Cheap AI models aren't good at this, anyway, and AI Overviews have to use cheap models since they get used so much. They would be a lot better (still need to check, but they'd be much less stupid) if they used something like GPT-5, but that's just not feasible right now.
I had the same issue with Kagi, where I'd follow the citation and it would say the opposite of the summary.
A human can make sense of search results with a little time and effort, but current AI models don't seem to be able to.