Statistical enrichments are always relative to some background set. If you do not provide a background set, then GREAT uses a whole genome background. Such a background is not always appropriate. You should provide a background set to account for any inherent biases in sampling.
For example, suppose that the input set of genomic regions consists of ChIP-chip peaks of regions found to be bound by a particular transcription factor. An appropriate background set is the genomic regions represented on the microarray chip. Some criteria was used to determine which of all possible genomic regions were present on the microarray, and a proper background set allows GREAT to account for inherent biases due to the selection criteria.
See Gill:relevant paper for a more detailed discussion of the importance of using a proper background.