Flattened Conversations: Reply-Structure Loss in YouTube Comment Data

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

If you study replies on YouTube, your choice of collection tool decides who answered whom. This report measures by how much, on identical comments, and says which tool to use for what. Two videos were collected both through the official YouTube Data API and through interface-level extraction (yt-dlp). The two methods assign different parents to 48.5% and 42.3% of replies. Every divergent case has the same shape: where the interface records that C answered B, the API records that C answered the top-level comment A. This is the API working as documented — across 43,896 replies its parentId points to a top-level comment in 100.0% of cases — but its size had not been measured. Across four corpora (n = 267,505), 10.8% to 65.5% of replies answer another reply, most in exactly the conversational material interaction research selects for. The interface method recovers that structure only up to a platform nesting cap at depth 3, so the usable gain over the API is 8.6–24.3% of replies, not 42–49%. Below the cap the recovered edges are the best-supported in the data: across all four corpora, the platform-inserted address marker agrees with the depth-2 parent in all but 3 of 18,114 resolved cases. At the cap, agreement falls to 48–82%. The two exports also differ in text. For the same comments, the API omits the space after an inserted address marker in 15.7% of marked replies, where the interface text keeps it — a trap for any tool that resolves mentions from API text. Practical rule: take time from the API, structure from interface-level extraction below the cap, record depth on every edge, and keep the two parent fields apart.