Pick a sensor by working backwards from the measurement you need, not from the part you have seen before.
Range and resolution first
The range must cover the worst case with margin, but an oversized range wastes resolution — a 500 kg load cell used for 5 kg readings puts your signal in the noise. Resolution is the smallest change you can see; check what your ADC can actually resolve after amplification, since a 12-bit converter over a 10 V span gives ~2.4 mV steps.
Accuracy is not precision
Accuracy is how close you are to truth; precision/repeatability is how tightly readings cluster. Control problems usually need repeatability; calibration and acceptance need accuracy. Also budget for drift (with time and temperature), hysteresis and non-linearity — these often dwarf the headline accuracy figure.
Output type decides your electronics
A raw bridge (load cell, strain gauge) gives millivolts and needs an instrumentation amp — see the load-cell tool. 4–20 mA is the right answer for long runs in a noisy plant: current is immune to voltage drop and a broken wire reads 0 mA, which is diagnosable (loop tool). Digital (I²C/SPI) skips analogue problems but is short-range. Thermocouples need cold-junction compensation; RTDs and NTCs need excitation and linearisation (thermistor tool).
Then the environment
Temperature range, IP rating, vibration, chemicals and EMI kill more sensors than overload does. Confirm the mounting does not preload or distort the sensing element, and plan the cable route away from motor drives (see grounding & EMI).