The problem is that FIMO is using a statistical comparison to determine if a bit of sequence is a more likely match to the background frequencies or to the expected symbols from the motif. The shorter the motif, the less reliably FIMO can make this distinction. By default FIMO applies a significance threshold (p-value) of 0.0001. You could try setting a less stringent p-value threshold
using the '--thresh' option. However, the higher you set the threshold, the more likely that FIMO will report matches that are simply due to chance. If you have a motif that is 7 characters wide, then the probability of perfectly matching that motif by chance is roughly (0.25)^7 = 0.000061. That may seem small, but if you scan the human genome with that motif you'd expect to find roughly 366000 perfect matches to the motif entirely by chance.