X Algorithm 2026: Inside the Open-Source Code (What Actually Matters)

X just open-sourced their recommendation engine. Not the 2023 version everyone keeps referencing — the new one. The one powered by Grok.

Most people are reading Twitter threads about it. I went straight to the source: cloned the repo, opened the Rust and Python files, and traced the logic myself.

Here's what's actually in there — and what it means for your reach.

TL;DR

• The old rules-based system is dead. Grok now predicts your behavior using your last 128 interactions.
• Likes are almost worthless. DM shares and link copies carry serious weight.
• Engagement pods don't work anymore. Each post is scored in complete isolation.
• There's a $1M monthly prize for long-form Articles that most creators don't know about.
• Negative signals hit hard. One block might hurt more than ten likes help.

---

The Old Algorithm Is Gone

In 2023, X used hand-coded rules. An engineer literally wrote logic like: "If a post gets a like, multiply score by X." If you knew the formula, you could game it.

That's history now.

The file phoenix/grok.py reveals a pure transformer model — the same architecture behind Grok itself. It doesn't follow "if this then that" rules. It predicts behavior.

The AI analyzes your last 128 interactions:

• What you clicked
• What you scrolled past
• Who you replied to
• What you ignored

It builds a behavioral map, then asks one question: "Given this person's history, will they engage with THIS specific post?"

No keyword tricks. No hashtag hacks. Just pattern matching at massive scale.

---

How Posts Actually Get Scored

Here's the part everyone gets wrong: there's no hidden "account reputation score." It doesn't exist in the code.

Instead, every post gets scored individually. The AI predicts 19 different actions someone might take:

Positive signals:

• Like
• Reply
• Repost
• Quote tweet
• Share (generic)
• Share via DM ← This one matters
• Copy link ← This one too
• Click on post
• Click author profile
• Expand photo
• Video quality view (50%+ watch time)
• Dwell time (how long you look at it)
• Follow the author

Negative signals:

• Click "Not interested"
• Mute author
• Block author
• Report

These predictions combine with weights. Positive actions add to your score. Negative actions subtract.

Key insight: A post with 100 likes and 10 blocks might rank lower than a post with 50 likes and zero blocks. Negative signals hit hard.

---

Thunder vs Phoenix: Two Different Worlds

The code splits everything into two sources — and understanding this changes how you think about growth.

Thunder handles your followers. When you post, your existing followers see it (ranked mostly by recency).

Phoenix is the discovery engine. This is the "For You" feed for people who don't follow you yet. To appear here, the AI has to predict that complete strangers will engage with your content.

Out-of-network posts get multiplied by a weight factor. If it's less than 1.0 (which it likely is), in-network content has an advantage.

Translation: Your existing followers matter. A lot. Growing genuine followers improves your reach for everything you post.

---

Why Engagement Pods Are Dead

Here's where gaming strategies fall apart.

I found something called "candidate isolation" in the attention mask. When the AI scores posts, each post is scored completely independently. Your post can see the user's history, but it can't see what other posts are being scored alongside it.

This means:

• Coordinated posting doesn't boost you
• Engagement pods can't help each other
• Each post lives or dies on its own merit

The AI asks: "Will THIS person like THIS post?" Not "is there buzz around this post right now?"

The game has changed. You can't manufacture virality through coordination anymore.

---

The Feed Diversity Rule

There's a scorer called AuthorDiversityScorer. It's not about how often you post per day — it's about how many of your posts show up in a single feed refresh.

If someone scrolls and three of your posts are eligible:

• The first gets full score
• The second gets reduced
• The third gets reduced even more

X wants variety in every scroll session. They actively prevent any one person from flooding someone's feed.

What this means: Posting 10 times in rapid succession doesn't get you 10 spots. Quality over quantity wins.

---

Want tactics like this delivered daily? Join 10,000+ indie hackers getting our curated digest of what's working right now. Subscribe to IndieRadar — it's free.

---

Video Has a Hidden Requirement

For video content, I found this logic:

if video_duration_ms > MIN_VIDEO_DURATION_MS:
    apply VQV_WEIGHT
else:
    weight = 0

Short videos don't get the "video quality view" bonus at all. You need to hit a minimum duration threshold.

The exact number isn't public, but if your videos are just a few seconds long, you're probably not getting credit for watch time. The algorithm wants people to actually watch — not just scroll past.

---

The $1M Article Play Nobody's Talking About

X announced a $1,000,000 monthly prize for the best Article. Not a one-time thing — it's recurring.

According to the official announcement, the goal is to reward long-form, high-quality writing. They're also:

• Doubling the creator payout pool
• Focusing payouts on "Verified Home Timeline" impressions

What this means: Views from Premium users count more for monetization. And X is clearly pushing toward long-form content over quick tweets.

I couldn't find an explicit "Article boost" in the code (that part might be server-side), but the incentive structure is obvious: X wants you writing Articles, not threads.

If you've been thread-maxxing, it might be time to consolidate that content into proper long-form pieces.

---

What Actually Works in 2026

Based on what's in the code, here's the playbook:

1. Write for the Share, Not the Like

The AI tracks three separate share types: generic share, DM share, and copy link. That's unusual — they clearly care about content that gets passed around.

Create stuff people want to send to their friends. Posts that get saved for later. Insights worth sharing in group chats. That's the new viral path.

2. Avoid Triggering Negative Actions

One block might hurt you more than one like helps you. Don't be annoying. Don't spam. Don't post stuff that makes people hit "Not interested."

The math is asymmetric. Negative signals punch above their weight.

3. Your Followers Still Matter

In-network content likely has a scoring advantage. Growing genuine followers improves your baseline reach for everything you post.

Stop treating follower count as a vanity metric. It's a distribution advantage.

4. Videos Need to Be Long Enough

Short clips might not qualify for the video quality view bonus. If you're doing video, make it worth watching past the halfway point.

Rule of thumb: If someone wouldn't share the video, it's probably too short to matter algorithmically.

5. Replies to Big Accounts Still Work (But Not Why You Think)

The AI sees your interaction history. If you're replying to high-engagement accounts and getting responses back, that interaction becomes part of your engagement graph.

It shapes what the algorithm thinks you're interested in — and associates you with those conversations.

---

What I Didn't Find

A few things people claim that aren't in the code:

No "shadowban" logic. Posts either pass the filters or they don't. There's no partial suppression based on secret scores.

No article-specific boost multiplier. The incentive is financial ($1M prize, creator payouts), not algorithmic. At least not in the public code.

No "post at 9am for maximum reach" timing logic. Recency matters for Thunder, but there's no magic hour.

No exact weight values. X excluded those "for security reasons." Anyone claiming to know the precise hierarchy (DM shares worth 50x, likes worth 1x, etc.) is guessing.

---

The Pipeline: How It All Works Together

For the technically curious, here's how the system processes your content:

1. Query Hydration — Fetches the user's recent engagement history and metadata
2. Candidate Sourcing — Retrieves posts from Thunder (followers) and Phoenix (discovery)
3. Pre-Scoring Filters — Removes duplicates, old posts, blocked accounts, muted keywords
4. Scoring — Runs the Grok transformer model predictions
5. Author Diversity Scoring — Prevents feed flooding
6. Selection — Top K candidates make it through
7. Post-Selection Validation — Final checks before serving

Each step has specific filters. Deleted posts, spam, violence — all get caught before they reach your feed.

---

FAQ

How do I get more DM shares? Create content people want to send privately. Tactics, insights, and "I thought of you when I saw this" posts. Useful beats clever.

Does posting time still matter? For your followers (Thunder), recency helps. For discovery (Phoenix), it's about predicted engagement, not timing. Post when you can engage with early replies.

Will engagement pods get my account flagged? The code doesn't show explicit pod detection, but the candidate isolation makes pods ineffective anyway. Each post is scored in a vacuum.

Should I stop posting short videos? Not necessarily — but know they might not get the video quality view bonus. If video is your strategy, make them worth watching past 50%.

Is there really no shadowban? Based on the code, no. Posts either pass filters or they're removed entirely. There's no "partial visibility" middle ground in what was open-sourced.

---

The Bottom Line

The algorithm isn't your enemy. It's a prediction engine trying to figure out what people want to see.

Stop gaming. The AI predicts intent, not just surface actions.

Stop spamming. The diversity scorer limits how much anyone can dominate a feed.

Start building things worth sharing. DM shares and link copies are how you go viral now.

The rules have changed. The question is whether you'll adapt.

---

Ready to Ship Smarter?

That's it. The algorithm rewards genuine engagement and punishes noise. If you're building in public, you need to know what's actually working — not last year's tactics.

We send insights like this daily to 10,000+ indie hackers. Join IndieRadar — it's free.