(The problem, in fact, is that the significance of monads in Haskell in terms of the side-effect model has nothing to do with category theory. The "side-effects" quite literally occur outside of these theoretical constructs.)
For example, it would be weird if we had three statements A B C that wrote to console, and sequencing A with B and then sequencing the result with C resulted in a different program than sequencing B with C and then sequencing A with the result.
Or, for non-IO things: it would be weird if we had three levels of nested lists, and concatenating two times "from the outside" resulted in a different list than concatenating each inner list first, and then the outer list.