An offset costs headroom on one side and buys nothing. A file sitting 2% above zero clips 2% earlier on the positive peaks and has 2% of its level committed to a constant no speaker can reproduce.
What it does to a file
Three things, all of them annoying and none of them obvious by ear.
A cut at any point steps from the offset to zero, and a step is a click. Zero-crossing snapping does not help, because with an offset the waveform never crosses zero.
Peak normalisation reads the offset as part of the peak, so a file with a 2% offset normalises 2% quieter than one without.
Compressors and limiters see the offset as constant level. The gain reduction sits slightly wrong for the whole file.
The filter
A one-pole DC blocker: a zero at DC to remove the offset, and a pole just inside it to set how far up the removal reaches. Two state words per channel, no resonance, and nothing to overshoot on a transient.
The corner is where the response is 3 dB down. At the 8 Hz default the settling takes about 20 ms, so an offset is gone before the first note.
Corner frequency
Higher settles faster and reaches further into the low end. Lower is gentler and leaves subsonic content intact.
For speech, podcasts, and most music, 8 Hz is invisible. For material with real content in the bottom octave, drop to 2 or 3 Hz and accept the longer settling. At 0 the step passes the signal through untouched.
In a chain
Put it first. An offset is a fault in the capture, and every measurement made before it is removed, including loudness and peak, includes it.