X Algorithm Update November 2026: Copy Links Cut
· 11 min read

On this page
X changed its ranking weights twice in November. On 4 November a like rose from 0.66 to 0.708, a share from 2 to 2.118, and the weight on time spent with a post from 0.006 to 0.0075, while clicks, video opens and the "not interested" penalty moved the other way. On 12 November X set the weight for sharing a post by copying its link to 0. It had been 20 since the code went public in August, the biggest positive number in the table.
So the heaviest positive actions on X are now a reply, a quote and a share by DM, at 5 each.
We read all 13 daily snapshots of xai-org/x-algorithm from 27 October to 13 November and dated each change by the sync time X stamps at the top of its weights file. Here is what moved, what it means for your posts, and what X built but hasn't switched on yet. If you missed the last round, our October update covers it.
- 0 Copy-link share weight. Down from 20 on 12 November
- 5 Reply, quote, DM share. Now the heaviest positive weights
- 0.708 Like weight. Up from 0.66 on 4 November
- −53.07 Not interested. Harsher for the third time since September
Source: github.com/xai-org/x-algorithm, home-mixer/params/param.rs, synced 12 November 2026
Key takeaways
- Copying a post's link no longer adds anything to its score. A share by DM still carries a weight of 5, so "worth sending to one person" is still the bar to write for.
- Likes count a little more again: 0.708, up 42% from 0.5 in early October.
- Opening a video adds nothing on its own now. The weight on time spent with a post rose 25%.
- "Not interested" moved to −53.07. Only a mute (−58.8) and a report (−234) cost more.
- X's spam check for tagging now counts people tagged in your photos, not just @handles in the text.
- Under the Hood reports can now list the exact posts behind each label and say how to appeal.
How to read these numbers
Each weight multiplies one viewer's predicted chance of doing something with your post. A like at 0.708 and a reply at 5 means a reply carries about 7.1x a like's weight on the same viewer's odds. It does not mean one reply equals seven likes on your post. X's own README makes this point, and we keep to it: people like posts far more often than they reply, so the two weights are scaled to very different odds.
Every weight that changed
Eight weights moved between the 26 October and 12 November syncs. Two config dates account for all of them.
Ranking weights that changed, 27 October to 13 November 2026. Dated by the config sync stamped at the top of home-mixer/params/param.rs
Signal | Before | Now | Changed |
|---|---|---|---|
Share by copying the link | 20 | 0 | 12 Nov |
Like | 0.66 | 0.71 | 4 Nov |
Share (any kind) | 2 | 2.12 | 4 Nov |
Click | 0.3 | 0.27 | 4 Nov |
Video open | 0.02 | 0 | 4 Nov |
Time spent with the post | 0.01 | 0.01 | 4 Nov |
Time spent after opening it | 0.4 | 0.39 | 4 Nov |
Not interested | -52.3 | -53.1 | 4 Nov |
Source: github.com/xai-org/x-algorithm, snapshots of 5 and 13 November 2026
Everything else held: a reply and a quote at 5, a share by DM at 5, a follow at 4, a repost at 1, a report at −234, a mute at −58.8 and a block at −31.2.
The copy-link share is switched off
X's ranking model, Phoenix, predicts three kinds of share separately: a share of any kind, a share by DM, and a share by copying the link. Since the code went public on 13 August, the copy-link prediction carried a weight of 20, about 30 times a like's weight in October. That made it the single strongest positive signal in the table, and a lot of advice (ours included) told you to write posts people would copy and send.
On 12 November X set it to 0. The model still predicts copy-link shares; they just add nothing to the score. The other two share predictions still count: 2.118 for any share and 5 for a share by DM.
X hasn't said why. One small detail if you go looking: a second file in the repo, vm-ranker/params.rs, still lists 20 as a fallback default. It reads the same live setting, and home-mixer/params/param.rs, which X mirrors from its live config every day, is the one that shows 0.
What it means for you: the advice changes less than the number suggests. A post good enough to copy and send is usually good enough to send by DM, and that still carries a weight of 5. What's gone is any reason to ask people to "copy the link." That was engagement bait before, and now it doesn't even point at a weighted action.
The heaviest positive weights now. Each multiplies one viewer's predicted chance of the action (bar chart)
Weight | |
|---|---|
Reply | 5 |
Quote | 5 |
Share by DM | 5 |
Follow the author | 4 |
Share (any kind) | 2.12 |
Repost | 1 |
Like | 0.71 |
Source: github.com/xai-org/x-algorithm, home-mixer/params/param.rs, synced 12 November 2026
Likes keep climbing
The like weight has now gone up three times since 12 October: from 0.5 to 0.6, to 0.66 a day later, and to 0.708 on 4 November. The October steps were 20% and 10%; this one was about 7%.
A like is still the lightest of the actions people think of as engagement. A reply or a quote carries about 7.1x its weight, a follow about 5.6x. But the gap is narrowing, so posts that people simply enjoy are doing slightly better than they were in September.
The like weight since early October (line chart)
Like weight | |
|---|---|
Before 12 Oct | 0.5 |
12 Oct | 0.6 |
13 Oct | 0.66 |
4 Nov | 0.71 |
Source: github.com/xai-org/x-algorithm, home-mixer/params/param.rs, dated by config sync
A video open no longer counts on its own
Opening a video was weighted 0.07 until 12 October, then 0.02, and since 4 November it is 0. At the same time the weight on time spent with a post went from 0.006 to 0.0075, a 25% rise, and the weight on time spent after opening a post eased from 0.4 to 0.385.
Two notes from X's README point the same way. On 8 October X said the ranking model now does a better job of estimating how long you'll spend on each post. On 13 October it said its retrieval model, the one that finds posts from accounts you don't follow, is now also trained to surface posts that people read all the way through.
And on 4 November X switched on two new predictions in the ranking model: how long a viewer will spend on your profile after seeing a post, and how long they'll keep watching videos afterwards. X calls these "home excursion" scores. Both carry a weight of 0 in the published score for now, but the ranker now reads them for every post it scores.
What it means for you: a thumbnail that makes people tap and leave earns nothing. The first seconds of a video and the first lines of a post are what hold someone, and holding them is what pays.
"Not interested" got harsher again
The "not interested" penalty was −43.2 before 28 September, then −47.52, then −52.272 on 13 October, and now −53.07. This step was small (about 1.5%), but it keeps moving in one direction. Only a mute (−58.8) and a report (−234) cost more.
Why the negatives matter so much: Before ranking, X offsets the score so that a post whose predicted negatives outweigh its positives for a viewer ranks below every post that scores above zero for that viewer (xai-value-model/scoring.rs). Waving a post off is far rarer than liking one, which is why the weight is so large: even a small predicted chance pulls a post down.
The usual causes are the ones you'd guess: repeating your own recent posts, posting in bursts (your second post in the same feed keeps 62.5% of its score, your third 43.8%), rage bait, and anything that reads as an advert.
Tagging: photo tags now count toward the spam check
X runs an extra real-time spam check on posts that tag several accounts. Until this month it counted words in the text starting with "@" and checked any post with two or more (grox/flows/ptos/task_filter.py).
The new code counts distinct handles: @mentions in the text plus people tagged in the post's photos, with the same account counted once however often it appears. The threshold is still two. That check can label a post as possible spam, which keeps it out of For You for people who don't follow you.
So a post that mentions one person and tags another in its photo now gets the extra check. A post that mentions the same person twice no longer does.
X also added a safety pass for posts with more than four images (counting a quoted post's images), which reviews them in groups of four so every image gets looked at (task_safety_ptos_many_media.py).
Under the Hood: the posts behind a label, and how to appeal
Under the Hood is X's monthly report on the visibility labels that touched your account and posts. Code published at the end of October adds two things to it.
The posts themselves. For each post label, the report can now list which posts carried it: every one when three or fewer did, none when more than 5% of your posts that month did, and at most the 50 most recent. Takedowns are always listed.
How to appeal. Each label can now carry X's own appeal steps. Labels for abuse, hateful conduct, violent speech and civic integrity say you can appeal from the post. Posts withheld in the UK, the EU, Australia or Brazil link to that region's appeal form, Japan has its own form for legal requests, and copyright takedowns point to the counter-notice form.
Both features sit behind switches that are off by default in the published code (underTheHoodReport.User.strato), so expect them to reach reports gradually. Our free shadowban check reads the new fields already: drop in the JSON file from Under the Hood and it lists the posts behind each label, with a link to each, and X's appeal steps.
X is also building a "why am I seeing this" for the feed itself. For Under the Hood requests, the ranker can now return the top three actions behind a post's score, each at least 10% of it, plus a trace of every stage the post went through. That is switched off for now.
Built, but switched off
These landed in the code during the same window, all off by default or with no effect on the score yet:
- Two whole-page rankers. Instead of scoring posts one by one, they choose a page of posts together. Both are off.
- A Grok head. The model can predict whether a viewer will ask Grok about a post. Its weight is 0.
- Ten candidate ranking models in shadow tests. Experiments 18, 19 and 22 to 29 are each scored on 30% of a copy of live traffic, up from two in October. Shadow traffic is scored but never shown.
- Topic and diversity inputs for the ranker, sending post topics and a viewer's recent history, both off.
- A like-rate floor for brand-new accounts: out-of-network posts shown to accounts under 30 minutes old would need at least 4 likes per 1,000 views. The threshold is set, but the filter itself is off.
One filter left: on 26 October X removed the Brazil 2026 election filter from For You, now that the election is over.
What hasn't changed
The new-author boost is the same as in October: one original post per feed, from an author with up to 50,000 followers, under two hours old and under 200 home-feed views, can be lifted to the 16th slot. Posts shown to people who don't follow you are still scored at 0.75x, and so are your replies and reposts. A reply from a mutual follow on an original post is still weighted 20 instead of 5.
What to do this week
- Write for the DM, not the link. One line someone would send to a single friend. Never ask for the share.
- Leave a real opening to reply to. Replies and quotes carry the most weight, and replies from people you follow back count four times as much.
- For video, earn the first seconds. An open alone is now worth nothing; time watched is what counts.
- Tag only the people the post is about, photo tags included.
- Don't give people a reason to wave you off. Space your posts out and don't repeat last week's.
- Check your Under the Hood report if your reach dropped, and read it with the shadowban check.
The interactive lab below runs on the 13 November weights. Set what 1,000 viewers do with a post and see how the ranker would score it.
Interactive demo: What the X algorithm rewards
Did X remove sharing by copy link?
No. You can still copy a post's link, and X's model still predicts how likely you are to do it. Since 12 November that prediction has a weight of 0, so it adds nothing to the post's ranking score. Shares by DM (weight 5) and shares of any kind (2.118) still count.
What is the strongest positive signal in the X algorithm now?
A reply, a quote and a share by DM are tied at 5. A follow is 4. A reply from someone you follow back, on an original post, is weighted 20.
Is a like worth more now?
Its weight rose from 0.66 to 0.708 on 4 November, the third rise since early October. The weight multiplies each viewer's predicted chance of liking, not the number of likes, so it is not a fixed exchange rate against replies or shares.
How do you know when each change shipped?
X publishes a snapshot of the code almost every weekday, and the weights file starts with the time it was synced from X's live config. We compare consecutive snapshots and date each change by that sync time.
Do I need to change how I post?
Not much. If you were writing posts worth sending to a friend, keep going: a share by DM still carries a weight of 5. Drop any call to copy a link, make video openings count, and keep tags to the people the post is about.
Sources
- home-mixer/params/param.rs, 13 November snapshot (config sync 12 November) and 5 November snapshot (config sync 4 November)
- README, Notable Updates for 8, 13 and 26 October
- home-mixer/scorers/phoenix_scorer.rs (home excursion scores) and vm-ranker/scoring/value_model.rs(top heads)
- xai-value-model/scoring.rs (how the weighted score is summed and offset)
- grox/flows/ptos/task_filter.py and grox/core/data_loaders/data_types.py (the tagging spam check)
- under-the-hood/strato/columns/underTheHoodReport.User.strato and underTheHoodLabels.strato(post lists and appeals)