It think it depends on what you want to study. You have indicated that you want to study the effect of various bias voltages, but it was unclear if the effect of the bias on the geometry was considered important. If so, i.e. if you need to re-optimize the structure for each bias, then you definitely need to design the flow carefully as it will be quite time-consuming, and you need the zero bias reference case to really be a minimum.
Moreover, there is always a fundamental question in a case like this if the distance between the gold surfaces should be "free", i.e. needs to be optimized, or considered fixed. Keeping it fixed, but of course reasonably, is typically fine; perhaps one can keep the atoms closest to the molecules free, esp. if it's a tip-like (pyramidal) shape.
Finally, note that the Gaussian optimize structure will probably not be an exact minimum structure when using the QuantumATK pseudopotentials and basis sets, so while that is a good start guess, it's only that.
Now given that
a) you already optimized the structure under different voltages.
b) the electrodes do not move as a function of voltage
I would keep the coordinates of the molecule as they are (since any internal rearrangement due to bias is captured in this) and add electrodes to the outside at a fixed distance from the molecular edge atoms in each case.
It should be generally noted that the absolute values for the current you compute are anyway not precise. What you are (I presume) interested in here is to see how the bias affects the transmission, i.e. to compute the I-V curve, and this is captured by this approach, at least to a first approximation, even if each structure is not strictly speaking an optimized structure in QuantumATK. As a second level simulation one could re-optimize each structure under bias, but keeping the gold electrodes at a fixed distance still.
