How the X (Twitter) Algorithm Works in 2026 — The For You Ranking, Explained From the Source Code
The For You feed builds a candidate pool that's roughly half posts from accounts you follow and half from accounts you don't, scores every candidate with a neural network that predicts which actions you'll take, then filters the ranked list for variety and quality. We know this because X open-sourced large parts of the system on March 31, 2023, and the code is still public at github.com/twitter/the-algorithm.
That repo is the reason this guide can be specific where most algorithm articles hand-wave. Below is what the code and its documentation actually show, followed by what it means for anyone trying to get read.
The three-stage pipeline
Stage 1: candidate sourcing. For each refresh, X assembles a pool of potential posts. The repo's own documentation puts the in-network share (accounts you follow) at about 50%, with the rest sourced out-of-network through follow-graph expansion and embedding similarity. That out-of-network half is where discovery happens: it's how strangers see your posts, and it's the only channel that turns readers into new followers at scale.
Stage 2: ranking. A model X calls the heavy ranker scores each candidate by predicting the probability of specific user actions. The model itself is published in a companion repo, the-algorithm-ml. The home-mixer configuration lists exactly which actions get weighted: favorites, retweets, replies, bookmarks, shares, video watch time, profile clicks, dwell, link opens, and on the negative side, reports and weak or strong negative feedback. The production weight values are set server-side rather than hardcoded, but the code and X's public statements are consistent on the ordering: replies sit near the top, an author replying to their repliers is a strong multiplier, likes are the lightest positive signal, and reports are heavily punitive.
Stage 3: heuristics and filtering. After scoring, the feed applies rules: author diversity so one account can't flood a feed, content balance, deduplication, and visibility filtering for policy-flagged content. A well-scored post can still be trimmed here.
Early velocity: the audition window
Your post's predicted-engagement scores start as guesses. Real behavior updates them fast. When a post goes out, the first sample of viewers generates the actual signal, and quick engagement relative to impressions expands distribution wave by wave. A quiet first 30 minutes usually ends the run.
Two posts of equal quality can land wildly differently for this reason alone. One published into an active audience window compounds. The same post at a dead hour gets a cold sample and stalls. Your analytics show when your impressions cluster, and publishing into those windows is the cheapest reach improvement available.
Dwell and conversation, the two signals worth optimizing
Dwell is the invisible one. The ranking configuration tracks whether people pause on your post, watch your video, or click through to your profile, and sustained attention counts even when nobody taps a button. This is why threads, longer posts, and native video punch above their weight: they hold readers long enough to register.
Replies are the loud one. A post that starts a genuine back-and-forth accumulates the platform's most valued signal, and the author joining the thread multiplies it. The practical version: end with a question you actually want answered, take positions worth pushing back on, and reply to every early response. Your replies under other, larger accounts do double duty, feeding the same signals while exposing your handle to someone else's audience.
Negative signals deserve equal respect. Mutes, blocks, "not interested," and reports carry heavy penalty weight in the code. Engagement bait that annoys ten readers to excite one is a losing trade.
Author reputation and the cold-start loop
The system also carries a sense of the author. Accounts with a history of posts that earn engagement get tested more confidently out-of-network, and new or empty-looking accounts get tested cautiously because there's no track record to bet on. Combine that with public follower and impression counts and you get the cold-start loop: a small account gets small test samples, small samples make velocity hard, and visible low numbers make strangers hesitant to engage with the account at all.
Breaking the loop is mostly work: a complete profile, a clear niche, a steady cadence, and weeks of showing up in other people's threads. Some accounts also establish a baseline follower count with a one-time purchase so posts and replies stop launching from an account that reads as abandoned. That raises the floor the loop starts from. It doesn't produce a single reply, dwell second, or repost, so it only makes sense alongside the habits above, never instead of them.
What the code settles, and what it doesn't
Worth being honest about the limits. The 2023 release was partial: no training data, no model weights, and the production weight values live in server-side configuration. What the repos settle beyond argument is the architecture (candidate sourcing, heavy ranker, heuristics), the signal list (the engagement types named in the code), and the design intent of predicting actions rather than counting followers. What still requires inference is exactly how much each action counts on any given day, since X tunes those values continuously. Any article quoting precise weight multipliers in 2026 is quoting a snapshot, not the live system.
The playbook this implies
- Publish when your audience is active, so the audition sample is warm.
- Make the first line carry the post. It decides the pause that becomes dwell.
- Ask real questions and reply to everyone in the first hour.
- Keep links out of the main post. Post the idea, link in a reply.
- Use threads and native video to hold attention.
- Spend daily time replying under bigger accounts in your niche.
The habit-level version of this list, with the account-setup and consistency work around it, is in our guide to growing on X. If your goal is the creator payout program, the eligibility gates are covered in the X monetization guide.
Questions the code answers
Why do my posts get almost no views even though I have followers?
Followers only define the initial sample. If that sample doesn't engage quickly, distribution stops before most of your followers ever see the post. Timing, a stronger first line, and prompting replies address the actual bottleneck.
Do links really reduce reach?
The ranking configuration tracks link opens as their own signal, and X's design favors keeping users on-platform. In practice, posts leading with external links consistently underperform, which is why link-in-reply became standard practice.
Is the algorithm on GitHub still current?
The architecture and signal types remain the best public map of the system, but X tunes weights server-side and has changed emphasis since 2023, notably toward video and Grok-adjacent features. Treat the repo as the blueprint, not a live mirror.
Next step
If the cold-start floor is what's holding your account back, FastSocial's one-time X packages at /buy/twitter start delivering within minutes of payment, need no password, and are refundable if delivery hasn't started within 24 hours. The rest of the loop is yours to run.
Sources: twitter/the-algorithm (GitHub), X's open-sourced recommendation system including home-mixer and candidate sourcing documentation, and twitter/the-algorithm-ml (GitHub), the published For You heavy ranker models.