daliah
learn · how feeds decide

How social media algorithms work: the five stages the platforms have published

I read what TikTok, Instagram, YouTube, X, Facebook and LinkedIn have put in writing about how their feeds pick posts. Here it is, stage by stage, with every claim tagged by how sure I can be.

by daliah·research by Jean-Paul Azzi·updated 21 Sep 2026·8 min read

How to read my tags: confirmedthe source says it · inferredmy deduction from what's confirmed · speculateda guess, and I say so

the short version

Each platform has published a different piece of the machine

No platform publishes its whole feed. Each has put out pieces: papers, engineering posts, help pages and, for X, code. I read them line by line.

First, the system reads your post and turns it into data

Where platforms describe this step, each post becomes data the system can compare: labels such as topic or safety category, and a long list of numbers that places it near similar posts. confirmed · 10, 14, 15

TikTok’s pages don’t say how it analyses a video’s content. No platform says how much this first read counts once viewers react.

Then millions of posts get cut to a shortlist

Scoring every post for every person would be too slow, so a fast first step builds a shortlist. confirmed · 1, 15

The main way to shortlist is similarity: you and each post become lists of numbers, and the system fetches the posts closest to you. confirmed · 1, 8, 14, 15 LinkedIn already filters here: no posts its safety checks rejected, from people you blocked, in languages you don’t read, or already seen. confirmed · 15

Ranking predicts what you’ll do, then weighs each action

This is what most people mean by “the algorithm”. A heavier model estimates the chance you’ll take each of several actions on each post, and those estimates are combined into one score. confirmed · 4, 7, 8, 9, 14, 16

Why weights matter: you’d probably like one post, then scroll past. You’re less likely to like another, but likely to send it to a friend. If sharing weighs more, the second can win. inferred

Facebook adds its predictions “in a linear way”, counting each by what people say in surveys is meaningful. confirmed · 9 Instagram’s Explore uses a “value model” that subtracts the chance you tap “see less”. confirmed · 8 It says tuning the weights lets it explore trade-offs between engagement metrics. confirmed · 8 LinkedIn combines its predictions linearly. confirmed · 16 X’s code keeps its weights as adjustable settings, “to enable experimentation”. confirmed · 14

So the model learns the chances and the company sets what each action is worth: confirmed for X and Instagram’s Explore. confirmed · 8, 14 For TikTok, YouTube and LinkedIn it’s likely, since each describes combining signals, but unconfirmed. inferred

Only X prints its numbers. In its 2026 defaults, a predicted like is worth 0.5, a reply 5 and a copied link 20. confirmed · 14 The negatives: “not interested” is −43.2, block −31.2, mute −58.8 and report −234. confirmed · 14 These weights multiply predicted chances, not counts, so one report doesn’t literally cancel 468 likes. confirmed · 14 But at equal odds, a negative reaction costs far more than a like earns. confirmed · 14

Instagram, TikTok and YouTube use negative feedback such as “Not interested”, reports or dislikes, but publish no weights for it. confirmed · 2, 3, 7

Last, rules reshuffle the list for variety, freshness and safety

No platform here publishes a time-decay rate or a fixed share of recommended posts.

What people do next becomes training data, sometimes within minutes

TikTok’s pages don’t say how fast its models update. Applying Monolith’s few-minute cycle to TikTok is my inference. inferred

What I’d take from this for your posts

These are my deductions, not instructions from the platforms.

What nobody outside knows, and four beliefs I checked

I watch videos all day and still can’t see TikTok’s weights. Nobody outside can. Every source here is a dated snapshot of systems that keep changing. confirmed · 1, 5, 7 Only X publishes weights, and it can change them. confirmed · 14 TikTok describes steps, not models. confirmed · 4 YouTube’s detailed paper is from 2016. confirmed · 1 Instagram and Facebook give no weights. confirmed · 7, 9 LinkedIn’s boldest ranking model is pre-production. confirmed · 17

Questions people ask

How do social media algorithms work?
The platforms that have published their designs describe the same stages: read each post, shortlist hundreds or thousands of candidates out of millions, score them by the actions you’re predicted to take, apply rules for variety and safety, then learn from what you did. confirmed · 1, 4, 9, 14, 15 What differs is which actions count most and how fast each system learns. confirmed · 1, 6, 7, 14
Do likes matter to the algorithm?
They count, but as one predicted action among many. confirmed · 4, 7, 14 In X’s defaults, a predicted like is worth 0.5, against 5 for a reply and 20 for sharing a link. confirmed · 14 Instagram names likes among its main predictions for Feed, Explore and Reels. confirmed · 7 TikTok says watch time generally weighs most for many users. confirmed · 5
Does follower count affect reach?
It depends on the platform. TikTok said in 2020 that follower count isn’t a direct factor in For You, though accounts with more followers are likely to get more views. confirmed · 3 Instagram lists the creator’s follower count as a Reels signal. confirmed · 7 In its 2026 defaults, X multiplies the score of posts from accounts you don’t follow by 0.75. confirmed · 14
Do platforms test new posts on a small audience first?
Instagram says it does, and it isn’t random: for recommendations, each eligible post first goes to a small audience Instagram thinks will enjoy it, and the best performers go to wider groups. confirmed · 18 TikTok and X rank each viewer’s feed by predicted reactions, as YouTube’s 2016 paper did. confirmed · 1, 4, 14 So a new post’s first viewers are mostly the ones it’s predicted to suit. inferred No platform publishes the size of that first audience or a “golden hour”; reactions do become training data, sometimes within minutes. confirmed · 1, 6, 10
Is the algorithm the same on every platform?
The stages repeat, but what each system rewards differs. confirmed · 1, 7, 14, 15 YouTube’s 2016 ranker aimed at watch time. confirmed · 1 Instagram’s Reels leans on reshares and full watches. confirmed · 7 X’s defaults weight a copied link at 20 and a like at 0.5, and LinkedIn’s shortlist model reads only text. confirmed · 14, 15
Is shadowbanning real?
Reduced reach without removal is documented: Instagram keeps some posts out of recommendations under its Recommendation Guidelines, YouTube demotes borderline videos, and X’s visibility filters include down-ranking. confirmed · 2, 7, 12 Instagram says it doesn’t suppress posts to sell ads, and it shows your recommendation eligibility in Account Status. confirmed · 7

Sources

  1. Covington, P., Adams, J., Sargin, E. (2016). Deep Neural Networks for YouTube Recommendations. Proceedings of the 10th ACM Conference on Recommender Systems (RecSys ’16). Google. doi.org · paper · productionThe shortlist (candidate generation) and ranking models behind YouTube recommendations in 2016; ten years old, so the details are historical.
  2. Goodrow, C. (2021, September 15). On YouTube’s recommendation system. YouTube Official Blog. YouTube. blog.youtube · official page · official statementSignals YouTube uses (clicks, watch time, surveys, shares, likes, dislikes) and demotion of borderline videos; no model details or weights.
  3. TikTok (2020, June 18). How TikTok recommends videos #ForYou. TikTok Newsroom. newsroom.tiktok.com · official page · official statementSignals, the weighting principle, the follower-count statement and variety and eligibility rules as of 2020; no model details.
  4. TikTok (undated; read September 2026). Introduction to the TikTok recommendation system. TikTok Transparency Center. tiktok.com · official page · official statementFive steps: selecting videos, predicting actions, ranking by a combined score, a similarity check and rules; no weights or model details.
  5. TikTok (undated; read September 2026). How TikTok recommends content. TikTok Support. support.tiktok.com · official page · official statementCurrent list of factors per feed; says watch time generally weighs most for many users and that weighting changes over time.
  6. Liu, Z., Zou, L., Zou, X., Wang, C., Zhang, B., Tang, D., Zhu, B., Zhu, Y., Wu, P., Wang, K., Cheng, Y. (2022). Monolith: Real Time Recommendation System With Collisionless Embedding Table. ORSUM workshop at ACM RecSys 2022; arXiv:2209.07663. ByteDance. arxiv.org · paper · productionByteDance’s real-time training system, which the paper says shipped in BytePlus Recommend; it does not name TikTok or describe the For You model.
  7. Mosseri, A. (2023, May 31). Instagram Ranking Explained. Instagram Blog. Instagram (Meta). about.instagram.com · official page · official statementSignals, predicted actions and rules for Feed, Stories, Explore and Reels; names actions but gives no weights.
  8. Vorotilov, V., Shugaepov, I. (2023, August 9). Scaling the Instagram Explore recommendations system. Engineering at Meta. Meta. engineering.fb.com · official page · productionExplore’s four stages and its weighted value model, including a subtracted “see less” prediction; no weight values; Explore only.
  9. Lada, A., Wang, M., Yan, T. (2021, January 26). How machine learning powers Facebook’s News Feed ranking algorithm. Engineering at Meta. Facebook (Meta). engineering.fb.com · official page · productionNews Feed’s ranking passes and the linear combination of predictions, weighted by what people say in surveys; 2021, no weight values.
  10. Meta AI (2023, June 29). The AI behind unconnected content recommendations on Facebook and Instagram. Meta AI Blog. Meta. ai.meta.com · official page · productionContent understanding, the shortlist funnel and minute-level model updates for recommended posts; high level, no numbers on weights.
  11. Naumov, M., Mudigere, D., Shi, H.-J. M., Huang, J., Sundaraman, N., Park, J., et al. (2019). Deep Learning Recommendation Model for Personalization and Recommendation Systems. arXiv:1906.00091. Facebook AI. arxiv.org · paper · researchA published model design with open-source code, tested on public ad-click data; it does not describe any live Facebook or Instagram feed.
  12. Twitter (2023). the-algorithm: X’s Recommendation Algorithm (repository README). GitHub. github.com · open-source code · productionThe services behind the 2023 For You timeline, including candidate sources, rankers and visibility filters; a 2023 snapshot.
  13. Twitter Engineering (2023). Twitter’s Recommendation Algorithm. X Engineering Blog. blog.x.com · official page · productionThe 2023 pipeline: about 1,500 candidates, the in-network and out-of-network split, a continuously trained ranker and final mixing.
  14. X / xAI (2026). x-algorithm: the For You feed algorithm (README and home-mixer/params/param.rs; production default values identical in the 12 August and 18 September 2026 syncs). GitHub. github.com · open-source code · productionCurrent For You pipeline and the default action weights; weights are tunable settings that can change, and some rules are unpublished.
  15. Ramanujam, S. S., Alonso, A., Kataria, S., Dangi, S., Gupta, A., Tiwana, B. S., et al. (2025). Large Scale Retrieval for the LinkedIn Feed using Causal Language Models. arXiv:2510.14223. LinkedIn. arxiv.org · paper · productionHow LinkedIn’s feed shortlists suggested posts from outside your network, tested in a live A/B test; covers the shortlist, not the ranker.
  16. Borisyuk, F., Zhou, M., Song, Q., Zhu, S., Tiwana, B., et al. (2024). LiRank: Industrial Large Scale Ranking Models at LinkedIn. arXiv:2402.06859. LinkedIn. arxiv.org · paper · productionLinkedIn’s production ranking models, including a feed model that predicts several actions and combines them linearly; no weight values.
  17. Firooz, H., Sanjabi, M., et al., 360Brew Team (2025). 360Brew: A Decoder-only Foundation Model for Personalized Ranking and Recommendation. arXiv:2501.16450. LinkedIn. arxiv.org · paper · researchA 150-billion-parameter ranking model that reads profiles and activity as text; its authors call it pre-production.
  18. Instagram for Creators (2024). Helping Creators Find New Audiences. Instagram, 30 Apr 2024. creators.instagram.com · official page · official statementSmall-audience-first ranking of eligible content, originality and aggregator rules; announced as rolling out over the following months.

This is the plain-language version. The full research — every stage written out formally — is Jean-Paul Azzi's, and it's becoming a book. Every claim here points at the platform's own paper, page or code; nothing comes from marketing blogs.

Want me to read your account like this?
give me your @ — I'll watch your videos with all of this in mind. no sign-up, no card.
Continue with Google
sign up with email · already with me? sign in →
By continuing you agree to the terms and privacy policy.
Welcome back.
you're already signed in on this browser — your desk is where you left it.
open daliah →

Want the same honesty about your own videos? That's the free account audit.

keep reading

The other feeds, from their own papers.

12 algorithm myths, checked against what the platforms actually published
Everyone has a theory about the feed. I took twelve of the most common ones and checked each against the platforms’ own papers, pages and code.
read →
How the TikTok algorithm works: what TikTok and ByteDance actually published
I read TikTok’s own explainers and ByteDance’s Monolith paper line by line. Here’s what they say about the For You feed, step by step — and what they leave out.
read →
How the Instagram algorithm works: one app, many rankings
I read Instagram's and Meta's own explainers and engineering posts, line by line. Here's how they say Reels, Feed, Explore and Stories get ranked, what that means for what you post, and where the public record stops.
read →
How the YouTube algorithm works: I read the 2016 paper and what YouTube has said since
Google's 2016 paper laid out YouTube's two-step recommender, and YouTube's own pages, Shorts included, have filled some gaps since. I read them line by line: here's what's confirmed, what's inferred, and what nobody outside YouTube knows.
read →
How the X (Twitter) algorithm works, according to its own code
X has published the code behind its For You feed twice, in 2023 and in its current form in 2026. I read it: here's how posts get ranked, and what the code can't tell you.
read →