Six Methodological Pitfalls in YouTube Comment Extraction: What a rendering decision does to a research corpus.

Abstract YouTube comment threads are widely analysed as interaction, on the assumption that the data record who replied to whom. They do not, quite. Reply structure is reconstructed twice before a researcher sees it — once by the platform when it stores a reply, once by the extraction tool when it reads it back — and this report documents six points at which that reconstruction misleads. The central finding is a nesting constraint. YouTube serves at most four levels of comment depth, numbered 0 to 3. Across 210 extractions of 133 distinct videos in which replies were collected — 480,925 comments, 238,700 of them replies — not one reached depth 4. Pooled across the collection the deepest level carries more comments than the level above it, which is the pile-up a constraint produces rather than the thinning tail a conversation produces. Replies to threads that run deeper are attached at the deepest available level regardless of whom they address, so edges at that level are censored interface attachments rather than validated conversational parent relations. Where the constraint does not bind, at depth 2, the address marker and the recorded parent agree in 13,435 of 13,439 resolved cases; one level down they disagree in 45.8%. A parameter sweep rules out the extraction tool as the source: asked for nine levels, it receives four. Five further pitfalls are documented with the same treatment — mechanism, a diagnostic that runs in minutes, evidence, consequence. They are: the official API's flattening of replies onto the thread root; timestamps that are back-calculated from coarse relative labels rather than measured; channel-handle migration silently breaking name resolution; orphaned replies re-attached to the thread root after a parent is deleted; and collectors that return only top-level comments while leaving a populated parent field behind. The last is documented from two of the author's own collections, together 258 videos and 89,074 comments in which the reply layer was absent and nothing in the files said so. The sixth pitfall is methodological. A coding pass specified in advance, on 100 randomly drawn divergent cases, correctly classified 78 as genuine address divergence — and the inference drawn from it was wrong, because a coder looking at one case cannot see where that case sits in a distribution. Manual validation rules out parsing artefacts. It does not rule out structural ones. This is a validation audit rather than a discovery report. Its unit of contribution is a set of checks, and nothing in it depends on being first.