The X algorithm for founders comes down to one idea from X’s published code: the For You feed predicts how likely each viewer is to act on your post, from replying and dwelling to muting you, then combines those predictions into a score. Founders earn reach by posting what the right people will answer, read and follow.
That summary comes from a primary source rather than from guesswork. X publishes the For You code in the xai-org/x-algorithm repository on GitHub, with a README that explains the pipeline in plain language. Twitter released an earlier version in 2023 as the-algorithm. This play reads the current README with a founder’s question in mind: what should I do differently on Monday?
Where For You gets its posts
According to the README, the For You feed is assembled for each request from two kinds of source.
- In network. A system called Thunder keeps recent posts from the accounts a viewer follows.
- Out of network. Phoenix retrieval and SimClusters find posts from accounts the viewer does not follow, and a popular posts source adds a shared list of popular recent posts.
Both kinds are ranked together by the same model. For a founder with a small audience this matters. Your followers see you through the in network path. Everyone else only meets you if retrieval decides your post is relevant to them, which is why posts that start conversations with people outside your followers carry so much weight early on. The play on how AI founders grow on X before launch builds on exactly this.
What the ranking model predicts
The ranking model, Phoenix, reads the viewer’s recent engagement history and predicts, for each candidate post, how likely the viewer is to take each action. The README groups the actions like this:
| Group | Actions the model predicts |
|---|---|
| Engagement | Favorite, reply, repost, quote, share, share via DM, share via copy link |
| Clicks | Post, profile, link, photo expand, video open, quoted post |
| Attention | Video quality view, dwell, dwell time, click dwell time, video continuation seconds, profile visit seconds |
| Author | Follow author |
| Negative | Not interested, mute author, block author, report, not dwelled |
A scorer then adds these up as a weighted sum. Positive actions carry positive weights and negative actions carry negative ones. The README adds a warning worth repeating: the weights scale the predicted probability of an action for a given viewer, not raw counts. You cannot read the weights as “one report cancels so many likes”. The prediction is driven by each viewer’s own behaviour.
What that means for a founder
- A post your buyers are likely to reply to, read for a while, or follow you after is the kind of post the model is built to surface.
- A post that makes people tap not interested or mute works against you, even if it gets attention from others.
- Likes are one action among many. A post engineered only for likes ignores most of the list.
The adjustments after scoring
The README lists three adjustments applied after the weighted score:
- Author diversity. Each post after an author’s first is multiplied by a decaying factor, down to a floor. Posting ten times in a row does not give you ten equal chances.
- Out of network discount. Posts from accounts the viewer does not follow are multiplied by a factor below one, as are replies and reposts from followed accounts.
- New author boost. Posts from authors whose impressions are below a threshold are lifted toward a target position.
For founders, the first point argues for fewer, better posts. The second explains why growing your follower base among real buyers compounds: you move from the discounted path to the in network one. The third is a small, documented reason a new account is not hopeless.
What filters remove
Before scoring, a set of filters removes candidates. The README’s list includes posts older than 48 hours, the viewer’s own posts, posts the viewer has already seen or been served, posts matching muted keywords, and posts from accounts the viewer blocks or mutes. After ranking, a visibility filtering system can drop a post or show it behind an interstitial, based on labels other systems attach to posts and accounts.
Two founder lessons follow. First, a post has a limited window in For You, so a launch post has to be ready when your audience is around. Second, behaviour that gets you muted or blocked does not just lose one viewer. It removes you from that person’s feed entirely. The README also describes a pilot transparency tool, Under the Hood, that shows aggregate labels on your account and posts.
How to post with this in mind
None of this is a trick. It is a reason to post the way good founders already do.
Write for a reader, not for the feed. Pick one kind of buyer for each post. Describe their problem in their words. The X content strategy for technical founders play has a weekly mix built on this.
Ask for an answer, not a like. End with a real question your buyer can answer from experience. Replies and conversations are part of what the model predicts.
Make the post worth dwelling on. A clear first line, a concrete example, a short demo. Threads still work when each part earns the next; see how to write a Twitter thread.
Reply where your buyers already are. Thoughtful replies on their posts put you in front of people who do not follow you yet, and give them a reason to.
Space out your posts. Because of author diversity, one strong post followed by replies usually beats a burst of similar posts.
Avoid what invites mutes. Engagement bait, repeated self promotion and tagging strangers all risk the negative actions in the table.
Founder account or company account
The ranking model scores posts for each viewer, and the follow author action is one of the things it predicts. People tend to follow people they want to hear from again. That is one reason many early AI startups lead with the founder’s account and keep the company account for announcements, docs and support. The play on founder vs company accounts covers how to split the work between the two.
Whichever account leads, keep the profile clear. A profile visit is one of the predicted actions, and the bio, pinned post and link decide what happens after someone clicks your name.
A weekly routine that fits the code
A simple week for a founder who wants reach from the right people:
- Two or three original posts on the problem your buyer has, each with one concrete example or a short demo of the real product.
- Daily replies on posts from people in your market, written to add something, not to promote.
- One conversation starter that asks buyers about their current workaround.
- One review at the end of the week: which posts earned replies, dwell and follows, and which drew silence or mutes.
Keep the routine for a few weeks before judging it. The feed is personal, so results build as more of the right people interact with you.
What the code does not tell you
The repository explains the structure and the kinds of signals. It does not tell you how a particular post will perform, and the README notes that the system changes over time with dated updates. Treat any thread that claims a secret formula with suspicion, and check claims against the source. Track your own results instead: which posts earned replies from buyers, which brought follows, which led to a site visit.
If you want a second opinion on whether your account and posts are set up for this, ask for a teardown. The content calendar is a simple way to plan fewer, better posts across a week.