Embedding clustering groups 23% of terms n-gram methods scatter

Embedding clustering groups 23% of terms n-gram methods scatter
Keyword grouping by shared words (n-grams) is fast but literal. Embedding-based clustering (grouping by meaning vectors) catches synonyms n-grams miss. We clustered 3,000 terms both ways.
— Embeddings merged 'laptop' + 'notebook computer' + 'portable PC': ▇▇▇▇▇ n-grams left these in 3 buckets
— N-grams over-split synonym families by 23%
— Embeddings can over-merge distinct intents (cost vs price vs free) without an intent guardrail
— N-gram grouping needs no model; embeddings need one API pass
So what: embed-cluster for topical structure, then layer a SERP or intent check to stop semantic-but-different terms from collapsing into one page.
Benchmark of the week: embedding clustering cuts the number of pages needed for a topic by ~20% versus exact-word grouping, before any quality loss.