I’ve noticed this too. A single result can determine the answer it gives. And removing the content from its context makes it harder to assess. Suddenly it’s “Gemini said …” rather than “some guy in the YouTube comments said”.