If you spend enough time doing mental simulation, you can get a reasonably close prediction of how the map will play. It'll save you development time in the long run.
True, but that can be a trap. Doing mental simulations can end up reinforcing bad choices, or creating areas that have only one strategy to deal with. I spend 75% of my time in Hammer flying around and asking myself what any particular class should or would do in a given area, and I have to make a conscious effort to step back from my assumptions once I get a demo/gameday feedback and see that what I would do in a situation doesn't necessarily translate to what most (or even some) people would do in that situation.
If a bot builds a sentry in a spot that seems bizarre, you should ask yourself why the bot thought that area was good. Examine it and compare to your intended sentry spots. It's about challenging your viewpoints, not looking for new sentry spots.
Again, better to have this tested with actual players.
And awkward areas, they might only exist because their nav tells them to take a weird direction. That awkward area might only be like that because its supossed to contain only 3 players instead of 10.
That's not what I meant by "awkward areas." I mean areas that feel fine when you walk through them and looked fine scale-wise in Hammer, but when you see what's happening there, when you see the various classes moving through it, it looks clumsy and poorly designed. Just seeing bodies in motion in the area can change your perspective compared to seeing a scale reference in Hammer in the same spot.
By the way, I don't count MvM as "testing with bots". You're not trying to simulate players in that mode because the bots are a feature there, not a method.
In response to both, you should note that I said very clearly
player testing is better in every way at the end of the post. But for someone who maybe can't get in a game day or impromptu, and is still very early in the design process, you can make use of bots as long as it's only for
very specific and limited testing.
It's a tool, like anything else, but one that is not as useful as others.