Optical character recognition reads pictures of text. Everything that makes text hard for you to read makes it harder for the engine.
Scan at 300 DPI or better
Below roughly 200 DPI accuracy drops sharply. We render PDF pages at 300 DPI before recognition for exactly this reason.
Keep it straight and high contrast
Skewed pages, shadows and low contrast all cost accuracy. A flatbed scan beats a phone photo.
Pick the right language
The language setting decides which character shapes the engine expects. Choosing the wrong one produces confident nonsense.