How the X Algorithm Works in 2026: The Signals That Actually Rank You

X's algorithm has been more publicly documented than any other platform's — the recommendation code was open-sourced — yet most creators still operate on folklore. The reality: X runs a prediction engine that scores every tweet for every user, and its signal hierarchy is both stranger and more learnable than people assume.

Here's how the For You feed ranks in 2026, every signal that matters, and the counterintuitive mechanics (like why outrage works and why that's a trap).

The core prediction: will you engage?

X's feed isn't chronological and isn't pure engagement-ranking — it's a prediction machine asking one question per tweet per user: how likely is this person to reply, retweet, like, or linger? Each interaction type carries a different weight, and the hierarchy inverts most creators' instincts: a reply is worth multiples of a retweet, which is worth multiples of a like. Engagement that creates conversation beats engagement that just endorses.

The practical hierarchy: replies first (conversation generates more conversation, which the feed reads as community value), retweets second (distribution to new networks), likes third (light endorsement), bookmarks and profile clicks as secondary quality signals. Creators optimize for likes because they're visible; the algorithm prices replies because they're scarce.

Signal 1: The reply economy

Replies are the heaviest-weighted signal on X — a reply is a public conversation that recruits other participants, and the algorithm treats reply-thread depth as the strongest content-quality proxy. This is why question tweets, opinion-with-humility posts, and 'what am I missing?' closers outperform announcements by multiples.

The dark side is well documented: anger and controversy generate replies cheaply, and the algorithm can't fully distinguish 'replying to agree' from 'replying to fight.' Outrage content does outperform on raw signals — briefly. The accounts that ride it build reply-volume and destroy audience quality, attracting the audience nobody wants and repelling the one everyone does. The sustainable reply economy is built on usefulness and genuine questions, not provocation.

Signal 2: Dwell time and the scroll-stop

X measures how long a tweet holds you — dwell time is a core input, and it changes format strategy entirely. A long tweet that expands ('show more') generates dwell; a thread holds dwell across minutes; a video pauses the scroll. The tweet that gets read for twenty seconds beats the tweet that gets liked in one.

The format implications: hooks that stop the scroll in the first line (the fold shows roughly 280 characters before truncation on mobile), threads that pull readers through with open loops, and media that demands a pause. The discipline is identical to other platforms — earn the first second, then the next thirty — but on X the dwell signal is more explicit in the code than anywhere else.

Signals 3–5: Recency, your graph, and the negatives

3) Recency: X weights fresh content heavily — the half-life of a tweet is hours, not days, which is why volume and timing matter more here than on LinkedIn. 4) Your graph: accounts you interact with surface more, which makes reply-networking (genuine replies to accounts in your niche) a distribution strategy, not just etiquette. 5) The negatives: mutes, blocks, 'not interested,' and report clicks — X's algorithm explicitly incorporates negative feedback, and it compounds: an account that generates habitual mutes trains the engine to stop showing it, invisibly, before the follower count ever reflects the decay.

The meta-lesson: X rewards the creators who generate conversation their audiences want to be seen participating in. Every strategic choice — topic, tone, format, reply behavior — filters through that single question.

Tips

• End tweets with genuine questions weekly; replies are the ranking fuel and questions are the cheapest way to earn them.
• Write threads with one open loop per post; thread dwell is the platform's most concentrated quality signal.
• Track replies-per-impression as your north star, not likes; it predicts reach on X better than any other visible metric.
• Audit your mute-block exposure quarterly: negative signals decay reach invisibly, long before follower counts show it.

Common mistakes

• Like-optimization — Creators polish tweets for endorsements when the algorithm prices conversation; a tweet engineered for likes underperforms a rougher tweet engineered for replies on the metric that actually distributes.
• Outrage short-termism — Controversy buys replies and destroys audience quality; the algorithm can't distinguish agreement replies from anger replies, but your future buyers can, and they leave quietly.
• Ignoring the mute economy — Negative signals compound silently; accounts lose half their effective reach to habitual mutes without ever seeing it in the metrics, and the fix (tone, topic calibration) only happens if you audit for it deliberately.

Example scenario