Candidate words are grouped in the browser by identifying contiguous runs of tokens that exceed the confidence threshold and merging their character offsets into unified anchor phrases.