How the X algorithm decides what people see

What X actually made public

In March 2023 X released the source for the recommendation system behind the algorithmic timeline, along with an engineering write-up describing the pipeline. That is unusual: most feed ranking is inferred from experiments, whereas here the candidate sourcing, the model layers and many of the heuristics were written down.

Two caveats matter before drawing conclusions. The release was a snapshot of one moment, and the system has been developed continuously since. It also documents mechanics rather than promises, so nothing in it commits the company to keeping any particular weighting.

Where candidate posts come from

Ranking begins by gathering a large pool of candidates from two broad sources. In-network candidates come from accounts the reader follows. Out-of-network candidates come from everywhere else, found through techniques such as looking at what accounts with similar following patterns engaged with, and through embedding-based similarity between communities.

That second path is why a post from an account with few followers can still travel. It is also why relevance to a recognisable community matters more than broad appeal: the system needs to place you near a cluster of readers before it can suggest you to them.

• In-network: material from accounts the reader already follows.
• Out-of-network: surfaced through engagement patterns of similar users and community similarity.
• The pool is then narrowed by a light scoring pass before expensive ranking runs.

What the ranking model is trying to predict

The heavier ranking stage is a learned model estimating the probability of various reader responses: replying, reposting, favouriting, opening the post, watching a video through, and so on. Responses that take effort and indicate genuine interest have historically carried far more weight than passive ones.

Understood properly this is liberating rather than mystical. You cannot optimise a coefficient, but you can write things that a specific person would plausibly answer or forward. Content that provokes a reply from somebody who knows the subject is doing exactly what the model is trying to detect.

The signals that quietly cost the most

Predicted negative reactions carry heavy penalties. A reader blocking or muting you, reporting a post, or marking it as uninteresting are strong statements, and the system is deliberately tuned to avoid repeating whatever produced them.

This is the honest argument against engagement-bait, reply-farming and mass-tagging. Those tactics do raise raw interaction, but they raise irritation at the same time, and irritation is weighted to hurt more than the interaction helps. The same reasoning applies to padding an audience with accounts that never respond.

• Blocks, mutes, reports and not-interested marks are strongly negative.
• Repeated near-identical posting looks like spam to both readers and filters.
• Reach lost this way recovers slowly, because the penalty applies before anyone sees you.

What this changes about how you post

Very little of this rewards cleverness about the system, and most of it rewards being legible to a particular community. Post consistently enough to be recognisable, keep to subjects where a knowledgeable reader might reply, and give a post enough context to stand alone when it appears in front of somebody who has never seen you.

The rest is hygiene: do not manufacture engagement, do not repeat yourself mechanically, and do not chase reach from audiences who will find you irrelevant. Recency still matters, which is why a predictable publishing rhythm helps, but freshness cannot rescue a post nobody wants to answer.