How AI Recommendation Algorithms Decide What You See and Buy Online

Nearly every major online platform now decides what to show each visitor using an AI-driven recommendation system. These systems determine the order of a Netflix homepage, the next video queued on YouTube, and the products listed under "customers also bought" on Amazon. Companies describe the mechanics differently, but the underlying approach follows a similar pattern: collect behavioral data, compare it with data from other users, and rank content or products by predicted relevance.

How the systems work, according to the platforms themselves

Netflix's help center states that its recommendation system evaluates a member's viewing history, how they rated other titles, and patterns from other members with similar tastes — a method known as collaborative filtering. It also factors in content metadata such as genre, actors and release year, plus contextual signals including time of day, device type and preferred language. Netflix explicitly says the system does not factor in demographic information such as age or gender.

YouTube's official "How YouTube Works" site and its blog state that the recommendation system draws on more than 80 billion signals, including watch history, search history, subscriptions, likes, and responses to satisfaction surveys where users rate videos from one to five stars. Only four- and five-star responses count toward what YouTube calls "valued watchtime." The company says recommendations appear mainly on the homepage and in the "Up Next" panel, and that for news and informational topics it aims to promote videos from what it defines as authoritative sources while reducing the visibility of borderline content, even when that content might otherwise generate more clicks.

Amazon's publicly documented recommendation technology, Amazon Personalize, generates recommendations primarily from "item interaction data," meaning records of users clicking, viewing or buying items in a catalog. According to AWS documentation, the service can also incorporate item or user attributes such as genre or price, and supports use cases including "customers who viewed X also viewed," personalized search ranking, and targeted marketing segments.

Why this is drawing regulatory scrutiny

In the European Union, the Digital Services Act imposes specific obligations on recommender systems. Article 27 requires platforms to explain, in plain language, the main parameters their recommender systems use and to let users adjust those settings when multiple options exist, according to legal analysis from the DSA Observatory, a research network tracking the law's implementation. Article 38 goes further for the largest platforms, requiring at least one recommender option that is not based on profiling. The DSA Observatory notes that platforms have interpreted the disclosure rules inconsistently in the absence of detailed regulatory guidance, and that systems optimized mainly for engagement — clicks, likes, shares and watch time — can create risks that go beyond simple consumer annoyance.

In the United States, the Federal Trade Commission published a staff report in September 2024 following a multi-year inquiry into nine social media and video streaming companies. The report concluded that these companies fed user and non-user personal information into automated systems, including algorithms and AI, while giving people little or no ability to opt out of how their data was used. The agency also found inconsistent internal testing of these systems across companies. Then-FTC Chair Lina Khan said the underlying data practices "can endanger people's privacy, threaten their freedoms, and expose them to a host of harms, from identity theft to stalking." The report recommended that Congress pass comprehensive privacy legislation.

What platforms say about data and limits

The open questions

Platforms generally present engagement-based personalization as a convenience that helps users find relevant content faster. Regulators focused on the FTC report and the EU's DSA have framed the same mechanics as a transparency and data-governance issue, pointing to limited user control over how behavioral data is collected and reused, and to inconsistent oversight of the automated systems involved. Both lines of description rely on the same underlying fact: recommendation engines rank content and products using historical behavioral data compared across large user populations, refined continuously as new interactions occur.