X's new feed ranking model (October 2026): what changed and what brands should do

X changed how its For You feed picks and orders posts in October 2026. Unlike most algorithm news, this change is documented in public code and dated notes, so you can separate what X said from what people guess.

In October 2026, X's Phoenix ranking model started predicting how long you'll spend on each post more accurately (logged October 8), and its retrieval model was trained to surface posts people read all the way through (logged October 13). Elon Musk announced both on X on October 19-20. For brands, that favors posts people finish over posts built for a quick like.

Below: the exact changes with sources, how the feed scores posts, what it means for brand accounts, and which popular claims are speculation.

What did X change in October 2026?

X changed two parts of the For You feed: ranking and retrieval. Both are logged in the "Notable Updates" section of X's public x-algorithm repository on GitHub (read November 12, 2026):

Musk's post went up on the evening of Monday, October 19, US time, which is October 20 in many other time zones. Tesla North quoted it in full and noted that X has not said exactly when the change reached all accounts or how time spent is weighed against replies and reposts.

Retrieval is the step that gathers candidate posts from across X. Ranking decides the order you see them in. The October update touched both, which means read-through now affects whether your post is considered at all, not only where it lands.

How does X's For You feed score posts?

X's repository describes the scoring openly. The Phoenix model reads a viewer's recent engagement history and predicts, for each candidate post, how likely that viewer is to take each action. The predictions are combined into one score:

Final Score = sum of (weight x predicted probability of each action)

The actions X lists, as of November 2026:

Three adjustments follow, per the repository:

• Author diversity. Each post after an author's first in a feed is multiplied by a decaying factor.
• Out-of-network discount. Posts from accounts the viewer doesn't follow are scaled down, as are replies and reposts from accounts they do follow.
• New-author lift. Posts from authors with impressions below a threshold are lifted toward a target position.

Posts older than 48 hours are filtered out of For You before scoring. X also warns against a common misreading: weights scale the predicted probability for that viewer, "they do not scale the raw engagement counts." A report weight hundreds of times larger than a like weight does not mean one report cancels hundreds of likes.

What role does Grok play in X's ranking?

X's engineering team said on January 20, 2026 that its open-sourced feed algorithm is "powered by the same transformer architecture as xAI's Grok model" (Decrypt). The current repository describes Phoenix as a transformer and does not name Grok. Musk's October post calls the new ranking model "AI/SI."

So the documented position, as of November 2026: the ranking model uses a Grok-derived transformer architecture. Whether the October models are Grok models, use Grok outputs, or "read" post content with Grok is not documented. Claims that "Grok reads every post and decides your reach" go beyond what X has published for the October change.

What does this mean for brands on X?

Posts that hold attention and get read to the end should gain, and posts that get scrolled past should lose. That follows directly from X's two notes. Five implications:

1. Read-through is now a retrieval signal. A post people abandon halfway may not enter the candidate pool for non-followers. Front-load the point and keep posts as long as the content needs, no longer.
2. Time spent counts, but so do negatives. "Not dwelled" is listed as a negative action. A hook that makes people stop and then leave quickly can hurt.
3. Replies and quotes still matter. They are predicted actions with their own weights. The October notes changed time-spent prediction, not the list of actions.
4. Volume has diminishing returns. Author diversity decay means your third post in a viewer's session is worth less than your first. Five weak posts a day can underperform two strong ones.
5. Fresh posts only. The 48-hour age filter means evergreen content needs a fresh post with a new angle every so often.

How should brands post on X now?

Post things worth finishing, written for the people who follow you, and measure read-through behavior where you can. These practices tie back to what X has said, not to rumors.

• Lead with the point. First line states the claim or the news. Detail follows. Readers who get value early keep reading.
• Use threads only when each post earns the next. A thread people quit after post two signals the opposite of read-through.
• Make video watchable without sound and worth finishing. Video continuation seconds and video quality views are listed attention signals.
• Write for your followers first. The out-of-network discount means your existing audience carries early distribution. Posts they engage with deeply are what travels further.
• Reply to replies. Replies are a predicted action, and conversations keep people on your post.
• Don't flood the feed. Space posts out and give each one a distinct idea.
• Check Under the Hood before blaming the algorithm. X's August 2026 transparency tool shows aggregate statistics about labels on your account and posts that can limit visibility. Since September 18, 2026, it also shows if posts were withheld in a country because of a legal demand.

If you publish to X alongside other networks, keep X's version separate. X's 280-character limit for standard accounts and its read-through signal reward a tighter post than the LinkedIn version of the same idea. Our guide to cross-posting to 13 social networks covers per-network versions.

What is speculation about the X algorithm?

A lot of what circulates after each change is not in X's statements. As of November 12, 2026, treat these as unverified:

The repository's code and parameters are public, so specific weights can be checked. They also change: X says production defaults are synced into the code by scripts and that experiments run on a share of traffic. A number someone read in October may not be the number running in November.

How do you measure whether the change helped or hurt you?

Compare like with like across the October 8 to October 20 window. Pull your X analytics for the four weeks before and after, and look at impressions per post, profile visits and link clicks, split by format (text, image, video, thread). If impressions fell on threads but rose on single posts, that fits the read-through change. If everything fell evenly, check Under the Hood for labels first.

Then follow the outcome that matters: did X bring visitors who became leads or customers? Tag links with UTM parameters so you can tell. Our guide to UTM parameters for social media posts gives a naming scheme, and the attribution guide shows how to match CRM contacts to the post that brought them in.

Next step

Take your last ten X posts, mark which ones people likely read to the end, and write next week's posts in that shape: point first, one idea per post, video worth finishing. Track the results with tagged links rather than likes alone. Tracee schedules X alongside 12 other networks with a version per channel and shows which posts brought in customers; see the pricing page for plans at the time of writing (November 2026).