Comparing Algorithmic Curation Across Social Media Platforms Using Sock Puppet Agents

Social media has become a primary channel through which people encounter news, including political news. What a feed shows is decided inside ranking and recommendation algorithms, and since their design is a black box, researchers have had difficulty studying what users are actually exposed to. This thesis examines three text-first social media platforms, X (formerly Twitter), Bluesky, and Truth Social, through an algorithmic audit based on automated sock-puppet accounts. Controlled personas with left-leaning, right-leaning, and neutral configurations were run on all three platforms under an identical design, and the feeds shown to them, 350,327 posts in total, were collected between December 2025 and May 2026 and classified with two complementary methods: bias ratings of linked news sources and large language model analysis of the post content. The feeds were then compared in terms of political exposure, ideological composition, diversity, content age, external linking, and thematic composition. The results show that the platform, rather than the persona, sets the baseline of political exposure. A no-signal feed is 20.7% political on X and 35.5% on Bluesky, but 78.5% on Truth Social, which remains heavily political and right-leaning for every persona examined. Bluesky is strongly left-leaning at baseline and responds only weakly to persona signals. X starts right of center and responds the most: a left-leaning persona shifts its feed by more than one full leaning category, while the other persona-platform combinations move far less. Truth Social also has the most concentrated and repetitive feeds, whereas the feeds on X are the most individualized and contain older content that keeps recirculating. The three platforms link to largely separate sets of external domains, and the personas see overlapping but distinct sets of topics. This ideological specialization at the platform level is consistent with users sorting themselves into platforms that align with their views. The audit pipeline is reusable on any ranked-feed platform, and the exposure data it produces is the kind of evidence needed for assessing the algorithms of online platforms.