Large language models (LLMs) often generate synthesized answers without offering easily inspectable evidence. We collected 489,283 assistant‑user response pairs from ShareChat and WildChat to examine how answer features relate to users' subsequent acceptance, checking, and challenging.\n\nBroad attribution language appears in 31.9% of qualifying responses, while explicit URLs appear in only 1.33%, revealing a gap between claimed grounding and actually inspectable support. Attribution cues, uncertainty expressions, and other signals tend to co‑occur with both checking and challenging; only uncertainty shows a modest statistical tilt toward checking.\n\nIn Technical Help interactions, uncertainty favors checking over challenging; in Seeking Information scenarios, refusals and reasoning traces tend to provoke challenging. Scrutiny is concentrated in a minority of conversations, and the apparent turn‑to‑turn persistence diminishes markedly when conversation‑level differences and response ordering are examined separately.\n\nThese findings suggest that supporting verification requires attention to the inspectability of evidence, the type of question the user seeks to resolve, and the accumulated history of the interaction.\n\nReview: The study highlights the impact of LLM opacity on user scrutiny and offers empirical guidance for building more verifiable conversational systems.