X Doesn’t Optimize for Likes. It Optimizes for Intent.

A lot of growth advice on X still sounds like it came from the era when "engagement" mostly meant likes.

That is not what the public recommendation repo suggests.

If you read the ranking pipeline closely, the stronger interpretation is this:

X is not optimizing for one shallow metric. It is ranking for a portfolio of predicted actions that reveal intent.

That distinction matters.

Because if the system is trying to predict whether someone will reply, repost, quote, click, expand, dwell, visit your profile, follow you, or share the post privately, then the winning strategy is different from "write like-bait."

It becomes: create posts that make people do something meaningful.

What the repo directly shows

The local ranking code combines multiple predicted actions into one weighted score.

In weighted_scorer.rs, the scorer includes signals such as:

• favorite / like
• reply
• retweet / repost
• photo expand
• click
• profile click
• video quality view (vqv_score)
• share
• share via DM
• share via copy link
• dwell
• quote
• quoted click
• continued dwell time
• follow author

It also includes negative signals:

• not interested
• block author
• mute author
• report

That is a much richer objective than "maximize likes."

The repo also shows a few other important things:

1. Filtering happens before scoring matters

Candidates can be removed for being duplicate, too old, previously seen, previously served, blocked, muted, keyword-muted, ineligible, or visibility-filtered.

So some posts do not lose because they scored poorly. They lose because they never meaningfully made it into the race.

2. In-network advantage is still real

The pipeline blends in-network content with out-of-network retrieval, and the out-of-network scorer explicitly downweights out-of-network candidates relative to in-network ones.

So yes, discovery matters. But your existing audience graph still has structural importance.

3. Author flooding gets damped

The author diversity scorer attenuates repeated-author scores within a single response.

Translation: posting five times in a burst does not mean you get five equal shots in the same feed response.

4. Video is not automatically favored

The weighted scorer only applies the video-quality-view weight when the video duration clears a minimum threshold.

So video seems to need enough substance to qualify for that extra path.

What that likely means for creators

If the platform is modeling a basket of intent-rich actions, then the highest-leverage content is not content that gets the fastest reaction.

It is content that creates a chain reaction.

A strong post often does several things at once:

• gets the initial stop
• creates enough curiosity to earn a click or expand
• gives enough substance to earn dwell
• has enough edge or relevance to trigger a reply, repost, or quote
• makes the viewer interested enough in you to visit the profile or follow

That is a very different bar from "people tapped like and kept scrolling."

This is why some posts with decent visible engagement still feel like they die early.

They may produce surface interaction without producing downstream intent.

And the opposite is also true:

A post can outperform its visible vanity numbers if it consistently causes deeper actions the system cares about.

The strategic mistake most creators make

Most creators optimize for reaction. The smarter move is to optimize for progression.

Ask this about every post:

What is the next meaningful action this post makes likely?

Examples:

• A sharp contrarian insight may trigger replies and quotes.
• A useful framework may trigger saves, copy-link shares, and profile visits.
• A strong teardown may drive long dwell and follow conversion.
• A compelling thread opener may earn the click into the rest of the thread.

If your post does not naturally lead anywhere, it is easier for it to become disposable.

A better content model: build for intent ladders

Instead of treating every post as a one-off piece of content, think in terms of intent ladders.

Each post should move people one rung deeper.

Ladder 1: stranger → reader

Use a hook strong enough to stop the scroll. Not fake intrigue — real signal compression.

Ladder 2: reader → engager

Give them something worth replying to, quoting, or sharing. That usually means:

• a sharp claim
• an unexpected distinction
• a credible pattern
• a useful framework
• a clear stake in the ground

Ladder 3: engager → profile visitor

Make the post imply depth beyond the post itself. The best posts feel like evidence of a mind worth checking.

Ladder 4: visitor → follower

Your account has to cash the check the post wrote. If the profile and recent posts do not reinforce the same quality, conversion leaks.

Ladder 5: follower → recurring audience member

This is where consistency matters. Not posting constantly — posting recognizably.

The repo’s in-network support is a reminder that repeated affinity compounds.

Practical implications for brands and operators

1. Stop using likes as your primary success metric

Likes are visible, easy to track, and emotionally seductive. They are also incomplete.

A stronger internal scorecard would look at:

• replies per impression
• reposts and quote behavior
• profile visits
• follows per post
• outbound clicks when relevant
• qualitative share-worthiness
• whether the post increases future post performance from the same audience

2. Write posts that can earn more than one type of action

The best posts often have layered utility.

For example:

• A crisp opinion earns replies and quotes.
• A tactical framework earns saves and shares.
• A curiosity gap with substance behind it earns clicks and dwell.
• A strong point of view tied to repeated expertise earns profile visits and follows.

One-dimensional content is easier to replace. Multi-action content is harder to ignore.

3. Reduce low-trust patterns

Because negative actions are explicitly modeled, annoying the viewer is not free.

That should make you more suspicious of:

• repetitive posting in short windows
• generic engagement bait
• manipulative outrage without substance
• forced controversy
• hollow hooks that do not pay off

Even when those tactics create visible activity, they may come bundled with negative feedback risk.

4. Build audience quality, not just reach

Out-of-network retrieval helps new people see you. But in-network preference suggests your existing audience is still strategically valuable.

The practical takeaway:

Do not just ask, "How do I go viral?" Ask, "What kind of audience am I training the system to associate with me?"

High-quality followers who repeatedly engage are not just social proof. They are part of the engine.

5. Post fewer, stronger shots when quality is uneven

The diversity logic is a quiet argument against flooding.

If your fifth post today is weaker than your first, volume may not save you. You may simply spread attention across more mediocre inventory.

Better to post when you have something with enough depth to generate multiple forms of intent.

A simple operating rule

Before you publish on X, run this test:

This post is strong if it is likely to create at least two of these:

• a reply
• a repost or quote
• a click
• a profile visit
• meaningful dwell
• a new follow
• a private share

If it is only likely to earn passive likes, it may still perform — but it is probably a weaker bet than it looks.

Bottom line

The most useful way to read the repo is this:

X appears to reward posts that create meaningful user intent, not just lightweight interaction.

That means the creator edge is not farming cheap engagement. It is producing content that moves people deeper:

from scroll → attention → action → curiosity → relationship.

That is a better growth model anyway.

Because when you optimize for intent, you are not just chasing distribution. You are building an audience with momentum.

---

This analysis is grounded in the local public X recommendation repository, especially the weighted scorer, author diversity scorer, out-of-network scorer, and top-level pipeline documentation. Where exact live production behavior is unknown, I’m treating the repo as directional evidence rather than claiming hidden weights or thresholds.