How the X (Twitter) algorithm actually works in 2026: the signals that decide your reach

In 2023 X open-sourced the core of its recommendation algorithm. That's rare - most platforms keep ranking a black box. It means we can stop guessing about "the algorithm" and instead read what actually moves a post up or down the feed. Here's the honest version.

The feed is a ranking problem, not a timeline

Every time someone opens X, the system assembles a set of candidate posts - from people they follow (in-network) and, crucially, from people they don't (out-of-network). Each candidate is scored by a model that predicts how likely you are to engage, and in what way. The post with the highest predicted engagement wins the slot. Your job as a creator is to raise that predicted score.

Not all engagement is weighted equally

This is the part most advice gets wrong. A like and a reply are not the same currency. The model weights predicted actions very differently:

• Replies you receive are weighted far more heavily than likes - a genuine conversation is the strongest positive signal.
• A reply the author engages back withis amplified again - the "reply-back" multiplier is one of the largest positive weights in the model. Being early in someone's replies, in your voice, is disproportionately valuable.
• Reposts and long dwell time signal the post is worth spreading and worth reading.
• Negative signals- "show less often", mutes, blocks, reports - carry heavy negative weight and can sink a post fast.

Out-of-network reach is where growth happens

Roughly half of a typical feed is out-of-network. That's the engine of growth: to reach people who don't follow you yet, your post has to clear a higher bar of predicted engagement. Replies into active conversations ("seed" replies) are one of the most reliable ways to earn out-of-network distribution, because they put you in front of an audience the original author already commands.

What actively suppresses reach

Some things quietly cap your reach regardless of quality: external links in the post body (X wants to keep users on-platform), engagement that looks low-quality or automated, and content that trips safety or spam classifiers. Writing a strong hook and putting links in a reply instead of the main post are two of the simplest fixes.

How to write for the algorithm without gaming it

The takeaway isn't "trick the model" - it's that the model rewards the same things a good post does: start a conversation, earn replies, be worth reading to the end, and avoid the patterns that trip negative signals. When you can see the weighted signals for a draft before you post, you stop guessing and start compounding.

X Radar drafts every post in your cloned voice and scores it against these exact signals in under a second, before you ever hit post. Sign in with X to try it free.