On October 20, 2026, Elon Musk announced two changes to X timeline recommendations: a new AI ranking model that predicts probable time spent, and a retrieval model trained to surface posts that people read all the way through. He said user active seconds is "growing exponentially and hitting all-time highs almost every day", and attached a chart. The post had about 1M views within hours.
For anyone who publishes on X, the useful part is not the growth claim. It is the training target. A feed that ranks by predicted time spent, and retrieves by finished reads, rewards a different kind of post than a feed that ranks by replies and reposts. This post covers what was said, what the chart shows, and what to test.
A short fuse with a green spark showing how X ranking now focuses on attention and time spent
TL;DR: the questions people ask first
| Question | Short answer |
|---|---|
| What changed? | A new ranking model (predicts time spent) and a new retrieval model (finds posts read to the end). |
| Who said it? | Elon Musk, in a post on October 20, 2026. No engineering write-up or code link came with it. |
| Is the growth claim true? | UAS is up, but the chart shows about 10 to 12 percent since late summer. It does not show exponential growth. |
| Do I need to change how I post? | Test it. Posts that people finish should do better. Padding and bait should do worse. |
| Is there a link to the model? | No. It is not in the public repository as of this post. |
| Does it help users? | Unclear. Time spent is not the same as value. |
What exactly did X announce?
Two models changed, and they sit at different stages of a recommendation pipeline.
Retrieval is the first stage. It pulls a few thousand candidate posts out of a very large pool. Musk says the new retrieval model is trained to surface posts that people read all the way through. In other words, the label it learns from is "did the reader reach the end", not "did the reader click like".
Ranking is the second stage. It orders the candidates for one user. Musk says the new ranking model is an "AI/SI" model that predicts probable time spent much better than the old one. The "SI" is not defined in the post, and he gave no model size, latency, or evaluation numbers.
Both changes point at one metric. X wants a longer active session per user, and it now trains both stages toward that goal. If you want the older picture of the signals, our May analysis of the open-sourced code (how the X algorithm ranks replies, reposts, and links) and the August follow-up on For You timeline ranking weights show the earlier weights. Those explicit weights now sit beside, or under, a learned prediction of time.
What does the UAS chart really show?
A river of cards with one green card selected, showing a feed choosing which post to show
The chart plots daily user active seconds for 2023, 2024, 2025 and 2026, each with a dashed trend line. We read the following from the image. These are estimates read off an axis, not published figures.
| Period | Approximate daily UAS | Note |
|---|---|---|
| 2023 to 2025 | 330B to 375B | Flat or slightly down through each year |
| Jan to Feb 2026 | 370B to 400B | Near the 2025 range, slightly above |
| March 2026 | About 400B to 420B | A step up after an early March spike near 480B |
| April to August 2026 | About 400B to 420B | Flat, with weekly swings |
| September 2026 | About 440B to 460B | Rising |
| Last points (October) | About 470B to 505B | Highest point on the chart, near 505B |
Three observations matter for a reader who wants to judge the claim.
- The rise is real. The 2026 line sits well above every earlier year, and the final weeks are the highest on the chart.
- It is not exponential. From about 450B to about 500B is a gain of roughly 11 percent. A single week-to-week swing is about 40B, which is nearly the same size. Exponential growth would need a curve that keeps steepening, and two months of data cannot show that.
- The timing does not isolate the cause. The climb starts in late August. Musk announced the models on October 20. A rise that started earlier cannot be fully credited to a change announced later, and X did not say when each model reached users. Other changes, such as the Android app rewrite he mentioned in a reply, could contribute.
We cannot verify the numbers independently. The chart is a screenshot from the platform owner. Treat it as a claim with a plausible shape, not as audited data.
Why does "read all the way through" change what works?
A retrieval model that learns from finished reads rewards content where the reader has a reason to continue. That has a few direct consequences for how a post should be built.
- The first line has to earn the second. If readers leave after one line, the post has a low finish rate, and the model learns that this kind of post is a poor candidate.
- Length is not the target. A long post that people abandon scores worse than a short post that people finish. "Read all the way through" favors posts whose length matches their value.
- One idea per post. A post that answers one question and stops is easier to finish than a post with three half-developed points.
- Native content gains over link-outs. A reader who taps away to an article leaves the app. We have no published evidence on how the new models treat links. Our May reading of the code found that external links lowered reach, and nothing in the October announcement reverses that.
This matches advice we already give for answer engines. Put the answer first, support it with a number, and stop. Our guide to generative engine optimization and the GEO guide for marketers use the same structure, because both a retrieval model and an AI answer engine pick the passage that finishes the thought.
What are users saying in the replies?
The replies under the announcement are a small and noisy sample, but they repeat a few themes. We list them as reports from replying users, not as findings.
- Several users asked what is being done about spam and bot accounts.
- One user said they are tired of seeing the same copy-cat posts over and over.
- One creator said the platform's content quality is high and that they like seeing the algorithm work with creators.
- One user asked for a feature that detects when someone has gone too deep down a video rabbit hole, and said they would pay to turn off doom-scrolling.
- Another reply joked that everyone on the platform is now reading the latest algorithm changes, which is the creator reaction in one line.
The rabbit-hole and copy-cat comments point at a real limit. A model that predicts time spent has no built-in view of whether the time was useful or wanted. If a repeated format holds attention, the ranking model will learn to show more of it. That is the same risk we describe in our piece on AI slop and content quality.
What should you do this week?
A rounded eye with coins flowing toward it, showing attention as the currency X ranks on
Test before you rebuild. X has published no model details, so everything below is a hypothesis drawn from the stated training goal.
- Rewrite your first line. Put the result, number, or answer there. Compare impressions and the share of viewers who expand the post, for the old and new opening, on the same topic.
- Cut one paragraph from each long post. Keep the part a reader needs to finish.
- Stop reposting headline text. If your post matches ten others, the model has no reason to prefer yours, and readers have said they are tired of copies.
- Keep links out of the first post. Put the link in a reply, as we advised in May, and see if reach changes.
- Track a 14-day window. Weekly swings in the chart are as large as the recent gain. A single week of data will mislead you in either direction.
- Do not chase length. Time spent per user is the platform goal. Your goal is a reader who finishes and acts.
If you run an AI news or education account, the practical effect is that a well-built explainer thread can compete with a viral meme on the metric the model cares about. It does not guarantee reach. We also covered how fast Musk's own posts shape trends in Musk's viral X post and Grok trending, which is a reminder that platform-owner posts bend the feed in ways ordinary creators cannot copy.
What is still unknown?
| Open question | Why it matters |
|---|---|
| Model type and size | "AI/SI" is not defined. We cannot say if this is a Grok-family model or a separate system. |
| Rollout date and share of users | Without these, the chart cannot be tied to the change. |
| Whether Premium or link signals still apply | Our May reading found both. The October post does not say. |
| Whether quality filters exist | Replies about bots and copies ask for this. No answer was given. |
| Public code | Musk said in July that the full codebase would be published after a security review. See X open-source codebase promise. The new models are not linked. |
If the code appears, the best check is a diff against the version we analyzed in May. We will update this post if it does.
What this means for people who build on AI
Recommendation models are now an AI product category on their own. X's move shows the same pattern that applies to coding agents and chat products: the training label decides the behavior. Train on time spent, and the system finds ways to increase time spent. If you build a feed, a search tool, or an agent that ranks things for a user, write down the label you train on and ask what it fails to measure. If you want to learn how to build and evaluate such systems, our live workshops cover evaluation and agent design. Related coverage of how X is mixing models into its own assistant is in Grok routing to Claude Opus 5.5 and other models.
For teams that care about the safety of AI-driven feeds and agents that act on a user's behalf, AgentBeam is the agent security platform from the explainx.ai team, and it stops agents before they take dangerous actions.
Related reading
- X algorithm open source: replies, reposts, likes, and links (May 2026)
- X For You timeline ranking weights (August 2026)
- X open-sources its entire codebase: what Musk promised
- Musk's viral X post and Grok trending
- What is AI slop and why content quality matters
- What is SEO and GEO in 2026
- GEO for marketers: answer-engine guide
- Grok will route to Claude Opus 5.5 and other models
Primary source: the October 20, 2026 post by Elon Musk on X, with a daily user-active-seconds chart.
Figures in this post are read from a chart screenshot and from a single announcement post, and are accurate as of October 20, 2026. X has not published the models or an engineering write-up.