One way is to model the call stack explicitly. I.e. you will need objects (or structs, in Go) representing a stack frame, and have a stack of those (the actual call stack). Evaluation works very different that way, since you can no longer piggyback on the implementation language's call stack. It does mean that you can now do the desired tail call optimizations, e.g. if you're evaluating (if cond a b), and you've evaluated 'cond', then you can replace the if's stack frame with one containing just a or b, and continue evaluation there.
One benefit of this is that continuations become trivial to implement (they're just a snapshot of the call stack at a given point).
Admittedly, almost all write-your-own-Lisp tutorials omit this, and use a bunch of recursive calls (in the implementation language) instead, which makes it very hard to do TCO, continuations or anything else that requires a real call stack (exception tracebacks, for example).
(I hope this answer makes sense; it's 3 in the morning here, so clarity might suffer a bit. ;-) But as it happens, I am writing a (hopefully non-toy) Lisp interpreter myself, in Go, so the question attracted my interest.