One thing I like about Kagi's FastGPT is that it always includes citations in the output. This emphasises that I'm still doing a search, just with a fancier interface, not consulting an all-knowing oracle.

When I ask an AI tricky questions, the responses usually point me in the right direction without fully solving the problem, so those citation links are really useful.

With Bard in particular, I find it often gives a confident but incorrect answer. When I point this out, it apologises then gives another incorrect answer, and so on.