How the X algorithm works in 2026

#Quick answer

The X algorithm ranks your For You feed by prediction, not by who you follow. Each session it collects about 1,500 candidate posts out of the roughly 500 million created that day, scores every one for how likely you are to react, and shows you the winners. Replies and reposts count for far more than likes, external links get suppressed, negative signals like blocks pull a post down, and early engagement in the first hour is the strongest lever you control.

X (formerly Twitter) open-sourced this system twice - the 2023 twitter/the-algorithm release and the January 2026 Grok-powered xai-org/x-algorithm rewrite - so this is not guesswork. Here is what the code actually shows, and what to do about it.

#The three stages: source, rank, filter

Every time you open the For You tab, X runs the same pipeline in real time.

1. Candidate sourcing. Out of ~500 million posts a day, X narrows to about 1,500 candidates for you. Roughly half come from your network - recent posts from accounts you follow, served from an in-memory store the code calls Thunder (the successor to the old Earlybird search index). The other half are out-of-network: posts from people you do not follow, surfaced by machine learning because people similar to you engaged with them.

2. Ranking. A transformer model - named Phoenix in the 2026 code, and powered by xAI's Grok - embeds you and each candidate post as vectors and predicts a probability for every action you might take: reply, repost, like, click, dwell, or a negative action like block or report. Those probabilities are combined into a single weighted score. The scoring weights are the heart of the algorithm, and they are literally a config file in the open-source repo.

3. Filtering. The top-scoring posts pass through a visibility layer that removes duplicates, applies diversity (so you do not see five posts from the same account), and enforces X's policies. A post here can be shown normally, placed behind an interstitial, or dropped entirely. The survivors are your feed.

Then it reruns on your next refresh. Nothing about the feed is static.

#What the algorithm rewards

The weighted-sum design means some actions are simply worth more than others. Across both open-source releases the ordering is consistent:

| Signal | Relative weight | Why it matters | |---|---|---| | Reply (author replies back) | Highest | A real conversation is the strongest positive signal | | Reply | High | Predicts a two-way interaction, not passive consumption | | Repost | Medium-high | Shares your post to a new network | | Dwell time | Medium | People actually reading, not scrolling past | | Like | Low | The weakest positive action - easy, so it counts least |

In practical terms: a post with a handful of genuine replies out-reaches one with many silent likes. If you want the algorithm on your side, write posts that make people want to answer, and answer them back. That is also why replying well on X is one of the fastest ways to grow - your reply is itself a ranked candidate in someone else's thread.

Native content is favored over anything that sends people away. Text and video that keep attention on X score better; a naked external link in the main post gets suppressed because it predicts people leaving. If you need to share a link, put it in the first reply, not the post.

#What the algorithm penalizes

Negative-action predictions are weighted heavily against a post. If the model thinks you are likely to block, mute, report, or tap "not interested," that prediction subtracts from the score - a few of those can bury an otherwise fine post. This is the mechanism behind so-called "shadowbans," which are not a real platform state: what people experience is visibility limiting, where a post or account is down-ranked or filtered under X's rules. If your reach drops suddenly, check whether your posts are triggering negative signals before you blame a ban.

Tactics that tend to backfire:

• Engagement bait ("like if you agree," "reply YES"): X has said it down-ranks manipulative bait, and hollow engagement teaches the model your audience is low-quality.
• Mass-tagging and reply spam: pulls reports and mutes, the worst signals.
• Delete-and-repost: resets the early-engagement window and looks mechanical.
• Over-scheduled, identical timing: a feed that posts at the exact same second every day reads as automated. Vary it and leave room for real-time posts.

#The role of timing and early engagement

The ranker reads velocity. How fast a post earns real replies and reposts in its first 30-60 minutes strongly shapes how widely it gets shown, because early signals are the model's best early predictor of overall engagement. A post that lands flat in its first hour rarely recovers.

This is why when you post matters - not as a magic minute, but because posting into an awake audience is what makes early engagement possible. The best time to post on X is a weekday mid-morning-to-afternoon band in the 2026 studies, but your own analytics beat any global chart. Timing is a multiplier on a good post, never a substitute for one.

#Does X Premium change the math?

Yes, in a defined way. Per X's own About X Premium page, Premium and Premium+ include tiered reply prioritization - your replies rank higher in other people's threads - with Premium+ getting the largest boost. Buffer's analysis of 18M+ posts found Premium accounts saw materially more reach on average.

That makes Premium a genuine amplifier, especially if replying is a core part of your strategy. But it multiplies whatever you already have: pay for Premium and keep posting weak content, and you have bought a bigger platform for posts nobody engages with. Decide with the numbers - our breakdown of X Premium tiers and price walks through when the boost pays for itself.

#What to actually do about it

This algorithm is complicated; working with it is not. Six habits cover most of the gain:

1. Write for replies. Open with a hook that invites a response - a question, a contrarian take, a specific claim someone will want to correct.
2. Reply back fast. An author reply is the highest-weighted signal there is, and it extends the early-engagement window.
3. Keep links out of the post. Put the URL in the first reply so the main post stays native.
4. Post when your people are awake. Use your X analytics to find your real active hours, not a global average.
5. Use replies to reach new networks. Thoughtful replies on bigger accounts are ranked candidates that put you in front of audiences you do not follow - a core tactic for getting followers on X.
6. Be consistent. The model rewards accounts that reliably earn engagement over time. One great week does not outweigh three quiet ones.

Doing all six by hand, every day, is the hard part. X-Autopilot runs your posting and replying rhythm from your own browser and residential IP - a lower detection surface than cloud tools that post from a shared IP pool - so you hit your windows and keep the conversation going without living in the app. It automates the cadence the algorithm rewards; it does not automate having something worth saying.

#The short version

The X algorithm predicts what you will react to and ranks accordingly. It rewards conversation over passive likes, native content over exit links, and fast early engagement over everything. Write posts people answer, reply back quickly, keep links in the replies, and post when your audience is awake - the open-source code and every 2026 study point the same way.

#Internal links

• How to reply to posts on X: replies are the highest-weighted signal in the ranker
• Best time to post on X in 2026: post into an awake audience so early engagement can fire
• How to get followers on X: use out-of-network replies to reach new audiences
• X Premium tiers and price: when the reply-prioritization boost is worth paying for
• X analytics: read your native dashboard: find your real active hours and your best-performing hooks
• Twitter shadow ban check: what visibility limiting actually is, versus the "shadowban" myth