What the Corpus Does Not Say
Reply structure, time and coverage in YouTube comment data — what linguists and other researchers should check before treating comments as conversation
PDFZenodo · latest versionVerification package
A living methods paper. Its premises — what yt-dlp and the YouTube Data API actually deliver — are re-checked every six months, and each result is published as a new version. Last check: 24 October 2026, all premises unchanged. Next check: May 2027. The DOI above always resolves to the latest version.
Abstract
Linguists increasingly read YouTube comment sections as conversation: who answers whom, how address is marked, what follows what, and when. None of this is observed directly. It is reconstructed twice — by the platform when it stores and serves a comment, and by the extraction tool when it reads it back. That thread position and actual address can come apart is established in CMC research, including for YouTube. This report adds measurements on the author's own YouTube corpora of how far, where and why they come apart under two extraction routes, and closes with checks that take minutes to run. For discourse and conversation analysts, pragmaticians and CMC corpus linguists, the result is a map of which readings of comment data — adjacency, address, sequence and time — the delivered fields can support, and which they cannot.
Six findings are reported. (1) The official Data API attaches every reply to the top-level comment; in two videos collected by both routes, 48.5% and 42.3% of replies were assigned different parents. (2) Interface-level extraction recovers deeper structure but is capped: across 210 extractions of 133 videos (480,925 comments) no reply was returned below depth 3, and at the capped level the address marker and the recorded parent disagree in 45.8% of resolved cases against 3 in 13,407 one level up. (3) The timestamp emitted by interface-level extraction is a back-calculation from a relative label such as “1 year ago”. (4) The address marker has artefacts of its own. (5) Corpora silently lose comments through deletion, reply limits set to zero and blocked runs. (6) A manual coding pass specified in advance confirmed 78 of 100 divergent cases as genuine, and the conclusion drawn from it was wrong, because the mechanism operated at the level of tree position.
A deposited verification package recomputes 72 published figures from text-free tables. Because tools and platforms change, the premises are re-checked every six months under a fixed protocol and each result is published as a new version. The first re-check, on 24 October 2026 with a newer yt-dlp release, found every premise unchanged.
This edition supersedes three earlier preprints by the author: Six Methodological Pitfalls in YouTube Comment Extraction, Flattened Conversations and Reconstructed Time. They remain available with their version histories.
The series
This is the first of three measurement papers on YouTube comment corpora. What the Corpus Can No Longer Find asks whether a deposited corpus can still be traced back to the platform; What the Corpus Still Holds asks how much of a deposited corpus survives and whether the loss changes its results.
Comments
Refutations, reviews and comments are most welcome: research@woidwide.net. If relevant work has been overlooked, I apologise and would be glad to be told.
AI use: the author designed the study and set all analytic decisions and decision rules; Claude (Anthropic) was used under his direction to draft text and analysis code and to compute the figures, and the author revised the text. The paper sets out the division of labour in detail.
Citation
Sedlmair, Max (2026). What the Corpus Does Not Say: Reply structure, time and coverage in YouTube comment data. Preprint, version 1.7. Zenodo. https://doi.org/10.5281/zenodo.23339873