A post has to pass three gates to reach a reader

The full pipeline has seven steps, but for anyone writing, only three gates matter. These three are in a multiplicative relationship — fail any one, and the result is zero, no matter how well you do on the other two.

Ranking and visibility are two separate systems. The former decides order, the latter decides whether something appears at all — two services, two inputs, two sets of rules. X lists this as an explicit design decision, and understanding shadowbanning starts with understanding this split.

Your replies can't reach strangers' timelines

Before scoring, there are 17 filters. Most are common sense — your own posts, people you've blocked or muted, posts you've already seen, things matching your muted keywords. But three of them directly decide whether strangers can see your posts, and the first one is especially counterintuitive.

The other two: 48 hours is a hard cutoff — past that, a post is thrown out of the candidate pool. This is a hardcoded constant in the config, not a tunable parameter. Then there's a rule about replies: for a reply to survive, the person being replied to must also be followed by this reader. Replying to your own thread is exempt — if all ancestors are the author's own posts, this filter passes it straight through.

The model predicts which actions you'll take

Older recommender systems assigned each piece of content a "relevance score." This one takes a different approach: the model doesn't guess whether you'll like something — it separately predicts the probability of each specific action you might take on a post. Will you like it? Reply? Copy the link and share it? Report it? Then a separate step blends those probabilities into a single number.

X lists this "blend into one number" step as a separate design decision, and the point is this: prediction belongs to the model, valuation belongs to humans. What this open-source release gives you is precisely the human-decided part.

Two details explain why some accounts seem naturally easier to push. First, a post's likes, replies, reposts, quotes, views, and age are bucketed and fed directly into the model as input features. The model sees these numbers, so early traction genuinely does self-reinforce. Second, each post is compressed into a semantic IDA post's text and images get compressed into a code representing "what this post is about": 6 levels, each selecting one of 256 codes. Posts on the same topic share the same leading levels, like a library call number — books in the same subject area share a prefix. — posts on the same topic share encoding prefixes, so a brand-new post with zero engagement can still be retrieved, as long as its topic matches what this reader regularly consumes. Also, the production retrieval model doesn't store a dedicated embedding vector for each user. Aside from coarse profiles like country and language, a reader is just the sum of what they've interacted with.

Likes are worth 0.5, copy-link shares 20

21 positive weights and 5 negative weights, all listed below.

Two facts jump out immediately: all positive weights add up to 43.324, and "copy link share" alone accounts for nearly half. Meanwhile, a single report has an absolute value more than 5 times the sum of all positive weights combined.

The score gets adjusted four more times before it's final

The weighted sum is just the starting point. Four more tweaks happen before the final score is set — the first happens during the summation, modifying one specific weight; the other three are multiplied against the total score. You can click through all four steps below.

Shadowbanning means being invisible only to people who don't follow you

After ranking, there's one final gate that can remove your post entirely. For each post and each reader, it produces one of three answers: show it, hide it behind a warning (readers can click through), or drop it.

The key is that rules come in two sets. One set applies to everyone. The other has 27 rules, only running when "this post is being recommended to someone who doesn't follow the author," and they can only result in a drop. The same post is shown to followers.

So shadowbanning on X has a very specific shape: your followers see you normally; you simply don't exist in the timelines of anyone who doesn't follow you. You can't detect anything unusual yourself, because what you see comes from the people you follow.

12 account labels, now checkable

The page is at x.com/i/under_the_hood. It shows labels applied to your account and posts over the past month, and lets you download the data. This is a gradual rollout — three conditions must all be true: account at least a year old, at least 10 posts in the past month, and randomly selected into the test group.

These labels aren't all manually assigned. A set of offline batch jobs watches how others respond to your posts — blocks, reports, spam reports — all measured relative to your like count, and uses that to label accounts. So "how many people have blocked or reported you" is a direct input to account-level decisions, not just something about individual posts.

There are 12 account-level labels, grouped into three categories.

These last three judge account status, not your content. But the effect is the same as the other nine: not recommended to non-followers.

Post-level labels have three tiers of effect: the lightest just makes you invisible to non-followers (spam high-recall, suspicious links, adult text, etc.); the middle tier adds a content warning and hides the post from minors, users without an age, and logged-out users; the heaviest tiers stop the post from being shown on X entirely, or restrict its discoverability to the author's profile page, with a "this post is restricted" note shown to everyone.

Two things were explicitly kept closed: the large-model prompts used by the content classifier, and some of the labeling rules. The reason is in the README: to reduce the risk of people gaming the system. X's compensation is precisely this page: you don't get the rules, but you get the results. You can see what these systems output for you, see if any manual labels were applied, and take those label names back to the code to see exactly how they affect you.

Weights shift: mutual-follow boost went from 20 to 15 in three weeks

Before you use any of these numbers, one thing you need to know: they aren't carved in stone. X wrote a dedicated document using July's widely-discussed mutual-follow boost as an example, explaining how to read algorithm updates going forward.

On July 10, an A/B test started. Users were randomly assigned to 5, 10, 15, or 20, with the vast majority at 0. On July 13, early results looked good, and it was rolled out to many users at 20. On July 24, after the full experiment results and real feedback from X, the document's example: during the World Cup, people wanted more football discussion, but most of those posts came from accounts they didn't follow — so the value was pulled back from 20 to 15.

The causal chain is complete: when the boost is too large, mutually-followed content crowds out out-of-network content; when a big event like the World Cup is happening outside that circle, the user experience immediately degrades — hence the five-point pullback.

The document also includes the full code diff from the time, with the intent stated up front: all future algorithm updates will be posted to this repo, and reading the diff shows you exactly what changed. For people who depend on X for a living, when parameters shift you'll see it the same day — no more waiting a month and reverse-engineering from declining metrics.

Based on the three gates, how to adjust your posting

Numbers change, gates don't. With the current parameters, all three gates together reduce to one sentence:

The full pipeline has seven steps, but three gates matter: candidate pool → scoring & ranking → visibility. The three are multiplicative — fail any one, and high scores elsewhere don't matter. Scoring and visibility are also two separate systems: one manages order, the other manages whether a post appears at all.

21 positive and 5 negative weights, all in the config file — readable plaintext.

After ranking, one final gate can remove a post entirely. Rules come in two sets: one applies to everyone; the other has 27 rules that only run when a post is recommended to someone who doesn't follow the author, and they can only result in a drop.