How the Facebook algorithm works in 2026: what Meta has actually published

Meta has published a lot about Facebook ranking. Much of it is old.

The clearest description of Facebook's Feed ranking I found is from January 2021. Meta engineers Akos Lada, Meihong Wang and Tak Yan explained how the production system collected, scored and ordered posts at the time. confirmed · 1 That's more than five years ago. Treat the details as a snapshot.

In June 2023, Meta published 22 “system cards”, 14 of them for Facebook, including Feed, Feed Recommendations, Reels, Stories and Video. confirmed · 3 Each card lists the ten most important prediction models, “rather than everything in the system.” Meta also says those models and signals “change frequently over time.” confirmed · 3 The same day, Meta's AI team explained how posts from accounts you don't follow get found and ranked. confirmed · 2

The newer, concrete sources are Meta's 2026 engineering posts on Facebook Reels confirmed · 5, 6, plus its newsroom and creator posts, from 2017 to 2026, on what gets less distribution. confirmed · 9, 10, 11, 12, 13 In 2024, the head of Facebook said Reels and Feed ranking had been upgraded with “a new model architecture,” first piloted on Facebook Reels. That model wasn't described in detail. confirmed · 20

Research isn't production. You'll see Meta's 2019 DLRM paper cited as “the Facebook algorithm.” It's a research paper and an open-source benchmark: a type of model for predicting clicks, tested on a public ad-click dataset. It never says it ranks your Feed. confirmed · 23 I worked from Meta's dated posts in the source list, not from the current text of the live system cards.

How your Facebook Feed gets built, step by step

1Gather what you could see. The pool starts with every post shared since your last visit by a friend, Group or Page you're connected to. Posts you missed in earlier sessions come back, and a post you already saw can return if your friends start talking about it. confirmed · 1 Recommended posts come from separate retrieval systems. They narrow billions of pieces of content to thousands, then to a few hundred, in hundredths of a second. confirmed · 2

2Understand the post. For recommendations, Meta's models read a post's images, text, audio and video to work out what it's about. confirmed · 2 A system Meta calls Meta Interest Learner matches new content to likely audiences “even when there are very few engagements.” confirmed · 2

3Run integrity checks. Integrity processes run on every post. confirmed · 1 Some posts stay up but get reduced distribution. confirmed · 16 Some content that's allowed on Facebook still isn't shown in places where Meta recommends content. confirmed · 15

4Predict and score. A light model picks about 500 of the most relevant posts. Then a main pass scores each one with neural networks that predict several actions, such as liking, commenting and sharing. confirmed · 1 Those predictions get multiplied by weights and added into one number. The weights differ by person: if you rarely comment, a comment prediction counts for little in your Feed. confirmed · 1 Meta names sharing as one factor confirmed · 4, and its signals include how many likes, comments and views a post has, and how long you watch. confirmed · 3 Meta also runs surveys: was this post “worth your time”, which friends are you closest to, how interested are you in a Page you follow. The answers go into ranking. confirmed · 1, 14 And Meta said that from 16 December 2025 it would use your chats with its AI to personalize the posts and reels you see. confirmed · 22

5Mix the Feed. A final pass adds context, such as content-type diversity rules, so your Feed isn't all one kind of post. confirmed · 1

6Learn from what happens. Watching a recommended video for a few seconds and clicking away tells the system you're not interested. Finishing a video, liking, hiding or snoozing, and “Show more / Show less” all shape what comes next. confirmed · 2, 4 Meta says these models can be updated with “minute-level freshness.” confirmed · 2

You can step outside the ranking. The Feeds tab shows posts without algorithmic ranking. confirmed · 4 Since March 2025, the Friends tab in the US and Canada shows only your friends' content, with “no recommended content.” confirmed · 18

Reels work the same way, plus one direct question to viewers

Facebook Reels follow the same pattern: retrieval, then early- and late-stage ranking with models that predict several outcomes at once. confirmed · 6 In January 2026, Meta's Reels team wrote that engagement signals like likes, shares and watch time “can be noisy.” confirmed · 5 So Meta asks a sample of viewers how well a video matches their interests, on a 1-to-5 scale, and trains a model on the answers. That score now feeds the final ranking formula, and it reshapes retrieval too. confirmed · 5 Meta reports more niche, high-quality videos and fewer generic, popularity-based recommendations as a result. confirmed · 5

Friends count on Reels too. “Friend bubbles” show which friends liked a reel. Behind them, a closeness model picks the friends whose activity matters to you, and retrieval pulls in videos those friends interacted with. confirmed · 6 In October 2025, Meta said its engine was surfacing 50% more reels published that day, and that “Not Interested” taps and saves change your recommendations. confirmed · 7

What Meta says it pushes down

• Engagement bait. Posts that push people to react just for reach, like “LIKE this if you're an Aries!”, get demoted. Pages that do it “systematically and repeatedly” get stricter demotions. Real requests for help, like a missing child report, aren't affected. (2017, updated 2019) confirmed · 12
• Reposted content. Accounts that keep reusing other people's videos, photos or text lose monetization for a while and get “reduced distribution on everything they share.” Duplicate copies of a video are shown less. (July 2025) confirmed · 9
• Minor edits and empty reactions. Re-uploads with borders, added captions or a speed change count as unoriginal, and so do reactions that only watch along or stitch clips “without adding anything meaningful.” Both get deprioritized. Accounts that keep posting mostly unoriginal content may become non-recommendable and demonetized. (March 2026) confirmed · 10
• Spammy captions. Long, distracting captions with an “inordinate amount of hashtags,” or captions unrelated to the post, get an account's content shown only to its followers, with no monetization. (April 2025) confirmed · 11
• Fake engagement. Comments Meta detects as coordinated fake engagement “will be seen less.” confirmed · 11
• Low-quality links. Posts linking to pages with little substantive content and disruptive, shocking or malicious ads get less distribution. (May 2017) confirmed · 13

It works the other way too. In January 2025, Meta said civic content from people and Pages you follow would be ranked more like any other content, based on signals such as likes and views. confirmed · 19

What this means for what you post

• Write for someone who's never heard of you. Reaching non-followers goes through recommendation, which has extra eligibility rules. confirmed · 15 Staying recommendable is how you reach new people. inferred
• Aim for what a real viewer would do: share it, write a comment worth writing, watch to the end. Weights differ person to person, so no single action is a sure lever. Something people would pass to a friend helps across many viewers. inferred
• Make it yours. Meta says content you filmed or produced gets the most distribution, while borders, speed changes and watch-along reactions on others' clips get deprioritized. confirmed · 8, 10 If you use a clip, add real information or analysis. inferred
• Keep captions short and relevant. Meta's creator guidance says captions with no links, few capital letters and no more than five hashtags perform best. confirmed · 9 A wall of hashtags can limit you to followers only. confirmed · 11
• Don't ask for likes, tags or shares. Ask a real question. Engagement bait is one of Meta's oldest named demotions, so it's an easy loss to avoid. inferred
• When reach drops, open post-level insights in the Professional Dashboard. Meta says they can show why a post may not be getting distributed. confirmed · 9 Rule out the named penalties before you change your whole strategy. inferred

What nobody outside Meta knows, and seven beliefs checked

Meta has published the shape of the scoring formula, not its numbers. confirmed · 1 None of the sources I read gives the weight of a like, a comment or a share, and Meta says its system cards show only the top predictions. confirmed · 3 I watch videos all day, and I still can't see those weights. Nobody outside Meta can. Today's share of recommended posts in Feed is unknown too: the last public figure, “more than 20 percent” across Facebook and Instagram feeds, is from 2023. confirmed · 2 And no one outside can see how the live models have changed since.

• “Links are punished.” No evidence of a blanket penalty. Meta demotes links to low-quality pages confirmed · 13 and says captions with no links perform best for creators. confirmed · 9 No published paper or official page I found says every link post is demoted. In Meta's first US report, from 2021, fewer than 13% of Feed content views were on posts with links. confirmed · 17 That describes how people use Feed, not a rule. inferred
• “Posting too often hurts reach.” No evidence. None of the Meta sources I read describes a frequency penalty. The Feed does apply diversity rules confirmed · 1, so several of your posts compete for the same person's attention. inferred
• “Facebook shows new posts to 5% of your followers first.” No published paper or official page confirms this. Meta describes matching new content to likely audiences by interest, even with few engagements. confirmed · 2 For Instagram, read the separate breakdown (see the Instagram page).
• “Page reach is dead.” Not supported, and not disproved for today. Meta calls followers your “core audience.” confirmed · 8 In that 2021 US report, 57% of the posts people saw came from friends and family. confirmed · 17 In 2026, Meta offered some creators “increased reach on eligible reels.” confirmed · 21 Current Page reach figures aren't published.
• “Shares matter more than likes.” Plausible, not proven. Sharing is a named prediction confirmed · 4, and Meta tells creators to make content people want to share. confirmed · 8 But the weights differ per person confirmed · 1, and no source publishes them.
• “Watch time is all that matters on Reels.” Contradicted. Watch time is a signal confirmed · 3, but Meta's Reels team calls engagement signals noisy and added a survey-based interest score to the final formula. confirmed · 5
• “Engagement bait and reposts get pushed down.” Confirmed. confirmed · 9, 10, 12