Word similarity compares shared words with all distinct words in each pair. Each term joins its closest matching seed above the chosen threshold; it needs two shared significant words where possible. Seeds start with the highest supplied volume, then alphabetical order. Groups do not expand through chains of loosely related terms.

Review guards separate supplied intent labels and English free, cost, comparison, guide, negation and buying modifiers. Other location, product and audience differences still need your review. Text in other languages is preserved and compared as Unicode words; stemming and modifier guards focus on English. Synonyms with different wording may stay separate.

Reference word match uses the internal tool’s local-source rule: two shared words (or fewer for short terms), simple English endings and the first matching seed. It ignores intent and group-size settings. Neither method uses AI embeddings or live search results.