Why Are My Replies Hidden on X? What X's Code Shows (October 2026)
X's public code, read October 16, 2026: an AI model scores replies from 0 to 3, a 0 labels the reply, a score only goes down, and replies skip non-followers.
X's open-source code, read on October 16, 2026, shows four systems that can keep your replies out of sight. An AI model scores replies from 0 to 3, and a 0 puts a spam label on the reply. Two other filters look for chains of self-replies and for coordinated spam. Separately, the For You feed removes replies from accounts a viewer does not follow.
You still see your own replies in all of these cases. That is why hidden replies are so hard to notice.
What does X do with a reply after you post it?
New posts on X go through a set of reply checks in X's Grox system. The code reads them from a stream of posts and sends each one to four plans: reply ranking, spam comment detection, coordinated spam and multi-step reply spam. Each plan first decides whether the post qualifies, then calls an AI model, then writes the result.
The writes run only in production. The code disables them in every other environment (DisableTaskForNonProd).
The code below was read at commit 92f0918, published October 15, 2026. The repository is republished almost every day, so a value can change from one commit to the next.
How does X's AI score a reply?
Free, no login, results in seconds.
X's code gives each qualifying reply a score from 0 to 3. A score outside that range is rejected as invalid. Which model reads the reply depends on the size of the conversation:
| Conversation | Model |
|---|---|
| Post you reply to and original post both from accounts with 250,000 followers or fewer | A Gemma model |
| Either of them from an account with more than 250,000 followers | A Grok model (reply ranking) |
A reply that quotes another post goes to the Grok scorer in both cases.
The model sees the conversation, follower counts and signals about the reply. The code logs signals such as whether the text was pasted, the app used, a risky label on the author's account and the number of legitimate blocks the author received in the past 24 hours. How the model weighs them is not public: the prompt is excluded from the repository, and X's README says Grox prompts are left out to reduce gaming.
Both scorers skip a reply when its author also wrote the post it answers or the original post of the conversation.
What happens when a reply scores 0?
A score of 0 puts the safety label RiskyHighVizReply on that reply. The label goes on the post, not on the account. The code skips the label for authors with a high credibility score (66 or more when a score exists), for accounts with a grey badge and for test accounts.
The score itself is stored with the model's reason. The code that orders replies in a conversation is not in the repository, so it does not show how much a low score moves a reply down. X's Help Center does list "downranking the post in replies" among the ways it restricts reach (Authenticity policy, April 2025).
Can a reply's score go back up?
No, not in the published code. Before storing a new score, X's write step reads the one already stored. If the new score is equal or higher, the write is skipped. A score can therefore only go down.
Editing a reply or posting the same text again does not repair an old score in this code. Your next replies are scored on their own.
Which other filters can hide a reply?
Two more plans look at whole conversations rather than one reply:
- The multi-step reply spam filter, added on September 4, 2026, targets chains where you answer your own reply, at least 2 levels below the original post, under someone else's post from an account with at least 1,000 followers. It needs at least 2 flagged posts by the same author.
- The coordinated spam filter checks replies at least 2 levels deep under posts from accounts with at least 180,000 followers. It skips replies by the original author and authors with a high credibility score or a grey badge.
Each post either filter flags gets the same RiskyHighVizReply label and a score of 0.
X's post enforcement rules add two more labels, both kept for 30 days. A post flagged as fast reply spam gets SpamHighRecall. A post flagged as low-effort AI text gets RiskyHighVizReply. X's Under the Hood tool describes a post with SpamHighRecall as "hidden from recommendations to non-followers".
Why don't my replies show up in other people's For You feed?
Replies from accounts a viewer does not follow are removed from that viewer's For You feed before scoring. X's OONRetweetReplyFilter drops them unless a setting allows replies found by X's retrieval model. That setting, EnablePhoenixOonReplies, is off in the published defaults.
This rule applies to every account. People who follow you can still get your replies in For You, and anyone can open the conversation and read them there. Our guide to what X's code says about reach covers this filter and the other For You rules.
Is a hidden reply the same as a ghost ban?
A ghost ban detaches your replies from the conversation they belong to: everyone else sees the thread without them. Reply deboosting means your replies are pushed to the bottom of a conversation or hidden behind Show more replies.
The labels and scores above are causes X's code makes possible. A ghost ban or reply deboosting is what other people actually see. One reply missing under one busy post proves little. The same pattern on your replies across several conversations is a restriction.
What does the code suggest doing?
These points follow from the published code only. The scoring criteria are not public.
- Reply to the post you mean to answer instead of adding replies under your own reply. Chains of self-replies under larger accounts have their own filter.
- Put what you have to say in one reply. X's rules give posts flagged as fast reply spam a label for 30 days.
- Write each reply for its conversation. The model reads the thread, and the code logs whether the text was pasted.
- Leave old replies as they are. A stored score never goes up, so rewriting or reposting does not repair it.
How do you know whether your replies are hidden?
Your own view cannot tell you: X shows you your replies in every case above. You need to see a conversation the way someone who does not follow you sees it.
A free check on Shadowban Radar does that for any public handle, with no login. It tests ghost ban and reply deboosting on your recent replies, plus search suggestion ban and search ban. Our guide to shadowban durations explains how long each restriction usually lasts.
Sources
Read on October 16, 2026, in the xai-org/x-algorithm repository at commit 92f0918, published October 15, 2026, and on X's Help Center.
- generators.py: the four reply plans applied to every new post.
- task_filter.py: which replies each plan checks, including the 180,000-follower minimum for coordinated spam.
- constants.py: the 250,000-follower split between models and the credibility threshold of 66.
- classifier_reply_ranking.py: the 0 to 3 range.
- classifier_simple_reply_scorer.py: the Gemma scorer, and quotes sent to the Grok scorer.
- task_rank_replies.py: the signals logged for each reply.
- task_write.py: the label at 0, its exemptions, the production-only write and the rule that a score only goes down.
- prompts.py: prompts excluded from the repository.
- enforcement_post.yaml: the 30-day labels for fast reply spam and low-effort AI text.
- underTheHoodLabels.strato: how Under the Hood describes
SpamHighRecallon a post. - oon_retweet_reply_filter.rs: the filter for replies from accounts a viewer does not follow.
- X Help Center, Authenticity policy (April 2025): "downranking the post in replies" in the list of reach restrictions.
Shadowban Radar is not affiliated with X Corp. For how X's own label report relates to a live check, see X's Under the Hood.
