X shows its hand on how posts get ranked

Not every action on X carries the same weight with its recommendation system. The company released the source code behind its "For You" algorithm on Thursday, giving users their clearest look yet at how posts are ranked, boosted, or buried.

This algorithm, named Phoenix, analyses the viewing behaviour of the user recently in order to understand how he would react to the posts made by the accounts he does not follow.

A prediction of link sharing using a copied URL has a value of around forty times compared to the value of like. On the other hand, reply, quote, and direct-message share have a value of five, ten times the value of, whereas a prediction of follow has an equivalent of eight likes, while repost has two.

Phoenix combines all these predicted probabilities multiplied by their respective weights to form one scoring rank.

A predicted report weighs 468 times as heavily as a like in the opposite direction. A predicted mute carries roughly 118 times the negative weight of a like; "not interested" clicks land around 86 times as heavily, and blocks about 62 times.

The disparity suggests that posts designed purely to provoke anger, often called ragebait, may not actually help accounts grow, despite X's reputation for hosting combative content.

Alongside the code, X introduced "Under the Hood", a pilot feature letting select users view aggregate data on labels applied to their account and posts.

The software also categorises content as spam or pornography, making its own separate visibility judgements. X mentioned that the software is currently being tested on a random sample of eligible users, and its wider rollout depends on the feedback from users.

In January, he referred to the platform's algorithm as "dumb" and pledged to make the system more transparent, saying that X stands apart from other social networking sites, as it enables the public to observe its operations in real-time.