify Netflix Prize users via IMDb.
  • September 21, 2009 - BellKor's Pragmatic Chaos wins with a 10.06% improvement.
  • March 12, 2010 - Netflix cancels the sequel contest over privacy concerns.
  • 2011 - Eli Pariser publishes The Filter Bubble.
  • 2016 - Google researchers publish YouTube's two-stage deep neural network recommender.
  • May 31, 2019 - Facebook posts the DLRM paper and open-sources the model.
  • October 27, 2022 - The DSA is published, including Articles 27 and 38.
  • July 27, 2023 - Science publishes the chronological-feed experiment on Facebook and Instagram.
  • August 25, 2023 - DSA obligations apply to the largest platforms.
  • February 27, 2024 - Meta researchers publish the HSTU generative recommender paper.
  • October 2, 2024 - Commission requests recommender details from YouTube, Snapchat and TikTok.
  • December 2, 2024 - Meta unveils the Andromeda ad retrieval engine.
  • September 11, 2025 - EDPB adopts guidelines on the DSA and GDPR, covering recommender choice.
  • January 2026 - X publishes its recommendation code.
  • February 6, 2026 - Commission preliminarily finds TikTok's addictive design in breach of the DSA.
  • July 10, 2026 - Commission issues preliminary findings against Instagram and Facebook.
  • August 5, 2026 - US court finds engagement-based feed ranking for minors is not protected speech.
  • September 17, 2026 - Commission proposes the EU KIDS Act with limits on recommenders for minors.
  • October 13, 2026 - Meta adds typed "Your algorithm" requests to Facebook.
  • October 19, 2026 - X adds viewing-time ranking and read-through retrieval models; Meta details GenRecs Tokenization.

Summary

Who: Platforms including Netflix, YouTube, TikTok, Amazon, Meta, X and LinkedIn build and operate recommender systems; advertisers and publishers depend on them for reach; the European Commission and national courts now supervise them.

What: Software that predicts which items a person will want and ranks them, using collaborative filtering, content-based filtering and, increasingly, deep learning and LLM-based models.

When: Collaborative filtering was named in 1992, the Netflix Prize ran from 2006 to 2009, deep learning took over from 2016, and the DSA's recommender rules have applied to the largest platforms since August 25, 2023.

Where: Home screens, feeds, product pages, search results and ad auctions worldwide, with the strictest obligations in the EU, where very large platforms must offer a non-profiling option.

Why: Catalogues are far larger than human attention. Recommenders choose what people see, which drives engagement, sales and ad revenue, and is why regulators now treat their design as a matter of public risk.