It would be very undesirable for the spatial index grid sizes to be recalculated automatically each time you added more features. That would add an enormous amount of overhead, and would not be likely to yield any improvement in the index itself.
Ideally, the grid sizes that are calculated should achieve a balance between a grid size that is too large or too small. If the grid size is too large, the result will be an index that is not sufficiently selective, that is, each grid cell contains an excessive number of features which means that a large number of false positives could be returned from the index scan. On the other hand, too small of a grid size causes the index to become very large, due to the fact that each grid cell intersected by the feature's envelope must have a separate reference for that feature. This slows down the index scan because it has to do a much larger amount of file i/o to read through the index.
The grid index calculation algorithm works best when the size distribution of the features currently contained in the feature class is representative. Ideally, all the features should be present, and then the algorithm is run. If only some of the features are present, and they are not representative of the true size distribution, the calculation will be thrown off. For that reason, the best workflow is to enter load-only mode, add all of your features, and then exit load-only mode. Exiting load-only mode triggers the grid size calculation and the loading of the index.
Here is a brief description of how the algorithm works:
The envelope of each feature is obtained. We look at the delta X and delta Y of the envelope, and use the larger of the two.
The resulting delta values undergo a logarithmic transformation, and the resulting values are used to create a histogram which represents the size distribution of the features.
The histogram is statistically analyzed and the initial grid size that is selected is approximately at the 66th percentile of the size distribution.
Depending on the number of features that are beyond the 66th percentile, it is possible that a second and even a third grid size might be calculated.
The algorithm seems to do a reasonably good job, and is vastly superior to the situation that existed prior to 9.2.
I hope this helps in understanding how the index works.