How the Facebook algorithm actually ranks your posts
Meta publishes the ranking pipeline, the prediction targets and the demotion list. Most guides ignore all three and guess instead. Here is the documented version.
8 minute read
Meta writes the outline of the pipeline down
Almost everything published about the Facebook algorithm is reverse-engineered guesswork, which is strange, because Meta describes the structure itself. Its Transparency Center sets out Feed ranking in four stages. Inventory is everything you could possibly be shown right now: posts from friends, from Pages and Groups you follow, and recommended content from accounts you do not follow. Signals are the data points read about each candidate post and about you, and Meta says there are thousands of them — explicit ones like likes, comments and shares, alongside inferred ones. Predictions turn those signals into probabilities, and Meta states directly that the Feed ranking system runs over 100 different prediction models. The last stage collapses all of it into one relevance score per post, and your Feed is that list, sorted.
What the system is actually predicting about you
The genuinely useful detail is which predictions Meta names, because that is the closest thing to a published objective. Its Feed ranking page lists how likely you are to comment on a post, how likely you are to watch a video inside it, and the predicted number of additional comments or replies a post will collect if you comment on it. Sitting in the same list is a prediction of whether the content is problematic and should receive reduced distribution. Two things follow from reading that carefully. The system scores a post against one specific person, so being good is not a property your post carries around with it. And demotion is not a separate penalty applied afterwards — it is one of the predictions made during ranking.
Your reach problem is an inventory problem
Meta's Widely Viewed Content Report is the least-quoted useful document it publishes. For the second quarter of 2026 it reports that 49.6% of Feed content views in the US came from connected sources — friends, Groups people had joined, accounts they follow — with 36.6% from unconnected recommendations and 13.8% from other sources. The previous quarter was close: 48.1% connected, 39.0% unconnected. So roughly half of all Feed viewing goes to content people deliberately connected to, and your Page competes for that half against every friend, every Group and every other Page each of your followers has. Organic Page reach in the low single digits is not a bug anyone can fix for you. Followers get you into the inventory stage, not into the ranking.
The demotion list is short, specific, and mostly ignored
Meta's Transparency Center has a page titled Types of content we demote, and its brevity is the story. There are four entries, grouped under three stated principles: responding to people's direct feedback, incentivizing creators to invest in high-quality and accurate content, and fostering a safer community. The four are clickbait links, engagement bait, fact-checked misinformation, and content likely violating Community Standards without quite crossing the line. That is the documented list in full. It is far shorter than the folklore lists of algorithm penalties that circulate, and what is absent is as instructive as what is present — no entry for posting too often, for hashtags, for editing a post, or for outbound links as a category. Meta adds that reduction size varies with repeat violations and prediction confidence.
Reels are ranked by a different system, with different targets
Feed and Reels share the same four-stage shape but not the same objectives, and Meta documents the Reels system on its own page. It describes roughly ten prediction models running together, and the list of predicted actions is revealing: how long you will watch, whether you will keep watching in full screen, whether you will click, like, skip or complete the reel, whether you will hide it or ask to see less content like it, whether it starts a viewing session, and whether you move on to the next one. Close to half of those are negative outcomes being predicted and priced in. This is why the same clip can sink in Feed and travel as a Reel — it is not the same competition.
Meta moved Reels ranking off pure watch time in January 2026
The most concrete recent change is documented on Meta's engineering blog rather than in any marketing post. In January 2026 the Facebook Reels team published its reasoning that likes, shares and watch time are noisy proxies for what people actually value, noting their prior interest heuristic reached only 48.3% precision. Their fix was to ask. A random share of viewing sessions gets a single daily question — to what extent does this video match your interests, scored one to five — and a model called UTIS is trained on those answers and layered over the existing ranking predictions. Offline, precision rose from 48.3% to 63.2% and accuracy from 59.5% to 71.5%. In testing across more than 10 million users, total engagement rose 5.2% and low survey ratings fell 6.84%.
Originality became an enforcement mechanic, not a virtue
In July 2025 Meta announced enforcement against unoriginal content on Facebook, and the numbers it gave are the part worth holding on to: action taken against roughly 500,000 accounts in the first half of 2025 for spammy behaviour, with penalties ranging from reduced distribution to loss of monetisation access, and around 10 million profiles removed for impersonating large creators. The ranking mechanic underneath it is specific. When Facebook detects duplicate videos it reduces distribution of the copies so views flow to the original, and it has been testing attribution links on duplicates that point back to the source. Reposting is not a neutral act with a neutral outcome; the copy is actively held back. Adding real commentary, analysis or editing is what keeps a post on the right side of it.
What the documented mechanics actually ask of you
Strip it back and the ask is narrow. Publish video people finish rather than merely start, because completion and continued watching are named predictions. Write posts that produce real replies, because reply volume is predicted explicitly — and note that engagement bait is the documented opposite of this, not a shortcut to it. Keep links out of the main rotation: Meta's own Q2 2026 figure is that just 1.1% of Feed content views were posts containing an external link. Post originals. None of this is hard to understand; it is hard to sustain, because original video every week is a supply problem before it is a strategy problem. Recording once and cutting a week of vertical clips is how people keep it going — FrameOS does that pass, 300 credits for 3 days · no card.
FAQ
How does the Facebook algorithm work in 2026?
Meta's Transparency Center describes four stages: inventory, signals, predictions and a relevance score. It gathers every post you could see, reads thousands of signals about you and each post, runs over 100 prediction models to estimate how you will respond, then sorts by the combined score. Reels are ranked by a separate system with around ten prediction models of its own.
What content does Facebook demote?
Meta's published list has four entries: clickbait links, engagement bait, fact-checked misinformation, and content likely violating Community Standards without breaking them. Separately, duplicated videos have the distribution of the copy reduced so the original gets the views. Meta says the scale of a reduction depends on prior violations and how confident its prediction is.
Why is Facebook organic reach so low?
Because ranking is a competition for a fixed slot count. Meta's Widely Viewed Content Report puts connected sources — friends, Groups and followed accounts combined — at 49.6% of US Feed content views in Q2 2026. Your Page competes for a share of that half against everything else each follower has connected to, which is why follower count does not translate into views.
Do Facebook Reels use the same algorithm as Feed?
No. Meta documents Reels as its own AI system predicting reel-specific actions: watch duration, full-screen continuation, skips, completions, hides and requests to see less. In January 2026 Meta added a survey-trained model that asks viewers directly how well a video matches their interests, which shifts ranking away from watch time alone.
Sources
- Our Approach to Facebook Feed Ranking · Meta Transparency Center
- Types of content we demote · Meta Transparency Center
- Facebook Reels AI system · Meta Transparency Center
- Widely Viewed Content Report · Meta Transparency Center
- Adapting the Facebook Reels RecSys AI Model Based on User Feedback · Engineering at Meta
Related pages
Keep reading
More Creator Growth →How to use Twitter (X) in 2026
Most guides to this platform describe features that no longer exist. Communities shut down in May 2026, DMs became Chat, and the home screen now has more than two tabs.
9 minute readHow the X (Twitter) algorithm actually works
X is the only large platform that publishes the code ranking its main feed. It also had to add a comment to that code correcting the most-repeated claim about it.
9 minute readHow to use Pinterest: the setup, the formats and the search mindset
Pinterest reported 640 million monthly active users in its Q2 2026 results, and almost none of them are scrolling a friend feed — they are searching. That changes what you publish, how you title it, and how long it keeps working.
8 minute readTurn one long video into a week of posts
300 credits for 3 days · no card
Start free