Social Media Algorithms — What You See and Why

Every time you open a social media feed, an algorithm has already decided what you will encounter. The content shown is not a random sample of everything posted by accounts you follow, nor a complete archive of all available posts. It is a ranked selection, assembled in real time from a much larger pool of candidate content, based on predictions about what you are likely to engage with.

Understanding even the general structure of recommendation systems — without needing access to the proprietary code that powers them — provides a more accurate model of how digital information actually flows.

The basic architecture

Most social media recommendation systems operate in roughly similar stages. First, a candidate generation step: from the enormous universe of posts that could potentially be shown, the system narrows down to a manageable set — perhaps several hundred or thousand items — based on broad signals such as who you follow and what you have engaged with recently.

Second, a ranking step: those candidates are scored according to predicted engagement, using a model trained on the historical behaviour of millions of users. The predicted probability of you watching, liking, sharing, or spending time on a piece of content is estimated, and candidates are ranked accordingly.

Third, adjustments: some platforms apply additional rules — diversity filters to prevent the same creator dominating your feed, safety interventions to demote content that violates policy, freshness boosts to surface new content over older posts.

Collaborative filtering

One of the foundational techniques in recommendation is collaborative filtering: the observation that users who have behaved similarly in the past are likely to engage with similar content in the future. If a large group of users who previously engaged with content you have also engaged with then watched a particular video, the system infers that you are likely to want to watch it too — even if you have never interacted with that creator before. This mechanism is responsible for the 'rabbit hole' effect: recommendations that lead progressively further into a niche based on each successive engagement signal.

Time spent as the fundamental signal

Across platforms, the most powerful underlying signal is time: how long users spend on the platform, on individual pieces of content, and on content from particular creators. Platforms are fundamentally attention businesses — their revenue depends on advertising, and advertising revenue depends on audience time. The algorithmic logic follows directly from the business model.¹

Footnotes#

1. The business model is worth noting because it creates a structural incentive that is independent of any intent by platform designers to cause harm. Content that generates engagement — attention, interaction, return visits — is advantaged by the system regardless of whether that engagement is positive or harmful. The platform's interest in user time does not automatically align with the user's interest in accurate or beneficial information. ↩