X Just Open‑Sourced Their “For You” Algorithm. Here’s What Marketers Should Actually Do
If you’ve ever said, “I don’t get X… sometimes my posts fly, sometimes they die,” you’re not alone.
For years, the “For You” feed has been a black box.
Now it’s not.
X has published a public repo that lays out the core recommendation system powering the “For You” timeline. It’s a real look at how posts are sourced, filtered, scored, and assembled.
So… can you “crack the code”?
Yes, but not in the way people think.
This isn’t a cheat sheet with a single magic weight you can game. It’s a blueprint of the machine. And once you understand the machine, you stop guessing and you start designing content with purpose.
Let’s break down what’s in the open-source release and translate it into a clean, practical marketing playbook.
What X actually open‑sourced (high level)
The repo describes a “For You” system that blends two types of content:
-
In‑network posts (from accounts you follow)
-
Out‑of‑network posts (discovered via ML-based retrieval across a global corpus)
Then it ranks those posts using a Grok-based transformer model that learns “relevance” from your engagement history.
And here’s a line that matters a lot for marketers:
The final score is a weighted combination of predicted engagement actions.
Meaning: it’s not “one relevance score.” It’s a bundle of predicted behaviors, rolled up into a single number.
The four big components (the “engine parts”)
Your analysis nailed the architecture. The repo itself describes four major pieces:
-
Home Mixer: the orchestration layer that assembles the feed using a pipeline (sources → hydrators → filters → scorers → selector).
-
Thunder: in-network candidate sourcing for posts from followed accounts.
-
Phoenix: ML retrieval + ranking using a transformer architecture (ported from Grok-1) with “candidate isolation.”
-
Candidate Pipeline: the reusable pipeline framework the system is built on.
If you’re a marketer, here’s the real translation:
The algorithm isn’t one thing. It’s a pipeline.
And pipelines have gates.
The feed is a nightclub, and filters are the bouncers
Before your post is ever “ranked,” it has to be eligible.
That sounds obvious… until you realize how many posts die before scoring.
TechCrunch’s overview of the system (based on the same repo) points out that the pipeline filters out content from blocked accounts, muted keywords, and content flagged as spam-like/too violent before it decides what to show.
Your source review goes deeper and shows practical “eligibility killers” that are worth treating as a pre-flight checklist:
-
Missing/empty text (in this implementation)
-
Aged-out posts (freshness matters)
-
Already-seen / already-served dedupe
-
Muted keywords
-
Social-graph exclusions (blocked/muted authors)
Marketing takeaway:
If you’re filtered out, your creative doesn’t matter because you never get ranked.
So the first playbook rule is simple:
Don’t publish posts that look like they belong in the filtered pile.
That means:
-
Add meaningful text context (even on video/image posts)
-
Avoid spam patterns (repetitive hashtags, bait loops, “DM me” traps that annoy people)
-
Stay brand-safe and audience-fit (muting is the silent killer)
-
Stop trying to “win attention” at the cost of negative feedback (more on that in a second)
The scoring model predicts many actions, not just likes
Here’s where most “how to go viral” advice falls apart.
Phoenix doesn’t just predict likes. In the open-source runners file, there’s a list of action types the model is explicitly set up to predict including negative feedback actions.
In the repo, the action list includes (among others):
-
favorite_score (like)
-
reply_score
-
repost_score
-
share_score
-
share_via_dm_score
-
share_via_copy_link_score
-
click_score
-
profile_click_score
-
follow_author_score
-
dwell_score and dwell_time
-
not_interested_score
-
block_author_score
-
mute_author_score
-
report_score
That’s the biggest “aha” for brands:
The system is optimizing for a portfolio of behaviors and it’s explicitly aware of negative signals.
So if your content drives clicks but also drives “not interested” or “mute,” you’re not “winning.” You’re training the system that your stuff is low quality for that audience.
Candidate isolation: you don’t get to “ride along” with someone else’s post
Phoenix’s documentation also calls out a design decision that’s easy to miss:
Candidates cannot attend to each other during ranking inference.
Translation: Your post isn’t scored higher because it appears near a viral post in the same batch. Your score is about that user + your post + that user’s history.
Marketing takeaway:
Stop chasing the trend of the day if it doesn’t match your audience.
Chasing trends can get you impressions, but it can also earn negative feedback (and the system is watching).
OK… so what should marketers do?
Here’s the simplest way to turn this into a usable strategy:
Build content that is likely to earn high-intent actions (and avoids negative ones)
Likes are nice. But the action list shows higher-intent behaviors exist:
-
Shares (DM / copy link) = “this is valuable enough to send”
-
Follows = “I want more of this”
-
Dwell time = “I stayed”
-
Replies = “this sparked a conversation”
-
Profile clicks = “I’m curious who you are”
So instead of making every post a generic “announcement,” design posts with a single primary objective.
An “action-based” creative playbook (steal this)
Below are examples that map directly to behaviors the model is built to predict.
1) Want shares (DM / copy link)?
Make it immediately useful.
Content patterns:
-
Checklists (“7 things to fix before you scale paid social”)
-
Templates (“copy/paste this outreach message…”)
-
“Send this to your team” framing
-
Simple visuals people save/share (one chart, one point)
2) Want replies?
Make it easy to respond.
Content patterns:
-
“Pick one: A or B (and why)”
-
“What’s your biggest challenge with X right now?”
-
“Hot take (brand-safe)… agree/disagree?”
Then (this part matters) respond quickly to early replies. Replies create more replies.
3) Want follows?
Make the account feel like a series, not a one-off.
Content patterns:
-
“Weekly teardown”
-
“Daily 60-second lesson”
-
“Part 1/3” content (but only if Part 1 is valuable on its own)
4) Want dwell time?
Tell a story, or build a thread that pays off.
Content patterns:
-
“Here’s what we changed, what happened, and what we learned”
-
Strong pacing: setup → tension → payoff → lesson
-
Video with captions + hook in the first 1–2 seconds
5) Want clicks (without getting penalized)?
Make the click make sense.
-
Clear CTA
-
Curiosity, yes; deception, no
-
Don’t bait people into irrelevant landing pages (that’s how you earn “not interested” and mutes)
In-network vs out-of-network: this is just AOK Game Theory in disguise
At AOK, we teach the funnel as phases: Cold → Warm → Hot → Converted.
The X algorithm structure maps beautifully:
-
In-network (followers) = Warm/Hot traffic (they already know you)
-
Out-of-network (recommendations) = Cold traffic (they don’t know you yet)
And just like any funnel:
Cold traffic needs trust + value fast
That means: utility, clarity, and zero spam vibes.
Warm/hot traffic can handle more nuance
That means: deeper POV, insider knowledge, product stories, behind the scenes.
This is why some accounts “grow” but don’t convert, and others convert but don’t grow:
-
Growth content wins cold traffic (shares, follows)
-
Conversion content wins warm/hot traffic (clicks, DMs, pipeline)
Your job is to design both… on purpose.
The “don’t do this” list (a.k.a. how to get muted)
If you only take one thing from the action list, take this:
Mute/block/report are explicitly modeled.
You don’t need a big percentage of these to hurt distribution.
So avoid:
-
Repetitive posting that feels like noise
-
Copy/paste hooks every day
-
Hashtag soup
-
Aggressive engagement bait
-
Off-topic swerves that confuse your followers
In other words:
Don’t win today’s impressions at the cost of training tomorrow’s suppression.
What you can’t learn from the repo (and what to do anyway)
Even with the open-source code, you won’t get everything:
-
Exact weights
-
Exact thresholds
-
All production integrations
Phoenix even notes that the code is representative of the internal model, but excludes specific scaling optimizations.
So how do you act like a pro anyway?
Run controlled experiments by “action objective”
Pick one objective per week:
-
Week 1: Reply-driven posts
-
Week 2: Share-driven posts
-
Week 3: Follow-driven series
-
Week 4: Dwell-driven threads/video
Track:
-
Reach / impressions
-
Follows per impression
-
Shares per impression
-
Replies per impression
-
Negative feedback signals (where available)
This turns “posting” into a measurable growth system.
The AOK “7-day X Algorithm Sprint” (simple, effective)
If you want a fast way to put this into motion, here’s a sprint plan:
Day 1: Audit your last 30 posts
-
Identify which posts earned shares, follows, replies, clicks
-
Identify which topics earned nothing
Day 2: Pick 2 content pillars
-
The system learns from engagement patterns; consistency matters.
Day 3: Write 3 share-first posts
-
Checklists/templates/tools
Day 4: Write 2 reply-first posts
-
“Pick one” / “What would you do?” / “Biggest challenge?”
Day 5: Write 1 follow-first post
-
Launch a simple series: “Every Friday we share X”
Day 6: Publish + engage hard for 60 minutes
-
Reply to every legit comment early
Day 7: Review results + double down
-
Keep what earned shares/follows/dwell
-
Cut what earned nothing (or felt spammy)
Bottom line
The open-source release doesn’t give you a “hack.”
It gives you something better:
A clear picture of what the system is trying to do:
-
Pull content from in-network and out-of-network sources
-
Predict multiple engagement and feedback actions
-
Combine those predictions into a final score
-
Filter aggressively for safety, spam, and preference constraints
Your marketing strategy, in plain English:
Be eligible, be valuable, optimize for high-intent actions, and avoid negative feedback.
About The Author
