How the X algorithm actually works

The short answer

X’s For You feed gathers candidate posts, about half from accounts the reader follows and half from accounts they do not, then scores each with a model that predicts how the reader will engage with it: like, repost, reply, open the conversation, visit the author, or react badly. The scores are weighted and added up, and the highest-scoring posts are shown. The Following tab skips all this and shows followed accounts newest first.

When X, then Twitter, published this code in 2023, the weights made the priorities plain: a reply counted 27 times a like, a reply the author answered counted 150 times a like, and a report counted hundreds of times against a post. X says it has adjusted the weights since, and Sprout Social’s 2026 guide reports that X has worked its Grok AI into ranking. The shape still points the same way: conversation counts, approval counts little, and annoying people costs the most.

What the open-sourced code weighed

These are the weights X published for its For You ranking model, as of 5 April 2023. Each is multiplied by the predicted chance of that action and the results are summed. X notes the production model may differ slightly and that the weights have been adjusted since.

Why a like counts so little

The weights are not a ranking of how much X values each action. X explained that it first set them so that, on average, each kind of engagement contributes about the same to a post’s score. Likes are common and replies are rare, so a reply needs a far larger weight to matter as much. Read the table as a measure of how hard each reaction is to earn, and how strongly earning it moves a post.

What that means for a post

Read the weights as instructions:

• Write to be answered. A reply was worth 27 likes, so a post that invites a response (a question, a clear view, an unfinished thought) beats one people only approve of.
• Answer your replies. A reply the author engages with carried the biggest positive weight of all. The first hour of a post is for talking to the people who replied.
• Give people a reason to open the thread. Opening the conversation and staying, or visiting your profile and engaging, each counted more than 20 likes.
• Do not provoke the mute button. A single “show less often” cancelled about 148 likes, and a report far more. Bait that annoys people costs more than it earns.

Half your reach comes from strangers

The code shows For You drawing about half its posts from accounts the reader follows, found and ranked through X’s search index, and the rest from accounts they do not follow, found through communities of people with similar interests and through what similar users engage with. A post that sparks conversation among people interested in a topic can reach others in that community who have never seen your name.

What has changed since 2023

The guides from Hootsuite and Sprout Social add three things for 2026. Premium subscribers get a boost: Hootsuite says Premium users are boosted with followers and non-followers alike, and Sprout Social that Premium accounts are generally favoured. Links hold posts back: Hootsuite reports from its own experiment that posts without links do much better. And Sprout Social reports that X has integrated Grok into its ranking and, since 2025, into the Explore tab’s summaries of trends.

What costs a post reach

The published weights show what costs a post on X. A reader choosing “show less often”, blocking or muting counted −74 against a like’s 0.5, and a report −369. Posts that provoke without giving anything (rage bait, misleading hooks, tagging strangers) can earn replies and still lose overall. Posting and leaving costs too: a reply the author answered weighed most of anything in the code.

Questions

Sources

1. X, X’s recommendation algorithm (open-source repository), March 2023.
2. X, Heavy ranker: engagement weights, April 2023.
3. Hootsuite, How the Twitter (X) algorithm works, October 2026.
4. Sprout Social, How the Twitter (X) algorithm works, October 2026.