Two posts with the same engagement counts can reach different people. This guide separates the published ranking mechanics from practical writing advice and claims the code cannot establish.
Written by Tushar MehtaUpdated October 18, 2026
Use the earlier repository for historical context, not as a current posting formula.
github.com/twitter/the-algorithmThe repository documents added configuration, filtering systems and model code.
github.com/xai-org/x-algorithm#august-13th-2026X explains why comparing weights as if they were engagement counts produces incorrect conclusions.
github.com/xai-org/x-algorithm#august-14th-2026X says reports now include account or post visibility limits required by law, including country-specific withholding.
github.com/xai-org/x-algorithm#september-18th-2026For You adds a shared list of popular recent posts from the most-followed active accounts.
github.com/xai-org/x-algorithm/commit/b412112Phoenix predicts video watch time and profile visit time. Visibility filtering gains a fourth answer: show the post with a notice.
github.com/xai-org/x-algorithm/commit/e62790cThe like weight moves from 0.5 to 0.66. It scales a viewer's predicted chance of liking, not a post's like count.
github.com/xai-org/x-algorithm/commit/17c96bfX pushed 13 code updates to the public repository between September 30 and October 17, about 85,000 added lines. The repository's Notable Updates section got no new entry for any of them. Most of the new code is video, ads, infrastructure and visibility-rule test data. These are the changes that touch how text posts are found and ranked.
For You now pulls from a shared list of popular recent posts, alongside follows, Phoenix retrieval and SimClusters. The list holds up to 500 posts from the last 24 hours, at most five per author, ordered by views adjusted for the post's age. Its authors are the most-followed accounts among those that posted in the past seven days: the top 0.005% by followers.
This is a route only the largest accounts can use. It adds candidates, not a ranking boost: these posts are scored by the same model as everything else.
Source codeThe like weight went from 0.5 to 0.66 on October 14. Like every action weight, it multiplies one viewer's predicted chance of liking the post.
It does not multiply your like count, and it does not set an exchange rate between likes and replies.
Source codePhoenix now predicts how many seconds a viewer will keep watching a video, and how many seconds they will spend on an author's profile after clicking through. In the published scorer, both weights are 0.0.
They are measured, not ranked on. A claim that profile time boosts reach is not supported by this code.
Source codeVisibility filtering used to return allow, interstitial or drop. It can now return notice: the post stays in the feed with a notice attached. The repository does not include the code that draws the notice.
If a post of yours carries a notice, it was not removed. Check Under the Hood before blaming the writing.
Source codeA like-count holdout that hides up to 15% of posts from a given viewer, scaled by each post's likes. A test that filters posts reaching a viewer only through a retrieval lane for posts with no likes. A video carousel. All three default to off.
X says the defaults in the code are set to the primary production values, and that it aims to publish experiments running on 10% or more of traffic. Off means not the main production setting. It does not mean never.
Source codeNone of this rewards a format, a hook or a posting time. For an account outside the top 0.005%, reach beyond your followers still comes from Phoenix retrieval and SimClusters, which match posts to people by what they engage with.
Imagine one post has 8 replies and 40 likes, while another has 2 replies and 400 likes. Which should you write next? Those totals alone cannot answer that question.
The first might have reached a small group of potential customers; the second might have reached people who will never use your product. Before choosing a format, inspect the conversation and the outcome you wanted.
For a founder, a useful reply can matter commercially without being worth a fixed number of likes in the algorithm.
The system gathers posts from followed accounts, retrieval models and, since October 2026, a shared list of popular recent posts from the most-followed accounts.
Phoenix uses recent engagement history to predict a viewer’s response.
Weights combine predicted actions or values. They do not multiply public engagement counts.
Eligibility to appear is distinct from ordering. Each post gets one of four answers: allow, notice, interstitial or drop. Blocks, mutes and system labels can affect visibility.
A screenshot of likes and replies is insufficient to reconstruct a ranking decision. Compare posts to learn, not to reverse-engineer a guaranteed score.
X’s September 18, 2026 update says its Under the Hood reports include whether an account or post had visibility limited to comply with law. X gives country-specific withholding after a legal demand as an example. If that information is available for your account, check it before treating low reach as a writing or ranking problem.
This is a narrow transparency check, not a complete explanation of distribution. The update covers legal-compliance limits; it does not say that Under the Hood identifies every reason a post may reach fewer people. No reported restriction also does not prove that ranking, audience interest, or other filters caused a particular result.
This is a simplified reading aid, not the full execution order. The published pipeline documents the components and their limits.
Name the person you want to help and what they should understand or do next. A launch post and a technical explanation can serve different jobs even when they discuss the same product.
Instead of “we made onboarding better,” explain which step changed and why it caused friction. Add a measured result only when you have one; an honest decision is useful without a growth number.
If the post depends on a document, demo or study, give readers a usable route to it. Test link placement against visits or relevant responses, rather than assuming a placement trick fixes distribution.
Keep a small record of topic, format, audience and outcome. Compare several similar posts over time. A single unusually strong result is a reason to investigate, not proof that you found a universal rule.
X publishes recommendation code at xai-org/x-algorithm. The older twitter/the-algorithm repository is a historical reference; neither should be treated as a complete snapshot of every live setting.
That is not a valid exchange rate. Action weights apply to predictions for a viewer, not raw engagement totals. Reply and like counts alone cannot tell you which post will rank higher.
A subscription does not establish how a particular post will perform. This guide does not infer a universal reach multiplier from historical code or plan names.
The sources reviewed here do not establish a universal first-hour cutoff. Respond when you can contribute something useful; do not treat a posting schedule as a ranking guarantee.
This review does not establish a universal link penalty or a workaround that guarantees full distribution. Put the source where readers can use it, then compare relevant outcomes for your own account.
The published code added a shared list of popular recent posts from the most-followed accounts, raised the like weight from 0.5 to 0.66, added two predictions weighted at zero, and added a notice outcome to visibility filtering. Several experiments are in the code but switched off by default.
Not in the published code. Since October 2026 Phoenix predicts how long a viewer will spend on an author's profile, but that prediction's weight in the scorer is 0.0.
Bangers Only generates four drafts from an idea. Compare the approaches, check the facts, and edit the version that best serves your reader.
Generate your next post