How does the X algorithm work

People ask how the X algorithm works as if there were a hidden lever. The honest version is narrower: quote what X has published, including the public ranking code, and refuse the rest. A README is not turned into a growth recipe here. If a specific post looks dead, diagnosis belongs on a planned impressions guide, not here.

What has X published about how ranking works?

X has published a public repository for the code that builds the For You feed, and that repo is the main official source quoted here. It describes in-network posts, out-of-network retrieval, filtering, and a transformer ranker. If a popular thread claims a rule those sources do not contain, treat it as unverified.

Do not blend guru numerology with X's words.

On 10 November 2026, America/Los_Angeles, this page opened github.com/xai-org/x-algorithm. GitHub lists the repo description as "Algorithm powering the For You feed on X." The README title is "X For You Feed Algorithm," and it opens: "This repository contains the core code that determines which posts a viewer sees in the For You feed on X." The latest commit fetched that day was 660cd7407c57, dated 9 November 2026, 5:53 PM PT, with the message "Open-source X Recommendation Algorithm."

A search on 10 November 2026 treated unofficial "how the algorithm works" explainers as common results. Those posts often list undocumented factors. This page stays with the repo and dated official material.

What does the public ranker code actually include?

The README, as fetched on 10 November 2026, says the For You feed is assembled from two places: posts from accounts the viewer follows, and posts found outside that network. Both are ranked together. Phoenix predicts how likely the viewer is to take each action, and those predictions are combined into one score with weights held in the code.

In-network posts are kept in thunder/. Out-of-network posts are found by Phoenix retrieval and SimClusters, plus a shared list of popular recent posts.

The same README is explicit about what is missing. Under "What's not in this repo?" it says some files are not currently published, "e.g. Grox prompts" and "Some botmaker rules," because "people could use it to try to game the system." It also says some deployment-related files may not be in the snapshot. The code is a source, not proof that every production scorer matches this checkout. Do not add weights that are not in the files you cite.

The README also points to an Under the Hood transparency tool that, it says, lets people see aggregate statistics about visibility-impacting labels on their account and posts. That is a published claim about a transparency surface, not a diagnosis of one quiet post.

What should a founder not infer from the code?

Do not treat a published scorer as a promise that doing one thing will multiply reach. Do not extract a secret posting time or a magic keyword list. The code can explain categories of signals. It cannot tell you that your demo video failed because of one undocumented penalty.

The README itself warns against a common misreading of the weights. It says the weights scale predicted probabilities of actions, "they do *not* scale the raw engagement counts," so it would be incorrect to see that "a report has 468 times higher weight than a like and conclude that e.g. '1 report cancels out 468 likes'." That sentence is the repo's own caution, not a growth tip. This guide does not reprint those weights as a playbook.

How do replies and original posts show up in that picture?

Quote only what the published material says. The README lists reply as one predicted engagement, alongside favorite, repost, quote, and others. An Out-of-Network Discount, it says, also applies to replies and reposts from accounts the viewer already follows. That is a published category, not a reason to stop replying.

How to write replies belongs on the planned reply-to-grow guide, which is a separate guide. If official Premium documentation discusses reply ranking, that question belongs on the planned Premium-reach page, covered in a separate guide, and no price is printed here.

If a specific post looks dead, go to the planned guide on why a post gets no impressions on X. That page owns diagnosis, including visibility labels and the shadowban question. This algorithm page stays at the published-system level and does not reprint those steps. A quiet post is not automatically proof that you decoded the ranker wrong.

Evenpost is an app for X (Twitter) for solo founders. You and your agent draft, you approve every post in the Evenpost app, and it posts with natural spacing, or right away if you choose Post now. Updates: @evenpost. It does not claim to beat, tune, or replace the For You ranker.

For a related question about whether posting times are a lever, see Best time to post on X. For X's link-post fee, see X's link post fee.

Read what X has actually published, including the public ranker repo, and ignore undocumented factor lists when a quiet post needs diagnosis instead.

FAQ

Is the public GitHub repo the live production algorithm?

The repo describes itself as the core code that determines which posts a viewer sees in the For You feed. It also says some files are not published, and that some deployment-related code may be missing. Treat the checkout you read as a dated snapshot, not as proof that every production scorer matches it today.

Does posting more often beat the algorithm?

The published README does not say that posting more often beats ranking. This guide does not invent that rule. If a posting cadence question comes up, start from what you have to say, not from a volume trick.

Do links hurt ranking?

This guide does not invent a link penalty. X's published link-post fee is a separate, live answer. The README lists a predicted link click among scoring signals, which is not the same as a documented penalty for posting a URL.

Does X Premium change how the algorithm treats you?

If official Premium documentation discusses reply ranking, that belongs on the planned Premium-reach page, and no price is printed here. The public For You README fetched on 10 November 2026 is not a Premium pricing page.

Can anyone really "hack" the X algorithm?

Not from this README, and not honestly. The repo even withholds some files because people could try to game the system. Undocumented factor lists are not a strategy. If one post looks quiet, diagnose that post instead of hunting for a secret lever.