There have been quite sophisticated database engines around to do case law searches for a while, haven't there? Which makes sense, basing it on a corpus of known content, rather than 'the kind of thing that plausibly could exist, but doesn't', which is the step taken by generative LLM.

But it seems to me, as you say, that the skill lies in convincing a judge why this particular case is compellingly relevant to the litigation at hand, despite its inevitable differences.

And why the cases cited by the opponent, with the same cogent arguments as to their applicability, should in fact be disregarded or interpreted in another sense.