How the X (Twitter) algorithm works (2026)

The short answer

X ranks the For You feed with a transformer model that predicts how likely you are to do each of about twenty things with a post, such as like, reply, repost, share it by DM or copy its link, follow the author, or hide it. Each prediction is multiplied by a weight and added up. X's open-source repository states it as “Final Score = Σ (weight × P(action))”.

In the weights X published, being shared by copied link, replies, quotes and DM shares count far more than a like, and the negative actions (not interested, mute, block, report) carry large minus weights. So the posts that travel are the ones people want to talk about and pass on, and the ones that lose are the ones people hide.

Where this guide's facts come from

X open-sourced its new ranking system on January 20, 2026 (TechCrunch), describing it as “powered by the same transformer architecture as xAI's Grok model”, and expanded the release on August 13, 2026 with its configuration weights and visibility filters (TechCrunch). The repository says scripts set its defaults to “the primary production values”. This guide uses that code as of November 7, 2026, plus X's own Help Center. Many guides still quote the 2023 Twitter weights; we show those separately and say which numbers are old.

Step 1: which posts get considered

According to the repository README, posts come from two places. In-network: recent posts from accounts you follow. Out-of-network: posts from accounts you don't follow, found by a retrieval model and by topic clusters, plus popular posts. “Both are ranked together by the same model.”

Before ranking, filters remove posts older than 48 hours, posts you've already seen, muted keywords, and posts from accounts you've blocked or muted. X's Help Center describes the two timelines simply: For you “serves posts from accounts and Topics you follow as well as recommended posts”, while Following “displays posts from only the accounts you follow”.

Step 2: what the model predicts, and the weights

The weights below are from home-mixer/params/param.rs in X's repository on November 7, 2026. X notes in the code that the weights multiply the predicted probability of each action, so a large weight on a rare action (like a report) does not mean one report cancels hundreds of likes.

Source: X's open-source ranking code, read November 7, 2026. X says it adjusts the weights over time.

Step 3: the rules applied after scoring

• Author diversity. “Each post after an author's first is multiplied by a decaying factor”, so ten posts in a row from you don't all get the same reach in one person's feed.
• Out-of-network discount. Posts from accounts the viewer doesn't follow are multiplied by a factor below 1, as are replies and reposts from accounts they do follow.
• New-author boost. Posts from authors with few impressions are lifted toward a target position. The parameters show a threshold of 200 impressions and a window of two hours for this boost.
• Time spent matters more now. Release notes in the repository say that from October 2026 ranking “predicts time spent much better” and retrieval is trained to surface “posts people read all the way through”.

How X ranks replies

X's Help Center says: “When ranking a reply higher, we consider factors such as if the original post author has replied, if a reply is from someone you follow, or if the person is a X Premium subscriber.” So replying to your own commenters helps the conversation rank, and Premium affects reply placement. X's public For You code has no Premium reach multiplier, and X gives no number for Premium's effect.

Do links get less reach on X?

Not in X's current published code. “Open a link” is a small positive weight (0.2), and the filter list has no penalty for posts with links. You'll still see the advice to put links in a reply instead; it comes from older behaviour, and no percentage you see quoted for a link penalty appears in X's own sources. Post the link where it reads best.

What gets your posts shown less

X's enforcement page lists how it limits a post that breaks its rules: “Excluding the post from search results, trends, and recommended notifications”, “Removing the post from the For you and Following timelines”, “Restricting the post's discoverability to the author's profile” and “Downranking the post in replies”. Since April 2023 such posts can carry a label, which you can appeal.

Outside the rules, the biggest drag is your own audience's reaction: Not interested, mutes and blocks are among the heaviest weights in the model.

The 2023 Twitter weights (historical)

Many articles still quote Twitter's 2023 heavy-ranker weights from the-algorithm-ml, labelled “current weighting of probabilities (April 5, 2023)”: like 0.5, retweet 1.0, reply 13.5, reply engaged by the author 75.0, good profile click 12.0, negative feedback −74.0, report −369.0. Those numbers are from the old Twitter system and are not what X uses in 2026; the same file says the weights were “periodically adjusted” afterwards.

What to do this week

PersonaXpand schedules your X posts with the rest of your networks and answers replies in your own voice, so conversations keep moving even when you're busy. See also our guides to posting to LinkedIn and X at once and the Instagram algorithm.

• Write for replies and sharing. Questions, strong opinions you can back up, useful lists and threads people want to send to a friend.
• Answer your replies fast, especially from people who follow you back: those replies carry the extra 15.0 boost.
• Don't post ten times in a row. Author diversity discounts each extra post in the same feed; spread posts through the day.
• Make people stay. Time spent and reading to the end now matter more, so give the point in the first line and the detail after it.
• Avoid what makes people hide you: repetitive promos, bait that doesn't pay off, and off-topic posts to an audience that followed you for something else.
• Use links freely when they help the reader.