X published its ranking weights and a label export

Every brand account has a theory about why its reach dropped. Almost all of them are guesses. X pushed a release on August 13 that swaps some of the guessing for source code: the X ranking algorithm repo now carries the actual scoring weights, the visibility-filtering logic, and training code for the model that ranks the For You feed. There's also a tool to download the labels applied to your own account.

What actually happened

The code lives at github.com/xai-org/x-algorithm under Apache 2.0. X has published versions of this before; the August 13 release is what got added. Per the repo:

• Configuration parameters for scoring weights — the numbers that decide how much a reply is worth versus a repost versus a favorite.
• Filtering decision code — the systems that determine whether a post gets shown, plus visibility filtering and labeling.
• Real training code for Phoenix, replacing a demonstration model.

Phoenix does two jobs. Retrieval: embed the viewer and every post as vectors, return the nearest ones. Ranking: predict the probability of each viewer action — favorite, reply, repost — then combine those into one relevance score that orders the feed. Candidates can't attend to each other in the transformer, only to the viewer context, so scores stay consistent and cacheable.

Alongside the code, TechCrunch reports a transparency tool that lets an account download aggregate stats as JSON showing which labels were applied over the past calendar month — piloting first for accounts at least a year old, and available to accounts posting at least ten times a month. Systems that use Grok to predict rule violations were held back.

Why it matters for your business

If X is a real channel for you, two concrete moves come out of this.

Pull your label file and stop arguing about vibes. A monthly JSON of applied labels turns "I think we got throttled" into a data point with a date on it. Check it against your posting calendar. If a label lands the same week you started dropping outbound links, you've learned something for free.

Read the weights, then act on the ratio — not the number. The value here isn't memorizing a coefficient that changes next release. It's the relative ordering: which engagement types the model predicts and scores highest. Optimize for the actions the ranker actually rewards. Then verify against your own analytics, because a public config is not a promise about your account.

The bigger operator lesson is the one we keep repeating: rented distribution is rented. This release makes the rules readable, not stable. A repo that updates on X's schedule can reprice your reach on X's schedule too. Keep the email list, the site, and the product data you own as the channel of record, and treat the feed as upside.

Sources: xai-org/x-algorithm on GitHub, TechCrunch.