Feynman diagrams conserve energy and momentum at the vertices, so at tree level, the one virtual particle's $q_{\mu}$ is determined by the real external particles. This is where the Mandelstam variables become useful (see fig).

Consider a $s$-channel process, such as:
$$ e^+ + e^- \rightarrow \gamma^* \rightarrow X $$
where the $*$ indicated virtual and $X$ is a fermion/anti-fermion pair.
In the lab frame:
$$ p^{\mu}_1 = (E, \vec p) $$
$$ p^{\mu}_2 = (E, -\vec p) $$
so the virtual photon has:
$$ q^{\mu} = p^{\mu}_1 + p^{\mu}_2 = (2E, 0, 0, 0) $$
So it's a massive photon at rest. Clearly virtual.
If you consider $t$-channel 180 degree scattering in the above case, then the final states are:
$$ p^{\mu}_3 = (E, -\vec p) $$
$$ p^{\mu}_4 = (E, +\vec p) $$
so that the exchanged photon has:
$$ q_{\mu} = p^{\mu}_4
- p^{\mu}_2 = (0, 2\vec p) $$
This is generally written as:
$$ Q^2 = -(q^{\mu}q_{\mu}) = 4||\vec p||^2 \approx 4E^2 $$
where the approximation is for $E \gg m_e$.
So here, the virtual photon has no energy, but large momentum (and negative mass-squared)...defintely virtual.
For any 2-particle EM scattering process, there is a frame (the Breit Frame) where the exchanged photon has no energy, and transfers 3 momentum.
Hence the interpretation is that the virtual photon probes structures with size $\hbar c/\sqrt{Q^2}$.
Finally, for identical particles you have $u$-channel in which the final state external lines are swapped. In 180 degree backscattering of indetical particles, you can work out that $q^{\mu} = (0,0,0,0)$, so no scattering occurs at all...that is, you can't distinguish forward and backward elastic scattering.
At higher order you have loops, and you have to integrate over all energy and momenta that are conserved at vertices, so you get way off-shell.