Flattened Conversations: Reply-Structure Loss in YouTube Comment Data

How Much Conversation Remains? Reconstruction Rates in Flattened Comment Architectures YouTube stores sequential structure only to the second level. Replies to replies appear in a flat list, so uptake is carried lexically rather than structurally, largely through @-address. Sequential work on such data therefore requires reconstruction — a procedure established for Twitter by Honeycutt and Herring (2009) and transferred here to a different platform architecture. What this paper adds is not a method but a measurement, and the measurement turns out to matter. Across a corpus of 14,087 comments, 20 per cent of contributions have a resolvable addressee, yielding 5,020 dyadic links, of which 361 — 2.6 per cent of the original material — form chains of at least three turns. The gap between corpus size and analysable sequence is thus close to two orders of magnitude. This has a direct consequence for how such studies are reported. A corpus of 100,000 comments says nothing about sequential tractability so long as the proportion sitting in recoverable chains remains unstated. The reconstruction rate places a ceiling on any claim about the frequency of an interactional practice, converts into a yield per thousand raw comments for advance corpus planning, and makes findings from differently structured platforms comparable. We propose it be reported as standard. The argument holds for any platform with limited nesting. The corpus construction, reconstruction pipeline and drafting of this paper were carried out with the assistance of a large language model. We report this in full, and note that the resulting materials may serve as a case for studies examining the quality of research produced by non-specialists working with AI assistance. We make no claim as to how that case should be judge