Most "beat the algorithm" advice is guesswork — confident threads built on vibes and survivorship bias. We can do better than guess here, because the part everyone treats as a black box isn't one. X's Heavy Ranker — the model that scores which posts you see — is literally in a public repo, with the engagement weights written out as numbers. You don't have to reverse-engineer it. You can just read it.
Stop chasing likes. X's own model scores a reply ~27× higher than a like, and a reply you reply to ~150× higher. Earn conversations, not vanity metrics.
What the code actually weights
These aren't estimates — they're the engagement weights pulled verbatim from X's open-source Heavy Ranker. Each number is the multiplier the model puts on a predicted action when it ranks your post.
- Reply the author then engages with75.0
- Reply13.5
- Profile click that leads to engagement12.0
- "Good click" (open + ~2 min dwell, or a reply)11.0
- Good click (v2)10.0
- Retweet1.0
- Like / favorite0.5
- Video watched ≥ 50%0.005
- Negative feedback (block / mute / "show less")−74.0
- Report−369.0
Read those numbers and the whole game reframes itself. A reply is worth 27× a like (13.5 ÷ 0.5). A reply the author responds to is worth ~150× a like (75 ÷ 0.5). Likes sit near the bottom of the positive list — barely above a half-watched video. And the punishments dwarf the rewards: a single report can erase the value of hundreds of likes. The model isn't asking "did people tap a heart?" It's asking "did this start a conversation people stayed in — without making anyone hit mute?"
One more structural fact worth knowing: roughly half of a "For You" feed is in-network (people you already follow). X's own repo notes that "~50% of posts come from this candidate source." So your existing followers and the conversations you have with them genuinely seed how far a post travels.
These weights are from X's April 2023 open-source release. X tunes them over time, so treat them as direction, not gospel — the relative order (replies ≫ likes, reports nuke you) is the durable lesson, not the exact decimals.
What people who grew actually do
The weights tell you what to aim for. These creators show what aiming at it looks like as a habit — each card links straight to the source.
X open-sourced the recommendation algorithm. The engagement weights above are pulled directly from the repo's README. You're not interpreting a leak or a guru's screenshot — you're reading the same scoring logic that ranks the feed.
Consistency beats virality. Koe writes in ~2-hour focused blocks, 4 days a week — steady output compounds more than sporadic viral hits. The engine rewards whoever keeps showing up and keeps starting conversations, not whoever lands one lucky banger.
Compounding through daily shipping. Butcher posts every day — the reach builds on itself and snowballs. Each post is another shot at the high-weight actions (replies, profile clicks), and the audience that accumulates makes every later post start from a bigger base.
Outbound links bleed reach — so move them to a reply. Musk: "Posting a link with almost no description will get weak distribution." X's Head of Product Nikita Bier explained the mechanism — "posts with links tend to get lower reach… the web browser covers the post and people forget to Like or Reply." The fix both endorse: keep the link out of the main post and let the post earn engagement on its own.
“Your 1,001st image isn't like a +1, it's like a +150.”— Jack Butcher, on why daily output compounds · via growthinreverse.com
Steal these 8 rules
The weights collapse into a short list of do-this moves. None are hacks — they're just posting in the direction the engine already rewards:
- Keep links out of the main post — drop the URL in your first reply. Link posts get covered by the in-app browser and lose the likes/replies that drive reach.
- Bait replies, not likes. End with a real question or a take people want to argue with — a reply outscores a like 27×.
- Reply to every comment in the first hour. A reply you answer is the heaviest signal there is (~150× a like). Your own comment section is the main event.
- Front-load a skimmable first line. The model rewards "good clicks" — an open plus ~2 minutes of dwell — so earn the expand.
- Upload native media. Images and video posted directly keep people on X; linked media sends them away.
- Post when your audience is awake. Early velocity compounds, so don't post into a dead window.
- Don't earn mutes, blocks, or reports. One report (−369) can erase hundreds of likes — skip the rage-bait and follow-for-follow.
- Be consistent. The engine rewards whoever keeps showing up and starting conversations, not one lucky banger.
Rules grounded in X's open-source ranker plus statements from @elonmusk and X's @nikitabier. X re-tunes the weights; the direction is the durable part.
What this looks like in real numbers
The weights aren't abstract. Plug two real posts into the math and the "smaller" one usually wins — because the feed scores conversations, not applause.
≈ 210 ranking points
≈ 576 ranking points
Common mistakes that tank your reach
- Optimizing for likes (0.5) over replies (13.5). You're farming the lowest-weighted signal there is.
- Chasing one viral hit instead of the daily reps that actually compound.
- Posting bait that earns mutes or blocks (−74). One report can wipe out the value of hundreds of likes.
- Low-dwell content — links people bounce off, videos watched under 50% — scores almost nothing.
- Forgetting ~half your reach is out-of-network. Write for people who don't follow you yet.
Audit your last post in 3 minutes
- Open your last post's analytics — note replies, likes, retweets, profile clicks.
- Run them through the Algorithm Score below to see the weighted total.
- If likes ≫ replies, your post was likeable but not reply-able — it didn't start anything.
- Rewrite the next one to provoke a reply: ask a real question, take a side, leave a gap.
- Then reply to every early reply — each author-reply is a 75-weight event.
Lesson 3 in five lines
- The weights are public — X open-sourced the ranker, so stop guessing.
- Replies ≫ likes: one reply is worth about 27 likes (13.5 vs 0.5).
- Author-replies ≫≫: a reply you answer is worth ~150 likes (75 vs 0.5).
- Negative feedback and reports are the real killers (−74 and −369).
- Consistency compounds — Koe's 2 hrs/day, Butcher's daily ship.