Information about token order, added to the embeddings because attention alone is order-blind.
Broad technical consensus on what this term means. Disagreements that invoke it are about the world, not the word.
Attention compares every position with every other and has no inherent notion of sequence. Without a positional signal a model would treat a sentence as an unordered bag of tokens.
Schemes vary, from fixed sinusoids to learned vectors to rotary encodings applied inside attention itself. The choice matters most for how gracefully a model handles sequences longer than the ones it was trained on.