Recommendation Algorithm: How Social Media Algorithms Rank

Reach on social media is not handed out, it is predicted. Every major network runs a system that estimates, for each combination of user and post, how likely a particular action is: watching, liking, commenting, sharing, swiping away, reporting. Those probabilities produce a score, and the score produces the order of the feed. What marketing teams experience as "the algorithm did not push my post" is the result of exactly that calculation.

The practically most important finding: the platforms differ in their weights, barely in their architecture. Once you have understood the shared mechanics, you learn one algorithm and five deviations from it.

How a recommendation algorithm works in four stages

The underlying problem is a compute problem. Milliseconds pass between someone opening the app and the finished feed, and in that time not every available post can be scored with a large model. So the set is reduced stage by stage and the scoring gets more expensive stage by stage. Each stage has its own knockout criterion, and a post that does not show up in stage one can no longer win anything in stage three.

Candidate Generation. A preselection of a few thousand candidates is drawn from the total inventory of posts. That does not happen by scanning through it but through embeddings: one neural network maps users, a second maps content, both into the same vector space (two-tower architecture). The selection is then a neighbourhood search, technically an approximate nearest neighbour query. This stage decides whether your post makes the shortlist at all. It is also the reason why thematic clarity works better than frequency: an account without a recognisable topic profile ends up in no clean neighbourhood.

Light Ranking. A cheap model roughly pre-sorts the candidate set and cuts it down to a size that can be scored expensively. Few, quickly available features count here: past performance of similar content, recency, rough fit between user profile and post profile. This stage is the reason why a post with a weak start is rarely turned around. It drops out before the precise model ever sees it.

Heavy Ranking. Now the actual prediction model runs. It scores the remaining candidates and predicts several actions at once instead of optimising a single target variable (multi-task prediction). The result per post is a bundle of probabilities.

Re-Ranking. The final stage corrects the pure score order: diversity (not the same sender three times in a row), business rules (share of ads, recency), trust and safety (moderation rules, reach limits for borderline cases). Whatever gets filtered out here does not appear in the feed, no matter how high the ranking score was.

The four stages also explain why diagnoses of reach contradict each other so often. A post can fail at very different points: because it connects to no topic (Candidate Generation), because the first few seconds deliver nothing (Light Ranking), because it gets clicked but not watched to the end (Heavy Ranking), or because a moderation rule kicks in (Re-Ranking). From the outside each of these four scenarios looks identical, namely like poor reach. Anyone who does not separate the cause optimises the wrong stage.

If you come from search or AI, you recognise the pattern immediately: cheap retrieval, expensive rescoring. The same split describes reranking in RAG systems, only with documents instead of posts.

The Value Model: what "good" means to the machine

The bundle of probabilities has to become a single number, otherwise there is no order. The Value Model handles that conversion, a weighted utility function over the predicted actions. The arXiv paper by Milli, Pierson and Garg describes that the aggregation happens as a linear combination of the individual signals, and it points to the ten actions whose weights X disclosed in 2023.

That is the lever platforms actually pull. When a network "changes the algorithm", the pipeline is rarely rebuilt; weights get shifted or a new prediction target gets added. LinkedIn documented this in its ranking paper LiRank: six predictions are linearly combined for the feed, among them a purpose-built long-dwell signal that detects passive reading. Dwell time is one of six targets there and is traded off against the other five; the threshold for it is deliberately context-dependent instead of being defined as a fixed number of seconds.

How much individual actions count is hard to measure from the outside. The analytics platform AuthoredUp, working from its own data set of more than 621,000 LinkedIn posts, arrives at the result that comments count roughly twice as much as likes. That is a third-party analysis, not a figure from LinkedIn, and should be read as a rough indicator accordingly. Considerably higher factors that circulate in advice pieces have not been independently reproduced anywhere.

The honest limit: apart from the code X published in 2023, exact ranking weights are not public, and even that snapshot is partly outdated. Which building blocks can actually be read out of it is covered in detail in the article on the X algorithm and its open-source code.

The mechanism behind it is old and uncomfortable:

"The more popular a post is, the more reach it gets. […] Anyone who understands this principle can exploit it for themselves and their own ends."

Björn Tantau, social media consultant, on LinkedIn. The sentence describes the feedback loop cleanly: early signals raise the predicted value, the raised value brings more delivery, more delivery produces further signals.

From the Social Graph to the Content Graph

For years distribution was a question of connection: you follow someone, you get their posts. That model has not disappeared, but it has moved into the second row. The default case today is interest-based delivery, where the system infers from the behaviour of similar users what might fit.

Instagram makes this split explicit. Adam Mosseri named the three central signals in January 2025 and distinguished between reach among followers and reach among non-followers: Watch Time, likes and sends are the most important signals, with likes counting somewhat more for content shown to existing followers and sends counting more for recommendations to non-followers. Same signals, different weights, two pools of reach. What follows from that in concrete terms for formats and signals on Instagram is covered by the article on the Instagram algorithm 2026.

Strategically the shift means this: your follower count only works at the first stage. It gives the system a cheap first set of candidates and with it fast signals. Whether a post travels beyond that is decided by the quality of those signals, not by the size of the account. In practice every single post competes against the relevant topic pool from scratch instead of inheriting an acquired base reach. In our assessment that is the most important reason why editorial plans built on constant reach per post are regularly wrong.

One side effect concerns republished content. Instagram excludes accounts that post other people's content without their own editing ten or more times in 30 days from the recommendation surfaces. When distribution runs through recommendations, originality becomes a condition of entry.

Foundation models are replacing the model collections

Until recently a ranking stack consisted of a large number of specialised models, one per task and surface. That structure is currently being consolidated. Meta describes GEM as the largest foundation model for recommender systems in the industry, trained across thousands of GPUs at the scale of large language models and 4x more efficient per unit of data and compute than the previous generation; for the initial Reels deployment Meta names 5 per cent more ad conversions on Instagram.

The same movement is under way at LinkedIn. The 360Brew paper describes a decoder-only model with 150 billion parameters that solves more than 30 prediction tasks which would otherwise each get their own model. The widespread claim that thousands of individual models are replaced by it does not appear in the paper in that form.

For practice less follows from this than the headlines suggest, but something concrete does: when a single large model semantically understands what a post is about, formal tricks lose effect and clarity of content gains. Keyword stuffing, hashtag clouds and format imitation address a generation of systems that is being replaced right now.

What the DSA changes in the EU

Recommender systems are regulated in the EU, and that has a direct effect on the availability of reach. Article 27 of the Digital Services Act obliges providers of online platforms that use recommender systems to set out the main parameters of those systems in clear and intelligible language in their terms. Article 38 goes further: very large platforms and very large search engines have to offer at least one option for each of their recommender systems that is not based on profiling. TikTok announced on 4 August 2023 that it would allow European users to switch personalisation off; For You then shows videos that are popular regionally and globally.

The operational consequence is unspectacular but real: a growing share of users can switch to non-personalised feeds. For you that shifts the value towards content that is found without personalisation, meaning searchable material and evergreens rather than pure timing bets.

What is mechanism and what is folklore

It is documented that new posts are first tested on small audiences and that early signal quality decides further delivery. The concrete figures circulating around this are not documented: fixed test pool sizes, exact thresholds for completion rates or a golden hour with defined limits. Those values come from agency posts, not from platform sources. The evidence on reach penalties for external links is equally inconsistent: on X the suppression is documented in the disclosed code, on other networks it is widely reported but not consistently confirmed empirically. Where documented moderation ends and what gets told as secret punishment begins is covered by the article What is a shadowban.

That produces a usable working rule: architectural principles have a long half-life, individual figures a short one. The four-stage pipeline, multi-task prediction and interest-based delivery were valid in 2023 and still are in 2026. Percentage weights, thresholds and format bonuses change constantly and are mostly interpretation anyway. Anyone who aligns content decisions with the principles and uses figures only as a hypothesis in their own tests works more steadily than anyone chasing the next report on weightings. Which platform deviates where and which formats pay into that is set out by the topic overview on social media algorithms and distribution. All figures given here are as of August 2026.