Spatial autocorrelation evaluates a set of features and an associated attribute to determine whether the pattern expressed is clustered, dispersed, or random. If the feature does not have any neighbors within the distance band it cannot compare that features values to its neighbors. As you increase the distance threshold the lone features may eventually get neighbors, but it will also increase the number of neighbors for features which are close together.
By using generate spatial weights matrix (Generate Spatial Weights Matrix—ArcGIS Pro | ArcGIS Desktop ) you have more control over what features are considered as neighbors. You can use the Conceptualization of Spatial Relationships - K nearest neighbors option so that every feature is given the same number of neighbors and no feature will have 0 neighbors. The spatial weights matrix can be used as input to the Spatial Autocorrelation tool (Spatial Autocorrelation (Global Moran's I)—ArcGIS Pro | ArcGIS Desktop ) but not the incremental version (which increases the distance to evaluate its effect).
There may be distance bands for which every feature has neighbors, that is indicated in the tool messages output. You can also identify the features with no neighbors to further investigate the impact they have in your data set.
