How the Facebook Algorithm Decides What Gets Seen?

Most explanations of Facebook ranking start with someone's theory. Meta's own Transparency Center starts with a process, and it is worth using that process as the map instead of guessing at one.

The first stage is inventory. This is every post you could possibly see right now: posts from friends, from Pages and groups you follow, from accounts Facebook thinks you might like. Nothing has been ordered yet. Your post enters this pool the moment you publish, alongside everything else published by everyone your audience is connected to.

The second stage is signals. Meta describes thousands of them, grouped into recognizable categories: who posted the content, when it was posted, what type it is (photo, video, Live, or link), your ties to the author, and your own behavior. Two of them are easy to overlook — how long you spend looking at a photo, and how long you spend reading comments. Content-quality flags, such as whether something has been marked false, sit in this stage too.

The third stage is predictions. Facebook does not score the post directly on those raw signals. It uses them to make personalized guesses about you specifically: how likely you are to comment, how likely your friends are to comment if you share it, how likely the post is to start a conversation at all. Meta says more than 100 models run at the same time to produce these predictions.

The fourth stage is the relevance score. Every prediction collapses into a single number per post, per person, and that number decides the order. Meta says the score is computed for roughly 500 posts each time the ranking runs. Your post is not competing with itself over time. It is competing with several hundred other candidates for one person's attention, refreshed constantly.

Here is the part that separates an honest explanation from a confident one. Meta names the stages. Meta names the signal categories. Meta does not publish weights.

There is no public document stating that a comment is worth a specific multiple of a like, or that a share carries a fixed numeric value, or that dwell time counts for a set percentage of the relevance score. Anyone showing you a table of Facebook ranking coefficients built it themselves.

The confusion usually traces back to a real event on a different platform. In 2023, X (formerly Twitter) open-sourced part of its recommendation code, and that release did contain concrete numeric weights. Those numbers spread quickly through social media blogs, and somewhere along the way the platform name fell off. They describe X's system. They have nothing to do with Facebook, and importing them will send you optimizing for arithmetic that does not exist here.

So when someone asks how does facebook algorithm work in exact numbers, the accurate answer is that the mechanics are documented and the math is not. You can work with the mechanics. Predictions are about interaction and conversation, so posts built to be responded to give the models something to predict. Signals include content type and time spent, so a photo that holds a look and a comment thread worth reading feed the system real input. That is as specific as the public record allows anyone to be.

The four stages explain something Page owners feel long before they can name it. A typical Page post reaches roughly 2 to 5 percent of that Page's followers, according to commonly reported industry benchmarks. Followers are not an audience you can address. They are a list of people who might qualify for a seat in someone else's inventory pool.

The decline is well documented across benchmark reporting over the years: around 16 percent of followers reached in 2012, around 6 percent by 2014, 5.2 percent in 2020, and roughly 2.2 percent in 2025. Nothing dramatic caused that curve. Inventory grew faster than attention, relevance scoring got better at predicting individual interest, and the space in the Feed stayed finite.

This is the mechanical reason a Page can post consistently and still feel invisible, and it is why so many owners start wondering whether the algorithm is working against you rather than simply ranking against a bigger pool. It is ranking. Every new account, group, and creator your followers connect to adds candidates to the same few hundred posts that get scored for a place in one person's Feed.

Ranking reads interaction more broadly than most creators track it. Reactions across all six types, comments, shares, and clicks all register, and clicks include link taps, photo expands, and other quiet actions that never appear as a visible number on the post. Those actions are part of the behavioral evidence the prediction stage runs on, which is why a post can travel further than its like count suggests.

Shares deserve separate attention because of what they do structurally. A share does not just add a count. It creates a new post in a different person's network, which enters that network's inventory and gets scored on its own. This is the mechanism behind the question of whether shares still multiply organic reach, and the answer sits in the stage model: a share is a new entry point into a pool you have no other access to.

That structural role is why added shares can help a post travel past your own followers and into feeds where your Page has no existing connection. It works best on a post that already holds attention, since every one of those new placements still gets scored on its own merits by the same prediction models.

Sudden reach drops rarely mean a penalty. Four mechanical explanations cover most cases.

Inventory shifted. If your audience followed a batch of new accounts, or a major event flooded their Feed, the candidate pool got more crowded and your relevance score has to clear a higher bar for the same placement.

Content type changed. Photo, video, Live, and link are separate signals, and switching formats changes which behavioral history the models draw on. A Page known for photos that suddenly posts links is being predicted with less relevant evidence.

Your returning viewers stopped returning. How often someone interacts with an author is an explicit signal. A few weeks of quiet, or a run of posts that did not land, weakens the tie that was carrying your distribution.

Or you followed a format that stopped working. Engagement-bait phrasing, reposted trend formats, and mass-tagging all had their moment before prediction models learned to discount them, which is how trends that quietly kill reach do their damage — they look active while the underlying signals hollow out.

The diagnostic question is not what you did wrong. It is which stage changed: the pool, the signals, or the ties.

Once the stages are clear, the practical question shifts from how does facebook algorithm work to what the models can actually read, because most of the work is giving them something legible to predict.

Post types your audience already has history with. The models lean on evidence, and evidence means past behavior with your specific content in that specific format.

Write openings that earn the caption expand, because expands and time spent on a photo are named signals rather than vanity numbers. Ask questions that require an actual answer instead of a reaction, since prediction targets include how likely a post is to start a conversation. Then reply in the thread while the post is still being distributed, because comment reading time is itself a signal.

Give people a reason to come back. Returning viewers strengthen the author-interaction signal that carries your next post.

Paid distribution fits the same logic rather than bypassing it. Promotion buys placements; the relevance model still decides what happens after the impression. A broader Facebook growth setup works when the paid layer amplifies posts that already hold attention, so the extra placements feed the models good signals instead of expensive indifference.

None of this is a hack, and none of it requires a weights table that does not exist. It is a matter of reading the documented process and posting in a way it can actually score.

No. Meta documents the four ranking stages and names the signal categories, but it has never published numeric weights for them. The specific coefficients circulating online come from X's 2023 open-source release, which describes a completely different platform.

Not by a published multiplier, because no such multiplier exists publicly. What Meta does confirm is that its prediction models estimate how likely you are to comment and how likely a post is to start a conversation, and that time spent reading comments is a ranking signal, so conversation feeds the system more to work with than a single tap does.

Usually because one of the ranking inputs changed rather than because of a penalty. Check whether the candidate pool got more crowded, whether you switched content formats, whether the people who normally interact have gone quiet, or whether you adopted a format the models have learned to discount.