Not many platforms open-source their algo.

But X does, and since we’re all here anyway, I thought about posting on how the For You feed works ↓
1. X mixes posts from people you follow with posts from accounts you don’t follow.
2. Then it ranks them based on what it thinks you’ll care about.
3. The flow is pretty simple:
your activity → possible posts → filters → engagement prediction → final feed
4. There are 2 main sources:
→ Thunder: posts from people you follow
→ Phoenix: posts from the wider X network
5. Phoenix is the part that helps posts reach new people.
If X thinks your post matches someone’s interests, it can show up even if they don’t follow you.
6. The algo looks at your behavior and finds similar posts from a huge pool.
Then it ranks them based on what you’re likely to do.
7. It predicts stuff like:
likes, replies, reposts, quotes, clicks, profile clicks, video views, photo expands, shares, dwell time, follows, blocks, mutes, reports, and “not interested.”
8. Good signals help. Bad signals hurt.
→ Good: likes, replies, reposts, shares, clicks, follows, dwell time
→ Bad: blocks, mutes, reports, “not interested”
9. Replies matter.
If people reply to your post, that’s a strong sign the post is starting a real conversation.
10. Dwell time matters too.
X doesn’t only care if someone likes your post. It also cares if they stop and spend time reading it.
11. Profile clicks and follows matter.
A post that makes someone check your profile is stronger than a post they scroll past.
12. Shares matter a lot.
If people send your post to someone else, DM it, or copy the link, that’s a strong signal.
13. The feed is personalized.
Your For You page is based on your own behavior, not some universal “best posts” list.
14. Staying on-topic helps.
If you post around a clear niche, X has an easier time figuring out who should see your content.
15. X also tries not to show too many posts from the same account in a row.
So even good accounts get spaced out.
16. A lot gets filtered before ranking.
Duplicates, old posts, muted words, blocked accounts, posts you already saw, spam, violent content, gore, etc.
17. The public repo is not the full live system.
It shows a lot, but production is bigger and trained on real-time data.
18. The exact weights are not fully public.
We know what signals X tracks, but X can still tune how much each one matters.
19. Timing still matters.
A post needs early signals from the right people before it gets pushed wider.
20. The real goal is training the algo.
Every good post teaches X who your content is for.
If the same type of people keep reading, replying, sharing, clicking, and following, the system gets better at finding more people like them.
Still pretty cool they open-sourced this much.