Guide
The X algorithm in 2027: how the For You feed ranks posts
X publishes the code behind its main feed. Since January 2026 the system that ranks the For You timeline, built around a transformer model, has been on GitHub with a dated changelog. This guide reads that code, X's Help Center and dated news reports to explain how the X (Twitter) algorithm ranks posts going into 2027.
Last updated: 2026-11-08
TL;DR: X's For You feed mixes posts from people you follow with posts found by machine learning, drops anything older than 48 hours, then scores each post by the predicted chance you will like, reply, repost, click, dwell or follow, minus the chance you will mute, block, report or scroll past. X does not publish the weights. Links are not formally penalised, per Musk in July 2026, and Premium buys reply prioritization.
By the Postify team at Businice S.R.L. Postify makes this site, and this guide covers it alongside other options.
How does the X algorithm work in 2027?
X's For You feed gathers recent posts from accounts you follow and posts found by machine learning, filters out ones you blocked, muted or already saw, and ranks the rest with a transformer model. That model predicts how likely you are to take each of a list of actions, and a weighted sum of those predictions sets the order.
X published this system as the xai-org/x-algorithm repository on GitHub in January 2026, after Elon Musk promised the release within seven days. TechCrunch reported that it relies on a Grok-based transformer that learns relevance from your engagement history, with no manual feature engineering for content relevance.
The service that builds the feed is called Home Mixer. Thunder holds recent posts from accounts you follow. Phoenix retrieval finds posts from accounts you do not follow by matching your interests to posts, SimClusters adds posts from communities of similar accounts, and a popular posts source adds what is widely read right now.
The README keeps a dated changelog. Updates in August, September and October 2026 added visibility filtering code, label transparency and, on October 13, 2026, training retrieval to surface posts that people read all the way through.
| Group | Predicted actions | Effect on score |
|---|---|---|
| Engagement | Like, reply, repost, quote, share, share via DM, share via copy link | Raises it |
| Clicks | Open the post, profile, link, photo, video or quoted post | Raises it |
| Attention | Video quality view, dwell, dwell time, click dwell time, video watch time, profile visit time | Raises it |
| Author | Follow the author | Raises it |
| Negative | Not interested, mute author, block author, report, not dwelled | Lowers it |
What is the difference between For You and Following on X?
Following shows only posts from accounts you follow, in reverse chronological order, and X says it is not personalised. For You is the ranked feed: it mixes posts from accounts and Topics you follow with recommended posts from accounts you do not follow, plus promoted posts. The app reopens on whichever tab you used last.
X's Help Center says For You recommendations are based on accounts and Topics you follow, how popular a post is and how people in your network interact with it. Its recommender page adds posts you like, posts liked by people in your network and accounts they follow, and says a neural network is continuously trained on your likes, reposts and replies.
The current open-source code applies a factor below one to posts from accounts you do not follow, so posts from people you follow start with an edge. The older twitter/the-algorithm repository, released in 2023, said about half of For You posts came from in-network sources. The 2026 code gives no such share, so treat that older figure as out of date.
What signals rank posts on the X algorithm?
Predicted engagement and attention. Each post gets a probability for likes, replies, reposts, clicks, dwell time, video watch time and follows, and these add to its score. Predicted mutes, blocks, reports, not interested taps and quick scroll-pasts subtract. X keeps the actual weights in a parameter file and does not publish their values in the README.
The README warns against a common misreading. It gives the example of a report carrying 468 times the weight of a like and says it would be wrong to conclude that one report cancels out 468 likes, because weights multiply your own predicted probability of each action, not raw engagement counts. Do not read that example as a published production weight.
After the weighted sum, three adjustments apply. Author diversity shrinks the score of each further post from the same author in your feed, down to a floor. Posts from accounts you do not follow, and replies and reposts from accounts you do, are multiplied by a factor below one. Authors with impressions below a threshold get lifted toward a target position, a boost for new accounts.
Before scoring, filters remove duplicates, your own posts, posts from blocked or muted accounts, posts with muted keywords, posts you have already seen and posts older than 48 hours. A separate visibility system can allow, label, put behind a warning or drop a post.
- Positive: likes, replies, reposts, quotes, shares by DM or link, profile and media clicks, dwell time, video watch time, follows.
- Negative: not interested, mute, block, report, and scrolling past without dwelling.
- Hand-set rules: author diversity decay, out-of-network discount, new-author boost.
- Hard filter: posts older than 48 hours are not candidates for For You.
Do links get less reach on X?
X says links are not penalised as such. On July 29, 2026, Musk replied that X had not penalised posts with links for over a year. The open-source code has no explicit link rule, but a link sends people off X, so link posts can earn less of the engagement and dwell time the model rewards.
This has changed more than once. In October 2025 X's head of product, Nikita Bier, acknowledged that creators complain link posts get lower reach, and X began testing an in-app browser that keeps the post and its reply buttons visible while people read. Musk said then that a link with almost no description gets weak distribution.
The practical answer is the same under every version: write a caption that is worth reacting to on its own, and treat the link as a bonus for people who want more. Link click is one of the positive predicted actions in the current code, so a link people actually open is not wasted.
What are the best practices for the X algorithm in 2027?
Write posts people stop to read, reply to and share, and avoid anything that makes them mute or report you. Post while your audience is online, because For You only considers posts from the last 48 hours. Spread posts out rather than bunching them, since each further post from you in the same feed scores lower.
Dwell time, read-through and video watch time are all predicted actions, and X's October 2026 updates improved how the model estimates time spent and trained retrieval on posts people read to the end. A post that earns attention without a like still scores.
Negative signals are the expensive ones. Engagement bait that people mute, block or mark not interested lowers your posts for those viewers. Stay on the topics your followers came for: retrieval matches posts to the interests a viewer has shown.
Replies count twice: as a predicted action on your post, and as a way to be seen in other people's conversations. If you are new, the new-author boost exists, so posting steadily in your first weeks matters more than a viral attempt.
- Open with the point; the opening is what stops the scroll.
- Give link posts a caption that stands on its own.
- Reply to every good reply in the first hours after posting.
- Use images and video that people will open or watch through.
- Spread posts across the day instead of posting several at once.
Does scheduling X posts hurt reach?
Nothing in X's open-source ranking code or Help Center treats a scheduled post differently from one typed in the app. What matters is timing: a post only competes in For You for 48 hours, so scheduling it for when your audience is active, and replying once it is live, is what helps.
Postify publishes to X as one of its ten networks, with text, photo, video and album posts and a separate caption for each network in one post. Its MCP server lets an assistant such as Claude schedule X posts and threads, add a first comment and read post metrics.
The honest limits: Postify's inbox does not include X DMs. X replies and comments come into the inbox on Plus and higher plans, and Postify fetches X replies only when asked, because X charges for each reply returned. For posting time, Postify's free best time to post tool covers X.
Frequently asked questions
- Is the X algorithm open source?
- Yes. X published the code behind the For You feed as xai-org/x-algorithm on GitHub in January 2026 and keeps a dated changelog there. X says some parts are withheld, such as certain prompts and some rule sets. The older twitter/the-algorithm repository from 2023 describes the previous system.
- Is the X Following feed chronological?
- Yes. X says the Following timeline shows only accounts you follow, in reverse chronological order, and is not personalised.
- How long does a post stay in the X For You feed?
- The current open-source code filters out posts older than 48 hours before ranking, so a post competes for For You placement for about two days.
- What is the character limit on X?
- Standard posts are limited to 280 characters. X Premium subscribers on every tier can write longer posts of up to 25,000 characters, and anyone can read them.
- Does X penalise posts with links?
- Musk said on July 29, 2026 that X had not penalised posts with links for over a year. Link posts can still get less engagement because people leave X to read them, so write a caption that works on its own.
- How much does Postify cost?
- Postify plans start at €9.99 a month, or €8.33 a month on yearly billing, which gives 2 months free. The top plan, Agency, is €199.94 a month.
Sources
- GitHub: xai-org/x-algorithm README, X For You feed algorithm (changelog to October 26, 2026) (checked November 1, 2026)
- GitHub: twitter/the-algorithm README, earlier recommendation algorithm (2023) (checked November 1, 2026)
- X Help Center: About your For you timeline on X (checked November 1, 2026)
- X Help Center: For You Home Timeline recommendations (checked November 1, 2026)
- X Help Center: About X Premium (checked November 1, 2026)
- TechCrunch: X open sources its algorithm (January 20, 2026) (checked November 1, 2026)
- Social Media Today: X tests in-app links to address link post penalties (October 19, 2025) (checked November 1, 2026)
- PPC Land: X drops year-old link penalty, quoting Musk's post of July 29, 2026 (checked November 1, 2026)
- Buffer: X Premium review, 18.8 million posts analysed (October 2, 2025) (checked November 1, 2026)
- Postify: how to post to X (Twitter) (checked November 1, 2026)
- Postify: X (Twitter) MCP server (checked November 1, 2026)
- Postify pricing (checked November 8, 2026)