The point
A decoupling capacitor works by supplying transient current faster than the supply rail can. Its effectiveness is set almost entirely by the loop inductance between the cap and the IC’s power pins - and loop inductance is loop area.
A 100 nF cap two centimeters from the pin, returning through a long ground trace, can present tens of nanohenries. At a 100 MHz edge, that inductance is worth more impedance than the capacitor is worth capacitance.
The rule that falls out
Place the cap so the current loop - pin, cap, and the return path underneath
- encloses as little area as possible. Via straight down to the ground plane, directly at the pad. Distance is inductance; inductance is the enemy.