X pushed new code on November 19, 2026 (commit b551517). The biggest change connects two parts of X that weren't connected before: the AI review that labels posts asking for engagement, and the ranking that scores every post. The connection is switched off for now. No ranking weight changed.

What are the three labels?
Grox, the AI review that reads every post when it's published, has three spam types for posts that push people to act: promotes replies, promotes shares, and promotes clicks. When it finds one, it puts a matching label on the post. That part isn't new. It's been in X's code since November 13 (commit 590f392), and it didn't change on November 19.
The three aren't treated the same today:
- Asking for shares gets the "promotes shares" label and the
SPAM_HIGH_RECALLlabel. The spam label keeps the post out of the feeds of people who don't follow you. It applies to every account, however big. - Asking for replies gets only the "promotes replies" label. There's no spam label.
- Asking for clicks gets only the "promotes clicks" label. There's no spam label.
Two older spam types, engagement baiting and engagement farming, also get the spam label on every account. X didn't publish the instructions Grox follows, so the code doesn't show which wording counts as asking.
What changed on November 19?
The feed can now pass those labels to the ranking service. For the 200 highest-scoring posts in a batch, the feed looks up each post's labels and sends the ranking service a flag for each one: "reply ineligible," "share ineligible," or "click ineligible." Each label sets only its own flag. A post labeled for asking for replies is flagged for replies, not for shares or clicks.

Three limits are in the code:
- It's switched off. The setting
VMRankerSendEngagementIneligibleis false inhome-mixer/params/param.rs, the file X mirrors from its live settings. - The lookup gets 80 milliseconds. If the labels don't come back in time, no flags are sent.
- What the ranking service does with a flag isn't published. The name says the post isn't eligible for that kind of engagement, so the likely use is to stop counting that part of its score. That's a reading of the name, not something the code shows.
If it works that way, a reply label would cost the most. A reply weighs 5, which is 7.0621 times a like. A reply from someone you follow back, on an original post, weighs 20.
What else changed?

- Muted words now stop mention notifications. If you've muted a word for notifications, a post that mentions you and uses that word won't notify you. A mute set to apply only to people you don't follow still lets people you follow through.
- An edit bug is fixed. X's filter could treat the newest version of an edited post as an old one. Now a version only counts as old when a later edit exists.
- Automatic enforcement can reach bigger accounts. X's automatic enforcement skips accounts above a follower limit. Each rule can now raise that limit, up to 1,000,000,000 followers. The real limit isn't in the published code.
- Ads can be spaced evenly. A new setting would space the remaining ads evenly when some are dropped, instead of leaving gaps. It's off.
- X saves its model's own predictions. The live model's predictions for the posts people are actually shown are now saved to train other models. X also added a long-term record of the authors each person engages with, for training. Neither changes how posts are ranked today.
None of the weights moved. A like is still 0.708, a reply 5, a share by direct message 5, and a copied link 20.
What should you do now?
- Never ask for shares. It's the one of the three that already gets the spam label, and that keeps a post away from everyone who doesn't follow you.
- Ask a real question instead of asking for replies. X's creator payout rules already say not to solicit engagement, which they describe as repeatedly asking people to like, reply, bookmark, follow, or repost. A question people want to answer doesn't need the instruction.
- Link to things worth opening, without "click the link." A clear reason to open it does the same job.
- Check your templates and pinned posts for lines like "share this" or "drop a comment," and take them out before X turns this on.
Every current weight is on The numbers, now read from November 19. The book's next drop on November 25 covers these changes.
Source: X's published code, commit b551517.
What to do about it is in the book. Updates say what changed. The book explains how the feed works and what to do, and it's revised every month from X's code.
See what it costs