# How the X Algorithm Works in 2026: The Practical Guide for Creators

By Pawnie (@pawnie_ on X). Growth on X. Published 2026-10-29. Canonical page: https://www.w3match.com/blog/x-algorithm-2026-for-you-guide

Learn how the X For You algorithm ranks posts in 2026, what the October Phoenix updates mean, and how to improve your content without chasing myths.

The X algorithm does not give every post one universal score. For each viewer, the For You system finds candidate posts, predicts how that person might respond, and ranks eligible posts using those predictions and other rules.

In October 2026, X announced two changes: its ranking model got better at predicting time spent on a post, and its retrieval model was trained to surface posts people read all the way through. The practical opportunity? Make posts worth reading and easy to pass along. The source code does not promise a reach boost for any particular format. [Source: X algorithm repository](https://github.com/xai-org/x-algorithm).

Earlier this year, I published [7 Growth Strategies for the New X Algorithm](https://www.w3match.com/blog/7-growth-strategies-for-the-new-x-algorithm-phoenix-ranking). Since then, X has opened more of the code. Some of my earlier conclusions need correcting: I treated estimated weights as if they explained actual engagement counts and suggested there were no manual rules. The public code now shows explicit weights and filters. A block cannot be converted into a fixed number of lost likes.

Here is what the newer code actually tells us — and what you can test as a creator.

## How does the X For You algorithm work?

Think of it as four steps:

1. **Find candidates.** Thunder supplies recent posts from accounts you follow. Phoenix retrieval and SimClusters help find posts from outside your network. A popular-post source adds candidates too.
2. **Filter.** The system can remove duplicates, posts you have already seen, or content affected by blocks and mutes.
3. **Rank.** Phoenix predicts possible actions for a specific viewer and post, including likes, replies, shares and negative feedback. A scoring stage combines predictions with configured weights.
4. **Select and check visibility.** Additional adjustments and diversity selection shape the feed. A separate visibility system can allow a post, place it behind a warning or drop it for a viewer.

This is a simplified description of the public For You pipeline. Search, replies, profiles and other surfaces have their own behavior. [Source: repository overview](https://github.com/xai-org/x-algorithm).

## What changed in the X algorithm in October 2026?

On **October 8**, X said Phoenix ranking had improved at estimating how long a viewer will spend on a post. On **October 13**, it said Phoenix retrieval was also trained to surface posts people read all the way through. On **October 26**, X removed its Brazil 2026 election filter from For You. [Source: official update notes](https://github.com/xai-org/x-algorithm#notable-updates).

The distinction matters. Retrieval decides which posts enter the candidate pool; ranking helps order them. A better time prediction and an additional retrieval training target are not proof that longer posts always win.

A short post can deliver its value in seconds. A long one can waste two minutes. Match the length to the insight.

## X algorithm weights: what do they actually mean?

The [public parameter file](https://github.com/xai-org/x-algorithm/blob/main/home-mixer/params/param.rs) showed these defaults when checked on October 29:

| Predicted action                       | Default weight                 |
| -------------------------------------- | ------------------------------ |
| Share via copy link                    | 20                             |
| Reply, quote or share via DM           | 5 each                         |
| Follow author                          | 4                              |
| Share / repost / like                  | 2 / 1 / 0.66                   |
| Click post / open link                 | 0.3 / 0.2                      |
| Not interested / block / mute / report | −52.272 / −31.2 / −58.8 / −234 |

These numbers multiply **predicted probabilities for a particular viewer**, not counts on your post. X also uses predicted continuous values for some time signals. Values are configurable and experiments can differ; the public defaults are synchronized with primary production settings according to X, not necessarily every user's settings.

Consider a simplified example. A 10% predicted chance of a like multiplied by 0.66 contributes 0.066\. A 0.2% predicted chance of copying the link multiplied by 20 contributes 0.04\. In that example, the larger weight produces a smaller contribution.

So no: **one copied link is not automatically worth 30 likes**, and **one report does not cancel a fixed number of likes**. Copy-link's high weight is interesting, but it is not a reason to turn every post into a request to share.

## 7 practical ways to improve your content for X

### 1\. Make the post understandable to someone who doesn't know you

A stranger seeing your post has no context for your last five posts. Name the topic and explain why it matters.

Weak: “This changes everything.”

Better: “X now trains retrieval to find posts people finish reading. Here is what that could mean for a Web3 creator.”

When you share a chart or screenshot, point to the number or detail the reader should notice. That helps a person understand the takeaway; it is not a guaranteed model-ranking trick.

### 2\. Create something a person would send to a friend

Ask: who would benefit from receiving this post?

A useful checklist, a comparison of two tools, or a clear explanation of a protocol update gives someone a reason to share it. For Web3, a post that explains the risks and links to the original announcement can be more useful than repeating “huge news.”

Don't confuse that editorial idea with proof that every copy-link action causes more distribution.

### 3\. Deliver the payoff your opening promises

The October changes make attention and completion worth testing. Start with a clear question, show the evidence, and get to the answer. Cut repetitive setup.

For example, instead of a 15-post thread listing every detail of an update, try a five-point breakdown: what changed, who it affects, what stays uncertain, where to verify it, and what to do next. If the subject needs more depth, give it the space it needs.

### 4\. Build real conversations

The public scoring configuration includes an extra reply weight for eligible original posts from mutual follows. That affects the model's **prediction** that a particular mutual will reply; it is not a multiplier on the replies you already have. X's [deployment history](https://github.com/xai-org/x-algorithm/blob/main/docs/BIDIRECTIONAL%5FBOOST%5FCHANGE.md) also shows that this setting has changed and been tested.

A practical response: talk with people in your field, respond when you have something to add, and ask questions you actually want answered. Mass follow-backs or generic “great post” replies are not proven growth hacks.

### 5\. Give original posts a place in your mix

The public For You pipeline describes a filter that can exclude replies and reposts from people a viewer does not follow. That applies to this recommendation path, not every place a reply appears. Replies can still be discovered in conversations.

Try turning a useful answer you wrote into a standalone post with context and a source. Keep replying where the conversation is valuable.

### 6\. Don't confuse post frequency with a hard penalty

The system includes an author diversity adjustment for multiple candidates from the same author in a viewer's feed. That does not establish a daily post limit or a rule to wait exactly two hours between posts.

If you publish several similar posts close together, test whether they reach the people you want. Compare results before changing your schedule.

### 7\. Check trust and visibility before blaming a shadowban

The ranking model predicts negative feedback, and separate visibility rules can restrict content in some contexts. Low impressions alone do not tell you which mechanism, if any, was involved.

If available for your account, check [Under the Hood](https://x.com/i/jf/under%5Fthe%5Fhood), X's pilot reporting tool for visibility-impacting labels. Its reports are aggregated and availability can vary. Then review the post itself: is the claim supported? Is the link explained? Could the opening feel misleading? For Web3 posts, clearly separate analysis from promotion and disclose partnerships.

## Does the X algorithm help small accounts?

The public code includes a cold-start mechanism for some original posts from smaller accounts. The checked defaults include an author threshold of 50,000 followers, fewer than 200 Home impressions and a post age under two hours, alongside other conditions and configuration. It selects an eligible candidate within a feed request; it does not promise a boost for every qualifying post. [Source: cold-start scorer](https://github.com/xai-org/x-algorithm/blob/main/home-mixer/scorers/author%5Fcold%5Fstart.rs) and [parameters](https://github.com/xai-org/x-algorithm/blob/main/home-mixer/params/param.rs).

For a small creator, the useful lesson is to make original posts that give a new reader enough context to care. Measure the result rather than assuming the boost applied.

## Do posts stop getting reach after 48 hours?

The public For You pipeline describes an age filter for posts older than 48 hours. That is a limitation of this path, not a platform-wide expiration date. A post can still exist on your profile, in search or in a conversation. [Source: official filtering documentation](https://github.com/xai-org/x-algorithm#filtering).

## A 14-day experiment you can run

Pick one audience and two repeatable formats. For example: a short Web3 news explanation and a practical checklist. Publish multiple examples of each, covering the same or closely related topics where possible. Record results at the same interval after publication, for example 48 hours.

Record each post's date, format, topic, impressions, shares, substantive replies, profile visits and follows when those numbers are available. Compare the median per post and shares or follows per 1,000 impressions. Note differences in topic and timing: this is an observational test, not proof that the format caused the outcome.

If X does not show you copy-link counts, completion or dwell, mark them as unavailable. A like or reply is not a direct measurement of reading time.

For a broader historical playbook, see [73 Practical Rules for Web3 Creators](https://www.w3match.com/blog/73-practical-rules-for-web3-creators-from-xs-new-algorithm-repo). Some recommendations there predate the newly disclosed weights and should be read alongside this update.

## FAQ

### Does X reward long posts?

The October updates concern time prediction and retrieval training for posts read through. They do not establish that long posts always beat short posts. Give the topic the length it needs.

### Is copying a link the most important X algorithm signal?

It has a high published default weight, but actual contribution depends on the predicted probability for a viewer and the rest of the system. The code does not establish a universal most important action.

### Do external links hurt reach?

The checked configuration has a positive open-link weight. That alone does not prove links boost or reduce overall reach: other predictions, filters and viewer behavior matter. Test useful links with context.

### Can I calculate how many likes cancel a report?

No. The published weights apply to predictions, not raw counts. X explicitly warns against that calculation.

### Is the entire X algorithm public?

No. The repository describes the For You pipeline and says some prompts and anti-abuse rules are withheld. Settings can also vary by experiment.

**Sources:** [Official X algorithm repository](https://github.com/xai-org/x-algorithm), [scoring parameters](https://github.com/xai-org/x-algorithm/blob/main/home-mixer/params/param.rs), [mutual-follow experiment history](https://github.com/xai-org/x-algorithm/blob/main/docs/BIDIRECTIONAL%5FBOOST%5FCHANGE.md), [Under the Hood](https://x.com/i/jf/under%5Fthe%5Fhood). Checked October 29, 2026\. This analysis used AI assistance and was checked against public documentation; the public `main` branch can change.