We have a tool to do exactly what you need. It's called Densify Sampling Network, and it's in the Geostatistical Analyst toolbox, under Sampling Network Design toolset.
You will use your kriging layer as input, then specify the number of new locations for lowering the overall prediction standard errors. It will create a point feature class that determines the best new locations for monitoring sites. However, this technique attempts to minimize the overall standard errors; it won't give preferences for particular locations (like cities).
If you're only interested in low standard errors near cities of interest, there's another technique you can try. The trick with this technique is that the standard errors only depend on the locations of the data points, not the data values themselves (unless you applied a transformation). In other words, if you add a new point and give it a value of 1, the standard error surface will be the same as using a value of 1 million (this is obviously not true for the prediction surface). So, you can create artificial new points and test how they will affect the standard error surface. The easiest way to do this is to copy your original point feature class, then append new points near cities that you're interested in (you can give them any data value you want). Then use the Create Geostatistical Layer tool. Give the tool the original kriging layer you created with the original points, then provide the appended point feature class for the input dataset(s). Run the tool, then look at the standard error surface and decide if it's acceptable.
But remember, this last paragraph will only work if you did not apply a transformation when doing the kriging. If you did, then the standard errors actually do depend on the data values, so you can't just make up new data values.
As for your idea, I don't think it will work. At least, I don't see how it would work, but maybe I'm just missing something.
