Sibling comments have good points but what made it click for me was the book "Exceptional C++" where all the possible sources of exceptions are pointed out -- it's usually twice as many as you would first guess by looking at the code, sometimes made more complicated by type-converting constructors and operator overloading and RAII.
Then consider that if any code you're calling is not exception-safe it makes the task of writing your code to be exception safe that much harder.
Then add on top of that the question of responsibility -- do you handle the exception close to where it gets thrown or farther up the call stack? Handling it too close to the issue may be missing some context and result in a lot of the same exception handling across many areas of the codebase. But handling it too far up misses context, too, and often leaves the program in an uncertain state where it's not clear whether the show must go on or if it's time to shut it down. It's not uncommon to see a mix of exceptions and returned error values to try and find a happy middle ground there.
Java tried to paper over some of these problems with fewer gotchas (more explicit separation between resource allocation and initialization) and compiler checks that exceptions are in the method signature and are always handled _somewhere_ up the call stack but this often results in handling of very vague exception types at a high level. Some developers, uncertain what the right way to handle an exception is, and not wanting to crash the program, will just silently drop the exception instead! Admittedly, you can do this in go-style and C-style error handling, too, but at least then it's not as far removed from the source of an error because it's annoying to have to keep passing an error response through so many function signatures.
I used to use exceptions a lot in C++ and Java. This changed when I started working at a place where a lot of C++ is used but without exceptions. It was ostensibly about runtime costs but when seniors were pressed on the issue it was clear there were a lot of these other reasons stated above, and ultimately about readability of the code (error handling being close to error source) and a philosophy of failing quickly when assertions fail (because keeping calm and carrying on leads to programs running that should have died before they could make more of a mess).
I know it's unpopular but, in light of some real problems with exceptions, I actually prefer the way Go makes you do it. It often encourages doing all the setup in one (or a few) function scopes, and it becomes pretty evident from a function's signature whether you should expect something to go wrong or not. Would I prefer it for a game-dev scripting language? no! But for a system language it is better IMHO