The Twitter Algorithm Update Everyone Got Wrong (October 2026)

Key takeaways

• Twitter changed three things in its published ranking code between 22 and 23 October 2026. No ranking weight changed.

• Equal trimming: when someone's feed has more candidate posts than the 2,800 Twitter's ranking model can score, every source is now cut by the same share. Before, the last source in line lost the most, and that is where posts under 2 hours old sat.

• A warning label: posts labelled low quality or misinformation now carry a notice in the published code. It is shown to everyone, not just non-followers, and on low-quality posts it blocks likes, replies, reposts, bookmarks and shares.

• The AI slop label: accounts Twitter flags as "LLM slop" get a spam label that hides them from people who don't follow them. It now lasts 7 days instead of 30.

• The viral threads got the big picture mostly right, but several details were wrong or overstated. The fix for small accounts is the same as before: post originals when you can be around for the first 2 hours, write things people read to the end, and don't post like everyone else.

Twitter's October 2026 algorithm update changed how the For You feed trims posts before ranking them, added a warning label to low-quality and misinformation posts, and shortened the "AI slop" account label from 30 days to 7. It did not change a single ranking weight. You can see all of it in the open-source code Twitter's parent company publishes at github.com/xai-org/x-algorithm, which updates almost every day.

The update went viral through two threads. Peter Mick's post on 25 October passed 2.7 million views, and Shann Holmberg's breakdown followed a day later. Both read the code and both got a lot right. But a few of the most-repeated lines don't match what the code says, and the biggest change for anyone using AI to post didn't get mentioned at all.

We compared the code from 21 October with the code from 29 October, line by line, and checked every claim in both threads. Here's what actually changed, what didn't, and what to do about it. (Disclosure: we make VoiceMoat, an AI social media tool for LinkedIn and Twitter. It comes up once, in the AI slop section.)

The short answer: three changes, no new scores

If you only read one line: nothing about how your posts are scored changed, but a little more of your newest content now gets a chance to be scored at all.

The six steps behind your feed, and the GitHub folder for each.

Change 1: the 2,800 cap and equal trimming

Every time you open the For You feed, Twitter gathers candidate posts from several places, then asks its ranking model to score them. The model can only score 2,800 candidates per request. That cap is in the code, and it has been there for a while.

Where the candidates come from, at the default limits:

• Accounts you follow: up to 1,200

• Popular posts: up to 500

• Posts liked by people similar to you: up to 800

• Twitter's own recommendation model: up to 1,000

• A "cold" pool of very new posts from that model: up to 200

That adds up to about 3,700, which is where the threads' number comes from. It isn't a single number in the code; it's the sum of these limits. Filters remove duplicates, posts you've already seen and old posts before the cap kicks in, so most real feeds start with fewer.

Before and after equal trimming, with every source at its default cap.

How it worked until 21 October. The sources were lined up in a fixed order: accounts you follow first, then popular posts, then similar-taste posts, then the recommendation model last. When there were more than 2,800, the code simply kept the first 2,800 in line. Everything after that was never scored, and a post that's never scored can't make it to the top of anyone's feed.

At full limits, that meant only about 300 of the recommendation model's 1,200 candidates got scored. Its cold pool is posts up to 2 hours old, so very new posts from people outside your network were the ones most likely to be dropped.

How it works from 22 October. A setting called EnableFairShareRanking was switched on, using a fair-share filter. Now, when there are too many candidates:

• Every source keeps the same share of the 2,800.

• Each source drops its own lowest-ranked posts, not whatever happens to be at the end of the line.

• If a feed has 2,800 candidates or fewer, nothing is cut.

At full limits, everyone keeps about 76%. The recommendation model now gets roughly 757 of its posts scored, plus about 151 from the cold pool, instead of about 300 in total.

The 2,800 cap and the switch that turned on equal trimming. Source: github.com/xai-org/x-algorithm

What nobody knows yet: how often a real feed goes over 2,800. That isn't in the code. If most feeds stay under the cap, this change does very little. If many go over, newer posts from outside people's networks get a noticeably fairer shot.

Where the threads went slightly wrong: "new posts used to be cut first" is close, but not quite it. The old cut was by position in line, not by age. Very new posts suffered because they sat inside the source that came last. The fix is real, but it is about sources, not timestamps.

Change 2: the warning label on low-quality and misinformation posts

The 23 October release added a rule that puts a notice on posts with one of these labels:

• Low quality

• Misinformation (civic, generic, medical or crisis)

What the notice does, according to the code:

• Who sees it: everyone. The rule's audience is literally Everyone. The tests expect it for people who don't follow you and for the post's own author.

• What it blocks on low-quality posts: replies, reposts, likes, reactions, sending by DM, bookmarks, pinning, viewing post activity, sharing and copying the link. Quoting is not on the list. This list comes from the code's tests, because the live settings file isn't published.

• What it blocks on misinformation posts: only what the specific label says. Some limit replies; some just show the notice.

• What it doesn't do: it doesn't remove the post from the feed or lower its ranking. It runs when posts are shown, not when they are ranked.

What the notice does, and what it doesn't.

The warning rule. Note the audience: Everyone. Source: github.com/xai-org/x-algorithm

Why this probably isn't new. The rule has a start date written into it: 1 April 2019. That matches Twitter's long-running "public interest" notices. The most likely story is that an old system was published in the open code for the first time, not that a new penalty switched on this week.

What the code doesn't tell you: what earns the label. That decision comes from Twitter's safety systems, and those aren't in the repo.

Where the threads went wrong: "non-followers can't like, reply or share" is wrong on one point. The published rule applies to everyone, followers included. And "misinformation labels limit your reach" isn't supported by the For You code: labelled posts are kept and ranked as normal; they just show the notice.

Change 3: the AI slop label (the one nobody covered)

This is the change that matters most if you use AI to post, and neither thread mentioned it.

Twitter's enforcement rules include a check for accounts flagged as llm_slop_user. In plain English: accounts its systems think are posting generic, mass-produced AI content. When an account gets that flag:

• It gets a spam label. That spam label is on the list of labels that remove your posts from the For You feed of people who don't follow you.

• As of the 23 October release, the label lasts 7 days. It used to last 30.

• Large accounts are skipped. The real follower threshold isn't published (the repo uses a placeholder "to reduce gaming").

• Individual posts can be flagged too, with a separate label that lasts 30 days.

What happens after an account is flagged as AI slop.

The AI slop rule: 604,800,000 milliseconds is 7 days. Source: github.com/xai-org/x-algorithm

The shorter label is good news if you get caught by mistake: a week instead of a month. But it also tells you this check exists, runs on accounts like yours, and costs you all your non-follower reach while it lasts.

What we don't know: how the AI slop check decides. The code shows what happens after an account is flagged, not what the detector looks for.

What it means in practice. The cheapest way to stay clear is not to post like a template. Posts that read like every other AI post (the same hooks, the same lists, the same "here's the thing") are exactly what a slop detector is built to catch. Use AI to help you get more done, but make sure what goes out reads like you.

What didn't change: the weights, the boost, the 48-hour limit

No ranking weight changed between 21 and 29 October (see the weights in param.rs). The last changes were on 13 and 14 October, when the weight on a like went from 0.5 to 0.66 and the weight on time spent went up by half, from 0.004 to 0.006.

Here's how the current weights compare with a like. These multiply Twitter's predicted chance of each reaction, so they're not points per action, but the ratios show what the feed values:

How much each predicted reaction counts compared with a like. Weights multiply predicted chances, not counts.

Shann's ratios check out to the decimal. One thing worth knowing: reports carry a huge weight because they're rare. A report is predicted more than 1,000 times less often than a like, so the weight has to be large to matter at all. For more on how these signals work, see our guides to the 25 signals that rank your posts and negative signals like mutes, blocks and reports.

The small-account boost is unchanged:

• Who: original posts from accounts with 50,000 followers or fewer.

• When: in the post's first 2 hours, while it has fewer than 200 views in the Home feed.

• How much: one eligible post per feed load can be lifted to about 16th place.

• How it's picked: a random draw based on each post's like rate shortlists two posts, and the one the ranking model scores higher wins. "The best like rate wins" is close, but not exact.

The reach rules are unchanged too:

• The 48-hour limit: after 48 hours, a post stops appearing in For You (why the 48-hour rule exists).

• Who sees what: replies and reposts only reach your followers; original posts can reach people who don't follow you, scored 25% lower for them by default (how out-of-network reach works).

• Posting several times: when your posts compete in one person's feed, the second keeps 62.5% of its score and the third 43.75%, by default (what that means for how often to post).

Viral claims vs the code

Most of what the threads said holds up. These are the lines that don't, or that need a caveat:

What small accounts should do this week

1. Post originals when you can stay around for 2 hours. The boost only works while a post is under 2 hours old and has fewer than 200 Home views. A post at 3 a.m. burns that window while your audience sleeps.

2. Lead with one strong post, not a burst. Your second post in someone's feed keeps 62.5% of its score. Spread posts out instead of stacking them.

3. Write for the read, not the click. Twitter's 8 and 13 October model updates reward time spent and posts people read all the way through. Put the payoff early and cut the filler.

4. Write things people pass on. A copied link counts about 30 times a like, and a reply, quote or DM share about 7.6 times. Useful, opinionated posts get shared; vague ones get scrolled past.

5. Don't post like a template. The AI slop label hides you from non-followers for a week. If you use AI, edit until it reads like you.

6. Check your pinned post. Pinning a post with a link Twitter rates as bad can put a 7-day spam label on your whole account.

7. Avoid what people hide. "Not interested" costs about 79 likes' worth and a mute about 89. Rage bait and engagement bait cost more than they earn, and engagement bait gets labelled no matter how big you are.

8. Measure it fairly. Compare reach from people who don't follow you in the weeks before and after 22 October. Give it two to three weeks before you judge.

Check it yourself: where everything lives on GitHub

You don't have to take our word (or anyone's thread) for any of this. The whole ranking pipeline is public, and every link below is pinned to the 29 October version (commit 606b254), so the line numbers won't move.

Twitter's open-source ranking code on GitHub, 29 October 2026. Source: github.com/xai-org/x-algorithm

To follow changes yourself, open the repo's commit history. Every commit has the same message, so compare two days with GitHub's compare view (for example 21 to 29 October) and look at home-mixer/params/param.rs first.

What we still don't know

• How often feeds hit the 2,800 cap, so how big the trimming change really is.

• What earns a low-quality, misinformation or AI slop label.

• Whether every value in the published code matches what runs live. The main settings file mirrors production defaults (last synced 28 October); some ranking settings are code defaults that production can override.

• Several new features arrived switched off this week, including a step that makes users who like everything count for less, and two whole-page re-rankers that add some randomness. None of them affect your feed yet.