Free trialTry Starter or Influencer free for 2 days.Start free trial →
X Growth

How the X Algorithm Works in 2026 (And How to Make It Work for You)

Short answer

The X algorithm builds your For You feed in three steps: it collects candidate posts from accounts you follow and from accounts you don't, scores each one with a transformer model called Phoenix that predicts how likely you are to like, reply, repost, dwell on, or report it, and then filters and reorders the top results. The code is public on GitHub (xai-org/x-algorithm). For creators, the practical lesson is that replies, reposts and time spent count for far more than likes, and negative signals like mutes and reports cost more than any positive signal earns.

Key takeaways

  • X published the full For You pipeline in 2026: candidate sourcing, a Phoenix ranking model, a weighted scorer, and the visibility filtering rules, under an open-source license.
  • Phoenix predicts a probability for about 20 actions per post, including favorite, reply, repost, share, profile click, dwell, follow author, and negative actions like mute, block and report.
  • The final score is a weighted sum of those probabilities. The weights live in a config file in the repo, and X notes that negative actions are weighted heavily because they're rare.
  • Posts older than 48 hours are filtered out before scoring, so reach happens in the first two days or not at all.
  • Out-of-network posts get a discount and new authors get a boost, which is why a stranger's great post can still beat a follower's average one.
  • Ranking and visibility filtering are separate systems: a label can hide a post from non-followers no matter how well it would have scored.

Twigor sells growth plans for X accounts. We clarify where our product fits and where it does not, and every claim about X's rules links to the source.

Most advice about the X algorithm is folklore. Post at 9 a.m. Never use links. Always reply to your own post. Some of it is right, a lot of it is guesswork dressed up as rules, and until recently there was no way to check.

Now there is. X published the code that builds the For You feed, first in January 2026 and then in a much larger release in August, including the scoring configuration and the filtering rules. We read it so you don’t have to. What follows is how the feed is actually assembled, in plain language, and what that means for anyone trying to get posts in front of people who don’t follow them yet.

What Does the X Algorithm Actually Do?

It turns millions of recent posts into the few dozen you see, in three stages: source candidates, score them, filter and reorder the top ones. Everything in the repository hangs off that pipeline.

The sourcing stage runs several fetches in parallel. Thunder keeps recent posts from the accounts you follow in memory, so in-network posts come fast. Phoenix retrieval embeds you and millions of posts as vectors and pulls the out-of-network posts that look closest to your interests. SimClusters groups people into interest communities and adds posts popular in yours. A list of broadly popular posts fills in the rest. Together these produce a candidate pool of a couple of thousand posts for one feed load.

Before anything is scored, a set of pre-scoring filters removes duplicates, posts older than 48 hours, posts from accounts you’ve blocked or muted, and posts you’ve already seen.

Then the scoring stage runs, which is the part people mean when they say “the algorithm.” Then post-selection filters apply, including visibility filtering, and a blending step mixes in ads, follow suggestions and prompts.

How Does X Decide What to Show First?

A transformer model called Phoenix predicts how likely you are to take each of about twenty actions on each post, and a scorer adds those predictions up with weights. That number is the post’s score, and the feed is roughly sorted by it.

The actions Phoenix predicts, straight from the README:

  • Engagement: favorite, reply, repost, quote, share, share via DM, share via copy link
  • Clicks: on the post, the profile, a link, a photo, a video, a quoted post
  • Attention: a quality video view, dwell, click dwell time, video continuation seconds, profile visit seconds
  • Author: following the author
  • Negative: not interested, mute author, block author, report, and not dwelling at all

The weights that combine them sit in a config file (home-mixer/params/param.rs). X added a note in August 2026 that matters for reading them: the weights scale predicted probabilities, not raw counts. A report is more than a thousand times rarer than a like, so its weight is large just to let the prediction register. You can’t read the ratio as “one report cancels a thousand likes.” You can read it as “negative reactions are expensive.”

After the weighted sum, three adjustments apply before the top posts are picked: an author diversity rule so one account doesn’t flood your feed, a discount on out-of-network posts, and a boost for new authors. Then a reranker makes a final pass.

This is a real change from the 2023 version, which stacked hand-built features (a reputation score for authors, explicit topic affinities, a separate heavy ranker) on top of each other. The 2026 pipeline learns most of that from your action history instead. There’s no published “author score” anymore. There’s a model guessing what you, specifically, will do with this post.

What Does the Algorithm Reward?

Actions that cost the viewer something. A like is one tap. A reply, a repost, a share to a DM, a click through to your profile, or thirty seconds spent reading: each of those is a stronger prediction target, and each contributes more to the score when the model thinks it’s likely.

That has a few concrete consequences for how you post.

Posts that invite a reply beat posts that invite a like. A question people want to answer, a take people want to argue with, a list people want to add to. Replies the author then answers do even better, because the conversation keeps generating actions.

Dwell matters, and dwell is honest. A clickbait first line that people bounce off doesn’t help you, because “not dwelled” is on the negative list. A post that takes twenty seconds to read and gets read all the way through does help.

Reposts and quotes carry the post to new in-network audiences, which feeds the next round of candidates. This is the compounding part: engagement from the right people doesn’t just raise today’s score, it changes who sees tomorrow’s post.

Profile clicks and follows are predicted too. A post that makes a stranger check who you are is doing exactly what the model is looking for.

What Does the Algorithm Punish?

Anything that predicts a mute, a block, a report, or a quick scroll past. The negative actions are rare, which is why they’re weighted heavily, and a post that triggers even a small predicted probability of them loses ground fast.

In practice the patterns that earn negative signals are familiar: the same post repeated, replies that read as promotion under unrelated threads, bait that doesn’t pay off, and anything that makes a stranger think “why is this in my feed.” The algorithm doesn’t know you’re a nice person. It knows what people did the last time they saw something like this from someone like you.

One more thing the code makes clear: ranking and visibility filtering are separate systems. A post can score well and still be dropped or hidden from non-followers because a label sits on it or on your account. If your posts reach followers and nobody else, check your Under the Hood report before you change anything about your writing. We covered how in our guide to shadowbans on X.

Does the Algorithm Treat Small Accounts Differently?

Partly. New authors get a boost, out-of-network posts get a discount, and the model has less history to work with, so early signals count for more. That cuts both ways.

The boost means a brand-new account’s post is not automatically buried. The discount means that to a viewer who doesn’t follow you, your post has to beat the in-network posts in their feed by a margin. And because Phoenix predicts actions from engagement history, an account with very little history is an account the model is unsure about. The first people who engage with you teach it who else might.

That is why early engagement from real people in your niche is worth so much more than a thousand bot followers. Bots don’t reply, don’t dwell, and don’t share, so they teach the model nothing. Worse, when X’s spam sweeps remove them, they take your follower count down with them, as we explained in Why Am I Losing Followers on X?

How Do You Make the Algorithm Work for You?

Give the model the actions it’s looking for, from the people it’s trying to predict. Nothing in the code rewards tricks. Everything in it rewards posts that specific humans respond to.

Write for the reply, not the like

End with something people can answer. Disagree with something specific. Ask for the one tool, the one book, the one mistake. Then answer the replies you get, quickly, in the first hour or two while the post is still fresh in the candidate pool.

Front-load the post and earn the dwell

The first line decides whether anyone stays. Make it carry real information, then make the rest worth the time. Long posts and native video do well not because length is rewarded but because time spent is.

A link isn’t penalized by name in the published code, but a post whose only action is “click the link and leave” predicts poorly on everything else. Put the substance in the post, the link in a reply or at the end, and give people a reason to stay before they go.

Reply where your readers already are

Thoughtful replies under larger accounts in your niche are seen by their audience, generate profile clicks, and build the engagement history the model learns from. Drive-by replies do the opposite.

Post inside the 48-hour window, then move on

A post’s recommendation life is about two days. Don’t delete and repost; don’t reply to your own post ten times to “bump” it. Post, engage with the replies, and write the next one.

Get your posts in front of the right people

Everything above depends on real people seeing the post in the first place. If you’re starting from a small audience, that’s the bottleneck. You can run X Ads, you can reply your way in over months, or you can use a promotion service that puts your posts in front of a targeted audience every day so that real engagement starts arriving. That last option is what Twigor does; we compared it with ads and with buying followers in X Promotion vs. X Ads vs. Buying Followers, including when ads are the better choice.

Conclusion

The X algorithm in 2026 is less mysterious than the advice about it. Candidates come from who you follow and what you’ve engaged with, a model predicts what you’ll do with each post, negative reactions cost more than positive ones earn, and a separate layer of labels can hide a post regardless of its score. If you want more reach, the code points at one thing: posts that make real people reply, stay, and come back. Build for that, keep the account clean, and let the model do what it’s built to do.

Frequently asked questions

Is the X algorithm really open source?

Yes. Twitter published a partial version in March 2023. X published a rebuilt version at github.com/xai-org/x-algorithm in January 2026 and expanded it in August 2026 with the scoring configuration, visibility filtering code, Phoenix training code and SimClusters. It's under the Apache 2.0 license.

What does the X algorithm reward most?

Actions that show real interest: replies, reposts, quotes, shares, profile clicks, and time spent on the post. Follows earned from a post are strong too. A like is the weakest positive signal. Mutes, blocks, reports and 'not interested' pull a post down hard.

Does the algorithm favor X Premium accounts?

Premium affects reply ranking (paid accounts' replies are prioritized) and adds features, but the open-source ranking code scores posts on predicted engagement, not on subscription status. A Premium account with posts nobody engages with still gets little reach.

How long does a post stay eligible for the For You feed?

The pre-scoring filters in the published pipeline drop posts older than 48 hours, so a post's window for recommendation is about two days. Engagement after that mostly comes from search, profile visits and reposts.

Why do my posts reach followers but not strangers?

Two possible reasons. Either a visibility label is hiding the post from recommendations to non-followers (check x.com/i/under_the_hood), or the post simply scores below the out-of-network bar after the discount applied to posts from accounts the viewer doesn't follow.

Can buying followers help with the algorithm?

Bot followers hurt it. They never engage, so your predicted engagement per viewer drops, and X's spam sweeps remove them anyway. What helps is real people engaging with your posts, which is what promotion to a targeted audience produces.

Sources

  1. x-algorithm: the code behind X's For You feed , xAI / X, GitHub
  2. Open-sourcing the For You timeline , X Open Source, August 13, 2026
  3. Twitter's Recommendation Algorithm (2023 release) , Twitter, GitHub, March 31, 2023
  4. X algorithm: how the newly open-sourced ranking system works , Multilogin, 2026
  5. Twitter/X Algorithm 2026: What xAI's Open-Sourced Code Shows , Postory, October 2026
Last updated
Writer, Twigor Blog

Web developer with nine years of WordPress work. Writes the Twigor blog's guides on growing and protecting accounts on X.