Algorithm updates

X's new-author boost: how a small clip account gets a test slot

X's ranking code can lift one fresh original post from an account with 50,000 followers or fewer into a For You slot. What the code does and its limits.

Updated October 4, 20269 min read6 sources

In short

Yes, sometimes. X's open ranking code, as read on October 4, 2026, contains a new-author boost that can lift one original post per feed request toward the 16th scored position when the author has 50,000 followers or fewer, the post is at most 2 hours old and it has fewer than 200 Home Timeline views. The code ties the boost to an experiment, picks among eligible posts partly at random, and X widened its limits on September 30, 2026, so treat it as a test you may get, not a promise.

On this page

X built a boost for small accounts into its 2026 ranking code
Five conditions a post must meet
How the code chooses one post: Thompson sampling
What the lift does, and what it does not do
The boost runs as an experiment
X widened the boost on September 30, 2026
What a small clip account should do now
Limits of this reading
Questions
Sources

X built a boost for small accounts into its 2026 ranking code

X has said for a year that small accounts are hard to discover. On October 17, 2025, Elon Musk wrote that Grok-based ranking "should address the new user or small account problem", where someone posts something good and nobody sees it. On April 11, 2026, Nikita Bier, then X's Head of Product, wrote that aggregator accounts posting stolen reposts had "hurt new author growth", and cut their payouts.

The open ranking code in the xai-org/x-algorithm repository now contains a mechanism for this problem. The code calls it author cold start. The README calls it the new-author boost. It lives in home-mixer/scorers/author_cold_start.rs, and its settings are in home-mixer/params/param.rs.

This article describes only what that code does, as we read it on October 4, 2026 (commit b412112 of October 3, settings synced on October 2). X has not published a separate explanation of the boost. Where we interpret the code, we say so.

Five conditions a post must meet

Each time someone opens For You, X scores a list of candidate posts. The boost then looks for posts in that list that meet every condition below. All values are the code defaults on October 4, 2026.

Condition	Value in the code	What it means
Post type	Not a reply, not a repost	Only original posts qualify. Quote posts are not excluded by this check
Author size	50,000 followers or fewer	Counted on the author's account at scoring time
Post age	7,200 seconds or less	The post is at most 2 hours old
Post reach so far	Fewer than 200 Home Timeline views	Counted per post, on the Home Timeline only
Score	Non-zero and not in the bottom 3%	The post must rank inside the top 97% of non-zero scores in that request

If no post meets all five, nothing happens. If several do, the code picks one (see the next section). So at most one post per feed request gets the lift.

One wording difference: the README describes the boost as lifting posts "from authors whose impressions are below a threshold". The code measures the threshold per post (Home Timeline views of that post) and adds a separate follower limit on the author. We describe the code.

How the code chooses one post: Thompson sampling

When several posts qualify, the default method is Thompson sampling, a standard way to test unknown options. In plain terms, the code guesses each post's like rate, adds random noise that is larger when the post has few views, and favors the posts with the best guesses.

The code does this in three steps:

For each eligible post, it draws a random number from a Beta distribution built from the post's likes and its Home Timeline views. The starting values (0.75 and 49.25) assume a like rate of about 1.5% before any data comes in.
It keeps the 2 posts with the highest random draws.
Of those 2, it picks the one with the higher ranking score.

Our reading of what this means for you: a post with very few views has a wide range of possible draws, so it gets a real chance even with zero likes. A post with, say, 150 views and a good share of likes keeps a high draw. A post with 150 views and no likes draws low and rarely wins. Early likes from real viewers therefore help a post win the slot, and the 200-view limit closes the window fast.

Before September 30, 2026, Thompson sampling was switched off by default and the code simply picked the eligible post with the highest score.

What the lift does, and what it does not do

The code takes the scores in the request, sorts them, and reads the score at a position drawn between the slot settings. With the current settings (minimum 15, maximum 16), that is always the 16th-highest score. The chosen post's score is raised to that value if it is lower. A post that already scores higher keeps its score.

This is a floor, not a fixed position. Several steps still run afterwards:

Out-of-network discount. If the viewer does not follow you, which is the normal case for a new account, the score is multiplied by 0.75 after the lift.
Author diversity. If you have several posts in the same request, each one after the first is scaled down.
Similarity reranking. A reranker reorders posts so that very similar posts are not next to each other.
Visibility filtering. Labels such as spam labels can still remove the post for non-followers after ranking.

If the request has fewer than 16 scored candidates, the code skips the boost. Also note what the boost never touches: replies, reposts and posts older than 2 hours. The general ranking still decides everything else; see what X's ranking weights reward.

The boost runs as an experiment

The code ties the boost to an experiment. It reads two settings that place the viewer in a control or treatment group of an experiment on X's retrieval model, or in neither (a holdout). It reads two more settings that place each author in a control or treatment group.

The rule is simple. A viewer in neither group can receive the boost for any eligible author. A viewer in the control group can receive it only for authors in the control group. A viewer in the treatment group can receive it only for authors in the treatment group. The code also counts how many posts it lifts per group, which is how a team measures an experiment.

The main switch for the boost is on (true) in the October 2026 defaults. But the README says many values come from a configuration system rather than the code, and that X runs experiments on part of its traffic. So the boost can be on for some viewers, off for others, and changed without a public announcement. Treat it as an experiment X is running, not a feature X has promised.

X widened the boost on September 30, 2026

The commit history of param.rs shows that the settings changed a lot in seven weeks. The commit dated September 30, 2026 (settings synced on September 29) changed five values at once:

Setting	Before September 30, 2026	From September 30, 2026
Follower limit for the author	1,000 or fewer	50,000 or fewer
Maximum post age	48 hours (24 hours until September 17)	2 hours
Home Timeline views on the post	Fewer than 1,000	Fewer than 200
Score range that can qualify	Top 85% of non-zero scores	Top 97% of non-zero scores
How one post is chosen	Highest score	Thompson sampling

The same commit set Phoenix retrieval, the system that finds posts from accounts a viewer does not follow, to request up to 200 "cold start" results per request, up from 0. The repository passes this number to the retrieval service but does not show how that service chooses those posts.

Our reading: X moved from a slow, narrow boost (tiny accounts, two days) to a fast, wide one (most small and mid-size accounts, two hours). For clip pages between 1,000 and 50,000 followers this is new: before September 30 they were outside the boost. The settings can move again, so check param.rs before you plan around them.

What a small clip account should do now

Post clips as original posts. Replies and reposts never qualify. If you share someone else's clip as a repost, it cannot get your boost.
Make the first 2 hours count. The boost window closes 2 hours after posting or at 200 Home Timeline views. Post when your audience is awake, and see posting speed and freshness.
Earn real early likes. Thompson sampling favors posts whose early viewers like them. Cut tight, put the strongest line first and write a clear post text. See hooks for clips.
Do not flood. Only one post per feed request gets the lift, and your extra posts in the same request are scaled down. Two strong clips beat ten weak ones.
Never buy or trade early engagement. Fake likes can raise the draw, but X's spam systems and campaign rules both target them. One spam label hides your posts from non-followers, which removes the reason for the boost. See flagged views and view bots.
Do not split one audience across many accounts. Several small accounts each stay under 50,000 followers, but X bans posting the same content across multiple accounts. See running several accounts.
Stay original. X's statements about this problem frame stolen reposts as the thing that hurt new authors. A small account built on reuploads works against the reason the boost exists. See reupload detection on X.

Limits of this reading

We read code, not production logs. Four limits apply. First, the code defaults are what X says it syncs from production, but experiments can override them. Second, the boost only sets a minimum score; later steps can still move the post down. Third, the retrieval service that supplies cold-start candidates is not fully in the repository. Fourth, X has made no public statement that names this boost, its limits or its results.

So the honest summary is: X's code gives fresh original posts from accounts with 50,000 followers or fewer a chance at a mid-feed slot, chosen partly at random, inside an experiment. It does not guarantee reach to anyone.

Questions

Does the boost apply to accounts with more than 50,000 followers?

Not under the October 4, 2026 defaults. The code checks that the author has 50,000 followers or fewer. Until the September 30, 2026 commit, the limit was 1,000.

Do quote posts count as original posts for the boost?

The eligibility check in the code excludes replies and reposts only. It does not exclude quote posts. Whether quotes get the boost in practice also depends on the other conditions and on the experiment group.

Does Premium or a blue check help me get the boost?

The cold-start code does not check Premium or verification. It checks post type, follower count, post age, Home Timeline views, score rank and experiment groups.

Can I see whether my post got the boost?

No. X does not show this in analytics. The code only records counts for X's internal metrics. A sudden burst of views from non-followers in the first hours fits a boost but has other causes too.

Is the boost the same as the old advice to post at the right time?

No. Timing advice is general. The boost is a specific rule in the code that only looks at posts at most 2 hours old with fewer than 200 Home Timeline views, so timing matters because of a hard limit.

Will deleting and reposting a clip give it a second chance?

It would create a new post that is under 2 hours old, but X lists "posting and deleting the same content repeatedly" as spam, and campaign platforms require posts to stay up. The risk outweighs the chance.

Sources

We checked every claim on this page against these sources on the date above. Platform rules and rates change often. Found an error? Tell us.

In short

Questions

Sources

Related guides