How the X (Twitter) Algorithm Ranks Posts

By Oliver Hayes · Updated October 15, 2026 · 6 min read

X is unusual among social platforms because it publishes the code behind its main feed. That means you can base your strategy on what the system actually does rather than on rumours. This guide explains the current set-up in plain English and what it means for your posts.

For You vs Following: which feed does the algorithm control?

The ranking system described here powers the For You tab, the default feed on X and the one where most reach happens. The Following tab is limited to accounts you follow. If your posts only ever reach your existing followers, you are mostly missing out on For You distribution.

How does the X algorithm choose posts?

In January 2026 X open-sourced a rebuilt For You system, and its public repository has since been updated with the production scoring weights (August 2026) and transparency tools (September 2026). Every time you open the feed, the system works through these stages:

1. Understanding you: it loads your recent engagement history, who you follow, who you block or mute, your muted keywords and what you have already been shown.
2. Finding candidates: in-network posts come from a service that keeps recent posts from accounts you follow in memory. Out-of-network posts come from a retrieval model that matches you with posts similar to what you engage with, from clusters of accounts with overlapping audiences and from a shared list of popular recent posts.
3. Filtering: duplicates, posts older than 48 hours, your own posts, posts from blocked or muted accounts, muted keywords and posts you have already seen are removed.
4. Scoring: a transformer model called Phoenix predicts the probability of each action you might take on each post.
5. Selection and safety checks: the top-scoring posts are kept, a separate visibility-filtering system can drop posts based on labels such as spam, and conversations are de-duplicated.
6. Blending: ads, “Who to follow” suggestions and prompts are slotted in around the ranked posts.

This is a big change from the version Twitter open-sourced in 2023, which relied on many hand-engineered features and a ranking model of roughly 48 million parameters; the new system ranks posts end to end with a transformer instead.

Which engagement signals matter most?

The model predicts around 25 actions – from likes and replies to dwell time, profile visits, follows, mutes and reports – and multiplies each prediction by a weight. Positive actions add to a post’s score; negative ones subtract. Here is how the main published weights compare with the 2023 version.

Two cautions from X’s own documentation. First, weights multiply predicted probabilities, not raw counts, so it is wrong to read the table as “one report cancels out hundreds of likes”. Reports are rare, so their weight has to be large for the prediction to matter at all. Second, the probabilities are personal: they depend mostly on each viewer’s own behaviour, which is why the same post can rank very differently for two people.

Replies between mutual followers

In July 2026 X tested and then rolled out a boost to the reply weight for original posts from people you mutually follow. It launched at 20 for many users on 13 July and was set to 15 on 24 July. In practice, conversations between accounts that follow each other carry extra weight.

What else changes a post’s ranking?

After scoring, three adjustments apply:

• Author diversity: each additional post from the same author in your feed is scored down by a decaying factor, so posting ten times in an hour does not mean ten slots.
• Out-of-network discount: posts from accounts you don’t follow, and replies or reposts from accounts you do follow, are multiplied by a factor below one.
• New-author boost: posts from authors with very few impressions are lifted towards a target position, giving small accounts a chance to be tested.

Freshness matters too. With a 48-hour cut-off, a post has a short window to collect early signals – our guide to why the first hour after posting matters explains how to use it. Separately, the code includes a filter that, for new accounts, holds back out-of-network posts below an engagement threshold.

Do Premium, links and visibility labels affect reach?

Premium: we found no Premium multiplier in the published scoring weights. What X does advertise is reply prioritisation – “larger” for Premium and “largest” for Premium+ – which affects where your replies appear in conversations rather than how the For You model scores your posts.

External links: the published weights do not include a penalty for posting a link; opening a link is actually a small positive signal. Whether link posts lose out because people leave X instead of replying or sharing is widely debated, but X has not published data that settles it, so test it on your own account rather than relying on rules of thumb.

Visibility labels: ranking and visibility are separate systems. Labels for things like spam or inauthentic behaviour can stop a post being shown at all, and X is piloting a tool called “Under the Hood” that shows the visibility-limiting labels on your own account and posts. If you suspect a hidden restriction, also see how to tell if you are shadowbanned.

How do you use this to grow on X?

• Write for replies and shares, not likes. Likes carry one of the smallest positive weights; replies, quotes and shares carry far more.
• Give people a reason to stop scrolling. Dwell time and clicks are predicted signals, so posts worth reading in full help.
• Avoid triggering “not interested”, mutes and blocks. Engagement bait that annoys people costs you more than it earns.
• Build real mutual relationships. The mutual-follow reply boost rewards conversations with people you actually talk to.
• Post when your audience is online and spread posts out rather than bunching them.

For a longer playbook, read our guide to growing a Twitter (X) account; our deep dive on bookmarks as a quiet ranking signal covers one more signal creators watch, and how X counts views explains the number most people judge their reach by.

Sources: X For You feed algorithm (GitHub), Twitter heavy ranker weights, 2023 (GitHub), Twitter Engineering: recommendation algorithm, X Help Centre: X Premium