Let's recall something briefly. A linear transformation of one vector space to another has an existence that doesn't depend on bases.
Bases are just "skeletons" of the vector space that we pick so that we can concretely write out the transformations in terms of matrices. Different choices of bases yield different matrices for a transformation, but the transformation itself does not require any of the bases. Only representations of the transformation require a basis.
We are interested in properties that all the representations share, because then we can rightly say it's a property of the transformation and not just a quirk about one particular representation. In a sense, when we pick a basis and represent a transformation with it, we have "zoomed in too far" and we might overlook information about "the big picture" (the transformation itself). That's why we study invariants, so we know what things in "the small picture" actually reflect the big picture.
So, for example, one thing we know is invariant between similar matrices is the determinant. This demonstrates that the determinant is a property of the transformation and not really the matrix. It can be computed from the matrix, but you'll get the same answer no matter what basis you picked.