What the Corpus Still Holds

Researchers often share YouTube comment data as comment IDs plus their own labels, and expect others to retrieve the comments through the YouTube Data API. Whether this works, and what is lost, has rarely been measured on real deposits. I tested three public deposits in October 2026, from the oldest to the youngest: 62,414 comments on COVID-19 videos (collected 2020) 173,012 comments on the U.S. Army's channel (2023); 16,652 comments on precision fermentation (2024). Every step was fixed in writing before it was run. Findings: The deposited IDs still work as keys. 74.4%, 93.2% and 94.8% of the comments can still be retrieved. The loss comes in blocks: whole videos and whole threads. On YouTube, a reply almost never outlives the comment it answers, so a reply that corrects a claim disappears with the claim. The more hostile the corpus, the more its hostile comments are lost: no difference in the low-conflict corpus, a slight tilt in the contested one, and 4.8 percentage points (5.4 on videos that still exist) in the oldest and most hostile one. Yet every deposited distribution moves by less than one percentage point. One deposit contains video IDs destroyed by spreadsheet export (#NAME?). For distributions, a loss of this size is a nuisance rather than a threat; for analyses of threads or individual users it may be one. The paper closes with five checks for anyone who shares or reuses YouTube comment data. Who it is for: linguists, discourse analysts and communication researchers who build, share or reuse YouTube comment corpora, and the data stewards who curate them. This is the third part of a small series on YouTube comment corpora. The three papers complement each other: What the Corpus Does Not Say (https://doi.org/10.5281/zenodo.23339873): what a corpus leaves out when it is collected What the Corpus Can No Longer Find (https://doi.org/10.5281/zenodo.23642703): whether a deposited corpus can still be traced back to the platform This paper: how much of a deposited corpus survives and whether the loss changes its results. All scripts, outputs and design documents are in the verification package: https://doi.org/10.5281/zenodo.23805596. Every offline step reproduces byte for byte from the original deposits. No comment text and no author information is contained in it. Not peer reviewed. Refutations, review, or comments are much appreciated: [email protected]