What Actually Moves the X Algorithm in 2026

The signal hierarchy has flipped

If you're still optimizing for likes, you're optimizing for a metric the ranker barely uses. Across the accounts we've audited in the last six months, the strongest correlations with home-timeline impressions are, in order: completed dwell time, reply velocity in the first 30 minutes, profile clicks, and only then likes and reposts. The shift started in late 2024 and accelerated through 2025 as X tuned the model toward time-on-platform rather than surface engagement.

The practical consequence: a post with 40 thoughtful replies and 80 likes will out-reach a post with 600 likes and 4 replies, every time.

What "dwell" really means

Dwell is not "scrolled past slowly." X measures the milliseconds between a post entering the viewport and the user either tapping into it, replying, or scrolling past. Posts that get expanded (tap to read more, tap to open thread, tap to view image) reset the clock and accumulate weight.

Three patterns reliably extend dwell:

• Cliffhangers in line 1. The first line is what the ranker uses to predict click probability. "Here's the framework we used to grow to 100k" outperforms "I want to share something I learned" by an order of magnitude.
• Visual asymmetry. Posts with one tall image, a chart, or an unexpected screenshot capture more dwell than text-only posts of the same length.
• Native formatting. Bold Unicode, line breaks, and bullet glyphs cause readers to scan rather than scroll. Scanning counts as dwell.

Reply velocity is the new viral lever

The 30-minute window after publishing decides whether a post enters expanded distribution. The ranker treats early replies as a high-confidence quality signal because replies cost more attention than likes. Two tactics consistently move this number:

1. End with a real question. Not "thoughts?" — a question with a specific scope ("which of these have you tried?"). Specificity invites a reply because the reader knows what to say.
2. Reply to your own first reply within 5 minutes. This kicks the conversation thread into the ranker's "active discussion" bucket and roughly doubles subsequent impressions for the next hour.

What stopped working

• Engagement bait swaps. "Like for part 2" / "reply 'send' for the doc" — the classifier flags these and silently caps reach.
• Reply guy farming. Replying to mega-accounts no longer transfers reach the way it did in 2023. The ranker discounts impressions earned in someone else's replies when scoring your home-timeline placement.
• Pure thread storms. Threads longer than 7 posts now see diminishing returns; the algorithm prefers a tight 4–6 with one strong reply continuation.

A playbook that ships today

Pick one of these per post and you'll outperform 80% of the feed:

• The framework post. Name a 3-step model, give one example per step, end with the question "what would you add?"
• The contrarian breakdown. State the consensus, name why it's wrong, give the evidence, propose the replacement.
• The receipt post. Screenshot + one paragraph of analysis. Receipts dwell because readers want to verify them.

Key takeaways

• Optimize for dwell and replies, not likes.
• The first 30 minutes decide the next 30 days of reach for that post.
• Engagement bait is a tax, not a lever — the classifier sees it.
• Native formatting (bold, line breaks, glyphs) is a free dwell multiplier.
• One strong post a day beats five mediocre ones.

If you're building a system to ship posts that hit these signals, Threadly was built for exactly this — Unicode formatting, thread reordering, and an AI assistant that rewrites for dwell, not for likes.