X open-sourced the algorithm that decides what you see. I read it. Here's what actually gets reach — with the real numbers.
On August 13th, xAI pushed a fresh drop of the For You feed code — and this release finally includes the actual weights. Not influencer folklore: the multipliers, from home-mixer/params/param.rs. I read it so you don't have to.
The score is just a weighted bet on what you'll do
Every post you might see gets a predicted probability for ~26 actions — will you reply, share, mute, report — and each probability is multiplied by a fixed weight and summed. That's it. That sum (with a couple of adjustments) is the feed order.
The weights, ranked
| Action | Weight |
|---|---|
| Copy-link share | 20.0 |
| Reply (from a mutual follow, original post) | 20.0 |
| Reply (base) | 5.0 |
| DM-share | 5.0 |
| Quote-tweet | 5.0 |
| Follow-the-author from the post | 4.0 |
| Generic share | 2.0 |
| Repost | 1.0 |
| Like | 0.5 |
| Click | 0.4 |
| Photo/video interactions | 0.05 |
| "Not interested" | -43.2 |
| Mute author | -58.8 |
| Report | -234.0 |
Read that table again. A copied link is worth forty likes. A like is nearly noise. And one predicted report erases the value of eleven replies.
What this actually means
-
The algorithm doesn't reward being liked. It rewards being SENT. The heaviest weight in the whole system is someone copying your link to share it somewhere else. Write the post someone forwards to their group chat.
-
Replies from mutuals are gold — but only on original posts. The +15 boost (reply 5 → 20) requires an original post and a mutual follow. Replies and reposts get NOTHING — in fact they're discounted 25% even to your own followers, and non-followers never see them at all. The reply-guy strategy is mathematically dead.
-
Engagement bait is negative-EV. "Repost if you agree" harvests mute (-58.8) and not-interested (-43.2) predictions. The math says bait is self-harm. There's also a separate LLM that literally grades posts for "slop."
-
Small accounts get a real break. If you have under 1,000 followers: an original post, under 1,000 impressions, less than 24 hours old, gets force-lifted to around feed position 15 by a built-in cold-start boost. The system deliberately gives new voices a shot — one good original post a day is the play.
-
Bursts cannibalize themselves. Your 2nd post in someone's feed session is worth 62% of the first, the 3rd 44%, floor 25%. Space it out.
-
Everything dies at 48 hours. Hard filter. Reach is a same-day game.
-
Two myths, busted by absence: there is no text-length signal anywhere in the code (long threads don't inherently rank), and no link penalty in the published ranking math (links only enter via click prediction).
The honest caveats
These are the open-sourced defaults — production weights can be A/B-tuned differently, and X runs closed layers on top (visibility labels that can silently cap you to followers-only). But the structure is real: what's rewarded, what's punished, the boosts, the filters. That's the game board.
I've already wired these rules into my own cross-poster. Every claim above cites a file in xai-org/x-algorithm — go read it yourself. That's the point.
Update (Aug 14): xAI updated the repo docs to correct a misreading going around — the weights multiply predicted probabilities of each action, not raw counts. So "1 report cancels 468 likes" (a take you may have seen) is bad math. That's the framing this post already used — every number above is a weight on a probability — but worth saying explicitly since the wrong version spread fast. Two other additions since I first read the code: a Brazil-2026 election filter (legal compliance, removes Electoral-Court-reported accounts from For You unless you follow them), and an "Under the Hood" tool that lets you see the visibility labels on your own account and posts. That last one is the sleeper: for the first time you can check whether your account is carrying an invisible penalty instead of guessing.