The COVID-19 pandemic set the stage for a range of new conspiracy theories and extreme beliefs. Social media in general, and YouTube in particular, play a decisive role in the spreading of fringe content by not suppressing or even boosting the spread of such information. Users are predisposed to consume extreme content, creators prey on the emotions of users and game the algorithm to increase their popularity, and recommendation algorithms create feedback loops of progressively extreme content or ‘radicalization pipelines’. All three forces together may create information ‘bubbles’ that bias individual beliefs. We know, however, very little about the degree of information segregation and the structure of recommendation flows on YouTube, despite the platform’s position as the second most visited website in the world.This study maps the structure of COVID-19 information as a network of videos connected by inter-video recommendations. We collected nearly 10,000 videos and up to 20 recommendation links per video using COVID-related search terms on the YouTube API. To understand the structure of the graph we employed a mixed methods approach, using, a combination of quantitative graph analysis techniques and qualitative descriptions. We manually labeled the content of over a third of all videos and enriched the data further by merging with ideological labels from the Recfluence dataset. We then mapped these labels to pockets of videos that are well-connected to each other, as revealed by a community detection algorithm. Descriptive analyses of topical homogeneity within clusters of videos are supplemented with qualitative characterizations of the communities and the most important ‘bridging’ and ‘switching’ nodes within them. Contrary to popular beliefs about YouTube’s tendency to create informational ‘rabbit holes’ or ‘radicalization pipelines’, we find no evidence for algorithmic bias that would push users towards the consumption of fringe content.