Draft: unlisted, not indexed

How the X algorithm works in 2026 (and what design posts that broke out have in common)

· Roy Quilor

Short answer

X's For You feed is ranked by a model that predicts what you'll do with each post: like it, reply, repost, share, click, watch, follow the author, or mute and move on. Each prediction is multiplied by a weight, the results are added up, and the highest-scoring posts are shown. X publishes the code and the weights on GitHub.

For a designer, that means one thing in practice: make posts that people stop for, reply to and share. A clear visual of your work, a story, or a question your audience wants to answer.

Based on X's published ranking code (version of 18 November 2026) and 151 design posts from X, 10–17 November 2026. Sources are at the end.

How the For You feed is put together

X put the code behind For You on GitHub in January 2026 (xai-org/x-algorithm, Apache 2.0) and keeps a dated log of changes in the README. Each time you open For You, it:

  1. Collects candidates. Recent posts from people you follow, plus posts from accounts you don't follow that are similar to what you engage with, plus a shared list of popular recent posts.
  2. Filters. It removes posts older than 48 hours, your own posts, posts you've already seen, and posts from accounts you've blocked or muted.
  3. Predicts your reactions. A transformer model X calls Phoenix (press reports describe it as built on Grok) estimates how likely you are to like, reply, repost, quote, share, click, watch, follow, or mute, block, report or scroll past each post.
  4. Scores. Each prediction is multiplied by its weight and added up. Negative actions have negative weights.
  5. Adjusts. A second post from the same author scores a bit less, then a third less again. Posts from accounts you don't follow are discounted. And posts from authors with few impressions get a boost toward a target position.
  6. Mixes. It reorders the top results so neighbouring posts aren't too similar, then adds ads and "who to follow".

The weights, in plain English

These are the default weights in X's published code on 18 November 2026 (param.rs). X says these defaults are kept in line with production, but it runs experiments on a small share of traffic, so treat them as a snapshot.

  • Like: 0.708
  • Reply: 5.0, with an extra boost when you and the author follow each other
  • Repost: 1.0 · Quote: 5.0
  • Share: 2.118 · Share via DM: 5.0 · Copy link: 20.0
  • Follow the author: 4.0
  • Click into the post: 0.274 · Open a link: 0.2 · Click the profile: 0

One warning from X's own code: these weights multiply the predicted chance of each action, not the raw counts. Rare actions get bigger weights partly because they're rare. So "one reply is worth 7 likes" isn't quite right. What it does say: a post people are likely to reply to, share or follow you from beats a post people only like.

What that means for your posts

  • Start conversations. A reply is weighted about 7× a like. Posts that ask something real, or tell a story people want to answer, score better.
  • Make it worth sharing. Sending a post in a DM or copying its link are among the highest weights. Useful, specific work gets shared; vague tips don't.
  • Small accounts get a lift. The new-author boost exists for authors with few impressions, and out-of-network posts are part of For You, so a small account can reach far beyond its followers.
  • Bait has a cost. Mutes, blocks, reports and "not interested" have negative weights. A post that gets reach by annoying people can lose more than it gains.
  • Links aren't penalised, but they don't help. I couldn't find a link penalty in the published code, but opening a link is weighted below a like, and people who leave X don't stay to reply.

What design posts that broke out had in common

I pulled design posts from X's Top results between 10 and 17 November 2026 and kept 151 that were clearly about UI, UX or product design. A post counts as a breakout when its views reached 10× or more the author's follower count.

By format:

  • Photo (a screenshot or shot of the work): 44% broke out (14 of 32)
  • Video: 34% (23 of 67)
  • Text only: 33% (16 of 49)

A still image of the work beat everything else. It's quick to read in a scrolling feed, and people stop for it.

Bookmarks outnumber replies. The typical breakout got about 17 bookmarks for every reply. Design posts get saved more than they get discussed, which is why the ones that do start a conversation stand out. The breakout with the most replies (405) was a UI director's story about the first game UI he designed, not a tip or a list.

Bait works, for a while. The 13 bait-style posts ("bookmark this", "N free resources") broke out at 62%, the highest of any group. They also teach your audience to expect freebies, not your work, and they're the kind of post people mute, which the ranking counts against you. I wouldn't build an account on them.

Sources and method

  • Code and weights: xai-org/x-algorithm, README and home-mixer/params/param.rs, read on 18 November 2026 (commit 9081b33). The repository was created on 19 January 2026.
  • Grok: the README calls the model Phoenix and doesn't mention Grok; the Grok link is from press coverage of the January release.
  • Design posts: 151 posts clearly about UI, UX or product design from four X searches (Top tab), 10–17 November 2026, at least 24 hours old, pulled through a third-party X search API. Breakout = views at least 10× the author's follower count on the day of the pull. "Bait" = posts matching phrases like "bookmark this" or "N free…".
  • Limits: one week of Top results, which favour posts that already did well. This is a different, stricter filter than the best-time post (344 posts), so the counts differ.
  • Last updated: 18 November 2026.

Frequently asked questions