Ok. So where scores are arranged on a normal distribution (a bell shaped curve), at a population level each test's scores cluster around the average (the mean). For every single subtest, 2 in every 3 people tested will get scores in that broad average range.
There are a number of tests that make up a cognitive assessment. For example, the WISC has 10 subtests, 7 of which go forward for calculation of FSIQ. So when an IQ is being calculated, whilst some variance in subtest scores is expected, most people will have a range that at at least some points hits or crosses the average band, because at a population level for each subtest there is a 67 percent chance of the score being in that band.
So a person who has a range of scores which might all be around 5-10th percentile, for example, has had 10 chances to hit the average band and yet they have not done so. This makes their learning profile rarer than someone who has some lower scores and some in the average band (the same is true for very high scores, where all are higher than the average band). This phenomenon at a population level is called regression to the mean because most people will have at least some scores that regress to the broad average (mean) band.
This is why it's really important to calculate FSIQ. Because a person with a range of scores around 3-10th percentile in all the subtests will get a full scale score that is lower, probably around 2nd percentile. Because unlike most people, they didn't hit that average band in any of the contributing scores, and so they are rare, and don't have particular relative strengths they can draw on to compensate for the areas they find more difficult.
I hope that makes sense!