I think Steve is right. You are probably doing unconditional simulation when you want to be doing conditional simulation.
To do conditional simulation, you must provide the dataset and the field that you use to create the kriging layer into the Gaussian Geostatistical Simulations tool. You provide them in the "Input conditioning features" and "Conditioning field" parameters.
The idea here is that there are an infinite number of surfaces that all have the same covariance structure (ie, the same semivariogram). When you do an unconditional simulation, you simply create several of them at random. However, these surfaces do not pass through the same set of points, and their high/low values will not occur in the same places. When you average over all these simulations, you will get close to a constant raster.
But there are also an infinite number of surfaces that all have the same covariance structure and are conditioned to pass through a given set of points. You can specify these conditioning points to be anything that you want, but the most common thing to do is to condition that the simulations must pass through the input points from the kriging layer. When you do this, all simulations will resemble the original kriging layer, and when you take an average of the simulations, you will get something close to the original layer. The more simulations that you perform, the closer the average will look to the original layer. Steve's picture shows how this works.
Both conditional and unconditional simulations have uses, but it can sometimes be tricky to tell which one you should use. In your case, it sounds like you want to be doing conditional simulations and conditioning on the features that were used to create the kriging layer.