Any matrix can be decomposed into a sum of tensor products where each term is of the form $X^a Z^b$ (where $a$ and $b$ are bits; they can be 0 or 1). For example, the 16 matrices of the form $(X^{a_1} Z^{b_1}) \otimes (X^{a_2} Z^{b_2})$ are a basis for the space of 4x4 matrices.
All errors can be thought of as an unwanted matrix applied to your system. And any unwanted matrix can be put into the XZ tensor product basis. So general errors decompose into linear combinations of $X$ and $Z$ errors.
What this means is that if you just focus on correcting combinations of $X$ and $Z$ errors, you end up correcting everything. And $X$ and $Z$ errors have all this structure that makes correcting them easy to think about. They're sparse, they only have real values, they're Hermitian, they're unitary, they're self-inverse, they're closed under Clifford operations; it's so convenient!