How do I run a fair content experiment with YouTube Studio's own tools, which variables to fix, how long, and why views alone cannot show cause?

Problem: deciding whether a change (hook, length, format, thumbnail) helps. Documentation (source_statement): YouTube's Tips page says to compare videos of the same format because audience behaviour differs across formats, and to group similar videos in Analytics and compare them in the watch-time report; the Audience page defines monthly audience as a rolling 28-day window and shows rolling windows of 7 and 28 days for other reports. The native thumbnail/title test (Test & Compare) is described in C3-B573-04. Method (inference by this worker, not a YouTube instruction): (1) write the hypothesis and one primary metric (for example average view duration or relative retention at 30 s, not raw views) before publishing; (2) change one variable and hold topic, format (Shorts vs long-form), length band, publish slot, title/thumbnail pattern and upload cadence constant; (3) publish at least several matched videos per arm, or use Test & Compare where it applies, rather than one video each; (4) compare over the same elapsed time (for example the first 28 days) not on different calendar windows; (5) check traffic source mix (search, browse, suggested, Shorts feed) because viewers differ by source; (6) log any outside event (a mention, a trend, a policy change) and discard contaminated videos; (7) report effects as 'associated with' unless randomised. Failure modes: comparing a Short with a long video; changing thumbnail and title and length together; reading a single outlier as an effect; survivorship (only keeping winners). Limitations: Studio gives aggregates, not user-level randomisation; small channels may never reach adequate n. Curator qualifications: All cited documentation excerpts located on the live pages by curator GET; no video sources. Limitations: The 7-step method is labelled inference; the Help pages do not provide an experimental protocol. Evidence status: statements located in the cited sources by a correlated desk review (Claude); not executed; not independently reviewed.

{
  "citations": [
    {
      "url": "https://support.google.com/youtube/answer/13616337?hl=en",
      "note": "When and how your viewers are watching; Actions you can take; Monthly audience"
    },
    {
      "url": "https://support.google.com/youtube/answer/9314416?hl=en",
      "note": "When and how your viewers are watching; Actions you can take; Monthly audience"
    }
  ]
}