In the picture above you can see two staircase. The top is modeled with 32 triangles to cleanly fill the gap beneath it. The lower is modeled with 2 triangles that are disconnected from the steps above it and serve to only fill the hole. When having a low polygon count is a requirement is there any reason other then not having clean topology to not create your models like the lower example?
If you care about the real time performance of your model, then forget topology and focus on removing every single unwanted face(poly). You may, when modeling keep good topology in order to expedite and simplify the process. At the end you can go back and remove the remaining extraneous faces before moving onto UV unwrapping and texturing. That is purely a matter of preference. If you don't care about the real time performance, then by all means focus on good topology.
It's much easier to modify the upper version afterwards than the lower one. the lower one would be totally messed up if you'd change anything in the mesh. I'd recomment to find something in between.I would propably use quads instead of triangles below the stairs in the upper example.