好文集ChatGPT Plus¥119

The 𝕏 Algorithm Explained, In Depth

文章目录

How the 𝕏 algorithm (actually) works, what it rewards, and the best practices to increase your reach. Tag your @bot to digest this entire breakdown.

Everything in today's article comes directly from the open-source code and my experience, so it's practical.

Why listen to me?

My entire job is to hack social media algorithms. On 𝕏 in particular, I have 1B+ lifetime impressions, and I've created and pioneered many of today's viral strategies.

Including: article + QT strategy, article + GIF meta, infographic meta & many more.

Not to mention, over the past 4 weeks, I started my own brand and have already grown to 8,000+ followers and generated 4M+ impressions - purely solo *(no big accounts boosting, engagement groups, etc.) *

文章配图

3M analytics (starting posting ~4W ago)

With that said, I am NOT from the 𝕏 algorithm team. This is purely my interpretation of the code.

文章配图

Check it for yourself: https://github.com/xai-org/x-algorithm

This article is pretty dense. For a condensed version, I recommend screenshotting this post and watching my video below at 2x speed.

嵌入推文(联网查看)

Your roadmap:

  • How to think about algorithms

  • The three primary 𝕏 algorithms: Thunder, Phoenix & SimClusters

  • The Judge: how posts get scored & weighted

  • The Bouncer: filters, account reputation & Under the Hood

  • Myths, busted

  • Random quirks

  • Practical advice


How to think about algorithms

A little bit of context for how I think about social media: at the end of the day, algos are essentially games and math. Especially the 𝕏 algorithm.

An algo's entire job is to "match" content with the right demographic to keep users engaged and on the platform. At any given time, a social media algorithm pushes various content, topics, hooks, and messaging angles based on user demand and data.

When people say, "*Algos don't matter, just post whatever you want!" *they're fundamentally ignoring the entire reason algorithms exist.

So no, you can't just post whatever you want. It doesn't work that way. You have to reverse-engineer your content to actually get pushed.

At a high level, here's how the 𝕏 algo works:

  1. The Scout: when someone opens 𝕏, three scouting algorithms go out and grab posts based on what the algorithm thinks that person might like.

  2. The Judge: 𝕏 guesses what each person will do with each post (like, reply, share, follow, scroll past, mute, etc.) and gives it a weighted score.

  3. The Bouncer: the highest-scoring posts get one final filter

Once a post passes all three stages, it lands on the For You page. That's where posts go really viral.

文章配图

Now let's break down each stage:


The Three Primary Algorithms

𝕏 runs three algorithms. I call them the Scouts. Their job is to find posts based on what the algorithm thinks each person might like.

1. Thunder (your home base)

Thunder pulls posts from people you follow. If you think about the way 𝕏 works, people are essentially put into little bubbles. There's a certain group of people you probably interact with on a regular basis, and that's your bubble. When you open up your For You timeline fresh into the app, Thunder pulls posts from these people.

2. Phoenix (where the virality lies)

Phoenix is a very interesting algorithm, and it's where this platform's virality really lies. It's essentially an algorithm that predicts what you'll want to see before you've ever followed the person who posted it.

3. SimClusters (your niche)

Outside of that, you have SimClusters: posts that are popular in your niche. SimClusters group accounts and posts by who engages with what. Basically, communities of people with shared interests. For example, I wrote an article on my page about Head of Content, which went viral a few weeks ago. It wasn't a platform-wide viral article, but it went viral within the tech startups, marketing, and content niche.

(Bonus: there's also a fourth, smaller source called Popular Posts, a shared list of popular recent posts, initially seeded from the most-followed active accounts.)

文章配图


The Judge: How Posts Get Scored & Weighted

Once the Scouts grab posts, they hand them to what I call the Judge.

For every post and every viewer, the Judge (Phoenix's ranking model, built on the same architecture as Grok) predicts how likely that person is to like it, reply, share it, read the whole thing, follow you, and so on.

It's essentially a gamified points and reduction system:


The actual weights

These are the production defaults as of October 30, 2026:

Positive actions
Share via copy link     20.0
Reply                   5.0
Quote                   5.0
Share via DM            5.0
Follow author           4.0
Share                   2.0
Repost                  1.0
Like                    0.66
Click into post         0.3
Open link               0.2
Dwell                   0.05
Photo expand            0.05
Profile click           0.0
Video quality view      0.0

Negative actions
Not dwelled (scrolled past)    -0.02
Block author                   -31.2
Not interested                 -52.3
Mute author                    -58.8
Report                         -234.0

文章配图

What this tells us:

  • Shares are king (again, 𝕏 wants people on the platform).

  • Conversations beat likes. Replies and quotes are weighted 7.5x more than a like.

  • Likes barely move the needle.

  • Negative feedback is absolutely brutal.

The adjustments: After a score is calculated, a few more things can push it up or down:

1. Posting too much: If you post too many posts, the algorithm basically doesn't give all of them a chance to push out.

2. Out-of-network discount: Posts from accounts you don't follow get their score multiplied by a factor below

3. The small account boost: This is actually super cool: small accounts get a bit of a boost now. I noticed this almost immediately. I started posting on 𝕏 about four weeks ago, and within three days I had my first viral post.

嵌入推文(联网查看)


The Bouncer: Filtering & Under the Hood

So the Scouts pull posts, the Judge scores them, and the "Bouncer" filters them.

Once a post passes all three stages, I interpret that as the highest possible chance of landing on the For You timeline.

A few categories of posts are consistently filtered out on 𝕏:

  • Posts older than 48 hours: 𝕏 is real-time. You generally have a 48-hour window to make a post pop.

  • Posts you've already seen.

  • Reposts and replies from people you don't follow.

  • Blocked accounts, muted accounts, and muted keywords.

  • Spam, flagged accounts, and bots

Account reputation

• Blocks and reports vs. likes: too many can get your account labeled.

• Account credibility: a PageRank-style score based on who follows and engages with you. Network quality matters (not just size).

• Inauthentic behavior: 𝕏 tracks your actions over time

Under The Hood

𝕏's transparency tool shows labels limiting your account's visibility, including posts marked as spam

嵌入推文(联网查看)

How to stay on the right side of the Bouncer:

  • Don't use AI to write and post 100% of your copy.

  • Get Premium, and verify your ID if you're comfortable.

  • Avoid negativity and content that gets you muted, blocked, or reported.


Myth Busters & Random Quirks Sections

Myths, busted

  • "Profile clicks are a huge signal." Not anymore. Profile clicks are currently weighted at 0. Conversions are what count.

  • "Likes are what matter most." A like is one of the weakest positive signals (0.66). Shares, replies, quotes, and follows matter far more.

  • "Big accounts get guaranteed reach." Every account has to fight to stay relevant. I've seen many big accounts die over the last ~18 months.

  • Small accounts can't go viral." False. Small accounts get a built-in cold-start boost on fresh posts.

  • Engagement groups no longer have the same effect they once did

Random quirks

  • Your feed is built from roughly 2,000+ candidates every time.

  • The algorithm reads your last ~1,000 actions. Your recent engagement history (likes, replies, shares, etc.) is the main input the model uses to predict what you'll want to see.

  • The algorithm uses signals like your installed apps, inferred gender, and location (via IP) to personalize your feed.

  • The weights often change. The like weight moved from 0.5 in August to 0.66 now, with no announcement.


Practical Advice

1. Make content people actually want to share

  • Remember, the algorithm wants to keep users on the platform. Shares, replies, and quotes are the most valuable actions.

  • Create posts and formats that are easy to scan and share.

  • Spark conversation. You want people replying in your comments.

  • Provide real value and original insights from your experience, your network, your job. Basically, your unfair market advantages.

2. Grab attention fast

  • Learn how to write hooks.

  • Learn copywriting.

  • If you're mainly a video creator, invest in a proper setup (camera, mic, lighting).

3. Post original content

  • Never copy someone one-for-one. Credit others, and use the Post Image/Video feature that 𝕏 offers,

  • Don't be reactive to what other creators are doing. Stay on the cutting edge.

4. One idea per post

  • If a post has multiple ideas, split it into multiple posts.

  • Every post needs a clear takeaway, a specific person it's for, and a specific CTA.

  • Mixed topics make it hard for the algorithm to know who to show it to, especially within SimClusters.

5. Avoid negativity

  • I personally recommend staying away from negativity and political content.

  • Don't go back and forth with trolls. Block or mute them instead. Arguing risks them muting or blocking you, which suppresses your reach.

6. Build a dense network of high-quality individuals

  • Engage genuinely with people in your niche. It strengthens your reach and your account credibility.

7. Test every format 𝕏 is pushing

  • Threads, articles, media, etc.

Closing

So, there you have it: the full decoding of the open-source 𝕏 algo. Note that 𝕏 is obviously keeping some aspects closed. Again, this is purely my interpretation.

If you found even one valuable point from this piece, a like, repost, and follow go a long way.

Every week, I write 1-2 articles on this account covering: GTM strategy, content, marketing, AI & related topics. If that sort of stuff floats your boat, let's connect.

I also post whiteboard-style videos that some people have liked recently.

Feel free to watch one:

嵌入推文(联网查看)

Keep crushing 👊🏼 - Ryan

全文完
查看原文

沿着这个主题继续读

从一篇文章,找到相关的教程、经验和阅读路径。

自媒体与内容增长用 AI 做自媒体:X、小红书、公众号与视频号的起号、涨粉与变现经验。