X Algorithm Read Time: What the October 20 Change Means

In this article

1. What Musk announced and what each model does
2. What is confirmed and what is still unknown
3. How to check the ranking code yourself
4. Why "one reply equals X likes" rules mislead
5. Writing rules for posts people read to the end
6. Lead with the specific claim
7. Cut filler, and accept the length trade-off
8. Make each thread post earn the next
9. A plan you can start this week

On October 19, 2026 (October 20 in UTC), Elon Musk said X's For You feed now uses two new models. A ranking model predicts how long you will spend on a post, and a retrieval model surfaces posts people read all the way through. For B2B writers, that raises the stakes on whether a reader stays with a post to its last line, a signal already in the public ranking code, though how much weight it carries, for which users and since when are all unconfirmed.

What Musk announced and what each model does

In his post, Musk listed two changes to timeline recommendations. The first was a "New AI/SI ranking model that predicts probable time spent much better." The second was a "Retrieval model trained to surface posts people read all the way through." He added that user active seconds were "hitting all-time highs almost every day," and Tesla North reports that he paired the post with a chart of daily active seconds.

The two models sit at different stages of the feed. TokenPost's explainer describes retrieval as the step that gathers candidate posts from accounts you follow and accounts you don't. Ranking then orders those candidates, and selection and visibility filters shape what finally appears.

That split matters if you want to know how to get more reach on X beyond your current followers. Ranking decides where your post lands among a viewer's candidates. Retrieval decides whether your post becomes a candidate for someone who has never seen your name. If retrieval now favors posts people finish, a post readers abandon halfway is less likely to be pulled into strangers' feeds at all.

None of this starts from zero. TokenPost notes that predicted viewing time was already part of ranking before the announcement. Crypto Briefing counts major ranking model updates in January, May and August 2026, and frames the October changes as a push on dwell-time prediction and retrieval of posts with high completion rates. Treat the X algorithm read time signal as a sharper version of something that was already there.

What is confirmed and what is still unknown

Most coverage blurs several separate questions together. Sorting them takes a minute and keeps you from rebuilding your whole approach on the strength of one post.

The October 8 README entry says the Phoenix ranking model "now does a better job of estimating how long you'll spend on each post." That is about 11 days before Musk's post, so the ranking half may have been running before it was announced. Reviewers have found no public code for the retrieval half.

Be careful with rollout language too. Tesla North describes X as "rolling out" both models, which goes further than Musk's own wording, while also noting he did not say when the change went live.

How to check the ranking code yourself

X's Engineering account announced the open-sourced algorithm and said it runs on the same transformer architecture as xAI's Grok model. According to OpenTweet's dated changelog, the code went up on GitHub on January 20, had its largest update on August 13, and is now updated most weekdays. When someone posts a confident claim about Twitter dwell time or For You feed ranking, you can usually check it faster than you can read their thread.

Open the repository and work through these in order:

• Notable Updates in the README. Read the dated entries from newest to oldest. The October 8 note on time spent is there, along with the August notes on weights.
• The predicted actions. The mirror of the README lists what Phoenix predicts for each viewer: engagement actions such as favorite, reply, repost and share; clicks; attention signals (video quality view, dwell, dwell time, click dwell time, active seconds); follow author; and negatives including not interested, mute, block, report and not dwelled. The final score sums each weight times the predicted probability of that action, with negative weights on the negative actions.
• The WeightedScorer and the weights config. The August 13, 2026 release added the configuration parameters that blend predicted values into a score. A Medium analysis from April found that the scorer applies DWELL_WEIGHT to the probability of dwelling and CONT_DWELL_TIME_WEIGHT to the predicted duration as two separate terms, though the author could not read the numeric values at the time.
• The filters. The README lists pre-scoring filters that drop duplicates, posts that are too old, posts from blocked or muted accounts, posts with muted keywords, posts already seen or recently served, and ineligible subscription content. A filtered post never gets scored.
• The commit history. Watch for anything touching retrieval dated after October 19. Until a commit like that appears, the only evidence for the read-through retrieval model is Musk's post.

Why "one reply equals X likes" rules mislead

The August 14 README note addresses a common misreading. Weights scale the viewer's predicted probabilities, or predicted continuous values such as dwell time, rather than raw engagement counts. The note says it would be "incorrect" to conclude that "1 report cancels out 468 likes" just because a report carries 468 times the weight of a like.

The same caution applies to dwell figures in creator blogs. OpenTweet's algorithm secrets post claims dwell time carries a weight of +10, equal to a bookmark and half a retweet, but it was published before the August 13 release put weights in the public repo. Compare any figure like that against the current config before you repeat it or plan around it.

Writing rules for posts people read to the end

These rules follow from what the code predicts. They are inferences, and X has not said how much time spent counts against likes and reposts, so treat them as hypotheses to test on your own account.

Lead with the specific claim

The negative action list includes "not dwelled," which suggests a quick scroll past can count against a post. The first line is where a reader decides to stop or continue, and on X it is often most of what they see.

Put the claim there. "Three lessons on hiring" gives a reader nothing to judge. "Drop take-home tests for senior engineering hires" states a position they can agree or argue with before reading on, and the reason, the evidence and the limit can follow.

Cut filler, and accept the length trade-off

Because the scorer treats the chance of dwelling and the predicted dwell duration as separate terms, it is tempting to pad posts so people stay longer. If retrieval rewards read-through, padding works against you, since every extra sentence is another place to quit. A short post most readers finish may do better than a long one most abandon, and nobody outside X knows where that line sits for your audience.

For each sentence, check whether it adds a fact, a step, a number or a reason. If it only restates, delete it. Warm-up lines and closing recaps usually go first.

Long-form sits on both sides of this trade-off. Premium long posts and Articles can hold readers longer, but they may be finished less often, so test both formats on your own account rather than relying on published multipliers.

Make each thread post earn the next

Adlibrary's guide, a practitioner source rather than X documentation, advises making each post in a thread strong enough to stand alone, since some users only see the first one, and putting the strongest content in the first one or two posts. A thread that saves its point for post six asks readers to sit through five posts of setup.

Write each post to deliver one complete point and end with a specific reason to open the next, such as the number you are about to explain or the step that comes after. Cut teaser posts that promise value without giving any, and drop the final post that only summarizes the others.

A plan you can start this week

1. Read the README's Notable Updates and the weights config, and note the date of the newest entry so you can see what changed when you check again.
2. Pull your last several posts on X. Copy only the first line of each and ask whether a stranger could tell what the post claims from that line alone. Rewrite the ones that fail.
3. Draft this week's posts with the claim in line one, then cut every sentence that adds no fact, step, number or reason.
4. For any thread, read each post in isolation. If one makes no sense alone or gives no reason to continue, rewrite it or merge it into its neighbor.
5. Pick one idea and publish it twice over the coming weeks, once as a short post and once as a long post or Article, keeping topic and posting time similar so format is the main difference.
6. Recheck the commit history in a couple of weeks for retrieval changes dated after October 19, and revise your rules only when the code gives you a reason.
7. HireSteward can take on the drafting in step 3 and the scheduling in step 5. It writes LinkedIn and X posts in your voice from your own notes about real work (it only builds stories from those notes), flags AI-sounding phrasing before drafts reach you, and publishes only after you review and edit. Steps 1, 2, 4 and 6 stay manual, as does deciding what your step 5 comparison means. Plans start at $35/month after a 7-day free trial with no credit card, and you can start a free trial to test it on this week's drafts.

Frequently asked questions

Did X really change its algorithm to favor read time?

Musk announced a time-spent ranking model and a read-through retrieval model on October 19, 2026 (October 20 in UTC). The public README has an October 8 note saying the Phoenix ranking model now estimates time spent better. The read-through retrieval model has not been found in the public code, and the rollout status is not established.

Does dwell time now matter more than likes and reposts?

Nobody outside X knows. X has not disclosed how time spent is weighted against likes, replies and reposts. The public code scores a post by multiplying each weight by a predicted probability and adding the results, so simple conversions like one reply equals so many likes are misleading.

Should I write longer posts on X to increase dwell time?

Not by padding. If retrieval favors posts people finish, extra sentences give readers more places to quit. Cut anything that does not add a fact, step, number or reason, and test short posts against long posts or Articles on your own account.

How should I structure threads after this change?

Make each post strong enough to stand alone, because some readers only see the first one. Put your strongest content in the first one or two posts. End each post with a specific reason to open the next one.

Where can I see the X ranking code?

It is in the xai-org/x-algorithm repository on GitHub. Start with the Notable Updates section of the README, then look at the predicted actions list, the WeightedScorer and the weights config added on August 13, 2026. Watch the commit history for any retrieval change dated after October 19.

Sources

1. Elon Musk on X: "Two important changes to timeline recommendations..."
2. X Just Changed What Shows Up in Your Timeline
3. Musk Outlines New Ranking and Retrieval Models for X Timeline
4. X updates timeline recommendations with new AI ranking model
5. GitHub - xai-org/x-algorithm: Algorithm powering the For You feed on X
6. xAI - Company Profile, Milestones & Funding - Venture Atlas
7. Elon Musk Changes X Algorithm Again To Focus on Read Time Instead of Just Retweets and Likes
8. Engineering on X: "We have open-sourced our new 𝕏 algorithm..."
9. Twitter Algorithm Changes 2026: Dated X Changelog
10. GitHub - Mu-L/x-algorithm (mirror of xai-org/x-algorithm README)
11. How the X Algorithm Actually Decides What You See in "For You"
12. x-algorithm/README.md at main · xai-org/x-algorithm
13. 10 X Algorithm Secrets Most Creators Miss in 2026
14. X (Twitter) Algorithm Explained 2026: How For You Ranks Posts