Start with a 10-by-10 square of tiles and remove a 3-by-3 square from one corner, leaving 100−9=91 tiles. Cut the leftover L-shape into two strips along the edge of the missing corner and slide one strip around: together they form a rectangle 10+3=13 long and 10−3=7 wide, and 13×7=91. That is a2−b2=(a+b)(a−b) with a=10 and b=3.
Algebra shows the same thing. Multiply (a+b)(a−b) and you get a2−ab+ab−b2. The two cross terms are +ab and −ab, so they cancel exactly, leaving a2−b2. Whenever you see a square minus a square, you can run that multiplication backward.
The squares can be disguised. 16x2−49 is (4x)2−72, so it factors as (4x+7)(4x−7): the coefficient 16 is part of the square. x4−81 is (x2)2−92=(x2+9)(x2−9), and x2−9 is again a difference of squares, so the full answer is (x2+9)(x+3)(x−3).
Two look-alikes do not fit. A sum of squares, x2+9, has no real factors: (x+3)(x−3) gives x2−9, and (x+3)2 adds a middle term 6x. And x2−6x+9 has a middle term, so it is a perfect-square trinomial, (x−3)2, not a difference of squares. Count the terms and check the sign first.
Think of it like this. Reshape a square garden into a rectangle with the same perimeter, b metres longer and b metres narrower, and you always lose exactly a b-by-b patch: (a+b)(a−b)=a2−b2. Run that backward and any square minus a square becomes such a rectangle. The trick depends on that missing patch, so it says nothing about a2+b2.
Core idea. When two perfect squares are subtracted, the expression factors into the conjugate pair (a+b)(a−b).