Measure, do not assume. Print a calibration part with a known dimension, measure it with calipers, and scale by intended ÷ measured.
Worked: a 100 mm dimension that comes off the plate at 99.55 mm has shrunk 0.45%. The scale factor is 100.452%, so you model the dimension at 100.452 mm — a correction of 0.452 mm.
Why the correction is not the same as the shrinkage
This is the arithmetic people get wrong. The part shrank by 0.45%, but you do not scale up by 0.45% — you scale by 1 ÷ 0.9955, which is 0.452%.
The difference is tiny at PLA percentages and grows fast. At nylon's 2%, the naive correction is 2% and the correct one is 2.041%, which on a 100 mm part is a visible error in the compensation itself.
Compensate per axis, not globally
Contraction is not isotropic on an FDM part. X and Y contract fairly freely; Z is restrained by the layer stack bonded to the plate.
So measure both. A 50 mm dimension coming out at 49.7 mm on Y is 0.6% and wants a 100.6036% scale on that axis alone — a different figure from the X result above, from the same print.
A single uniform scale factor applied to all three axes will over-correct height every time.
What the measurement actually captures
Everything, which is the point. The observed difference includes material contraction, flow calibration error, extrusion width error and your machine's dimensional accuracy — all folded into one number that describes what your setup really does.
That makes a measured factor strictly better than any published shrinkage percentage, including the ones on this site.
When to re-derive it
After a material change, a nozzle change, a significant temperature change, or a flow calibration. Those are the four things that move it.
A factor derived for PLA at 210 °C does not transfer to PETG at 245 °C, and applying it anyway is how a compensated part ends up further out than an uncompensated one.
Design around it where you can
Slots instead of holes, clamps instead of interference fits, adjustable rather than fixed spacing. A part that tolerates a tenth of a millimetre is easier to make than one that needs a scale factor to be exactly right.
Derive your own factor in the shrinkage calculator, and see the full method in the shrinkage and tolerance guide.