How do you measure semantic similarity in topic clustering?

Short answer: Use three measurements together: cosine similarity between text embedding vectors, SERP overlap between queries, and query families that already land on the same page in Search Console. Relying on a single method produces wrong clusters.

Practical measurement method

• Embedding base: vectorize queries and page titles with an embedding model and form candidate clusters via cosine similarity; algorithms such as k-means or HDBSCAN scale the work.
• SERP overlap test: if two queries return largely the same first results page, one page suffices; if they differ, separate pages are needed. This test captures intent differences best.
• GSC validation: queries already earning impressions on the same URL are natural cluster evidence.
• Editorial control: algorithmic clusters get human confirmation; commercial intent differences are often invisible in the vector.

Report per cluster: impressions, clicks, and query coverage per cluster on a single line, and re-evaluate pages that drift.