'Surely, you don;t(sic) mean that you would dismiss something because it's "not always relevant"?
If it applied on only 5% of the cases, it's still useful then, right?'
I think we can quibble for ever over the usefulness of a metric that is not applicable in 95% of cases. Of course we can use our skill and judgement to evaluate the importance of a low cache hit ratio is meaningful. But how much investigation is required to determine whether it matters? Is that the best use of our time? In most tuning situations I come across these days it doesn't seem relevant. YMMV (Don's obviously does).
Funnily enough one of the other threads I mentioned in my previous article demonstrated a situation where a low cache hit ratio was indeed an indicator of a bad situation. When the OP finally posted the relevant init.ora parameters we discovered
buffer cache size = 32 mb
shared pool = 76 mb
As Laurent Schneider observed, that's barely enough memory for a pocket calculator.