*This was merely too long for a comment.*

There is a certain preferable choice in a special situation, where one can choose an orthornormal basis of a rather specific nature.

Let $H$ be cocompact in $G$, then $L^2(H \backslash G)$ decomposes into a direct sum of unitary, irreducible subspaces $V_\pi$. Now assume furthermore that $G$ admits a large compact subgroup $K$ in the sense that $Res_K V_\pi$ decomposes with finite multiplicity, then you can construct an orthonormal basis of $V_\pi$ in terms of irreducible representations of $K$. This is often done in the theory of automorphic representations, where these orthonormal vectors are then called automorphic forms (or rather finite linear combinations of them).

For example, consider $H$ to be a uniform lattice in $G=SL_2(\mathbb{R})$ and $K=SO(2)$. The vectors introduced above will then be real analytic, but have no finite support.

In general, there is no hope for a canonical basis in terms of representations. Already for $H =SL_2(\mathbb{Z}) \subset G = SL_2(\mathbb{R})$, the decomposition will involve a direct integral of irreducible representations.