Flattened Conversations: Reply-Structure Loss in YouTube Comment Data

Research on YouTube comment sections routinely reconstructs who replied to whom. This report measures what happens to that reconstruction when comments are collected through the official YouTube Data API v3 rather than through interface-level extraction. For one video collected by both methods (n = 13,484 comments present in both exports), the two files encode structurally different reply graphs. The API export assigns every reply to the top-level comment of its thread; the interface-level export assigns 3,032 of 6,248 replies (48.5 %) to a parent that is itself a reply. None of these immediate reply relations is recoverable from the API export. Across four corpora (n = 267,505), the share of replies whose parent is itself a reply ranges from 10.8 % to 65 %, and is highest in precisely the conversational corpora that interaction research selects for. This is not an API defect: the two-level comment model is documented. The contribution is quantitative. A blind coherence validation on 50 hand-checked pairs supports the interface-level structure (90.0 %, 95 % CI 76.9–96.0), with self-reply edges scoring lower than cross-author edges. Implication: studies that infer immediate reply-to-reply structure from official API parent relations risk systematic flattening of nested interaction.