X’s most visible content recommendation feature is the "For You” feed, accessed via the “For You” tab at the top of the home screen. The code that powers this functionality, along with other X feed generation features, has been open sourced, resulting in a variety of guides for optimizing posts to be more algorithm-friendly. This article looks at the algorithm from another direction: what is the “For You” tab presently serving to me?
Data for this experiment was gathered by scrolling through the “For You'“ feed until 200 posts (not including ads) had been displayed, and recording the account names and post content. This process was repeated on five occasions on five different days, resulting in 1000 recommended posts from 579 unique accounts. Slightly less than half of these posts (46.7%) were from accounts that I follow, with posts from accounts I do not follow making up a majority of the content (53.3%).
The account whose posts were most frequently recommended to me was Elon Musk’s official @elonmusk account, comprising 5.3% of the content in my “For You” tab over the course of the experiment. Since I do not follow Musk, his account is also the most frequently recommended account that I do not follow. Most of the non-followed recommended accounts are liberal influencers, but right wing content (including many of the Musk posts) does turn up here and there, along with content from random meme accounts such as @s8n. Interestingly, at least four of the posts I was recommended are from accounts that block me; it is currently unclear whether this is intended behavior or a side effect of recent changes to X’s block feature.
Due to the lack of the free X/Twitter API of days of yore, a full analysis of the recommended accounts is cumbersome to perform. Nonetheless, through use of archive sites such as Wayback Machine (along with a few Nitter instances), sufficient recent content from each account was gathered to plot a repost network. Almost all of the accounts recommended in the “For You” tab are connected via reposts via one or more degrees of separation to one or more accounts that I do follow that were recommended. Only 18 accounts (3.1%) have no discernible connection to accounts that I’ve followed; these 18 accounts are mostly crypto-themed, with a few random personal accounts with viral posts thrown in here and there.
One type of content that was almost entirely absent from my “For You” feed: AI-generated images. Although the experiments are not exactly parallel, this is a significant difference between the outcome of this analysis and a similar analysis of recommended Facebook content performed a few months ago. Another difference: The Facebook recommendations were almost entirely devoid of political content, which was very far from the case with the X “For You” tab.