A new paper shows user trust in LLM output is meaningfully influenced by citations— and that while the quality of those citations matters, the quantity doesn't