Topic Embeddings & Site Radius
Google computes vector embeddings for entire websites. Understanding siteEmbedding, siteFocusScore, and siteRadius gives you a mathematical framework for topical authority decisions.
What you'll learn
- Understand what vector embeddings are and how Google uses them for sites
- Learn what siteRadius measures and how to minimize it
- Know how siteFocusScore is affected by off-topic content
- Understand cosine similarity and how to use it for content decisions
- Learn how to detect and fix TopicalDrift using embedding logic
Vector Embeddings: What Google Computes for Your Site
Every page on your site has a semantic vector — a mathematical representation of its meaning in high-dimensional space. Google also computes a siteEmbedding: the average semantic vector of all pages on the domain. This site-level vector represents "what is this site about" as a single mathematical point.
The siteEmbedding is not just metadata — it directly influences how Google evaluates the relevance of individual pages on your domain. A page about "CFD trading leverage" on a site with a strong CFD trading siteEmbedding gets a credibility boost from the domain-level signal. The same page on a general blog with a diffuse siteEmbedding does not.
siteFocusScore: How Topically Consistent Is Your Site
siteFocusScore measures how tightly clustered your pages are around the site's central topic. A site with 100 pages all about CFD trading has a high focus score. A site with 80 pages about CFD trading and 20 pages about cooking recipes has a lower focus score — the off-topic pages pull the siteEmbedding away from trading and reduce Google's confidence in the domain's topical identity.
The practical implication: publishing off-topic content doesn't just fail to help the site — it actively hurts the authority of the main topic cluster. Every page on the site contributes to the siteEmbedding calculation. An off-topic page that gets traffic from other sources still dilutes topical focus.
siteRadius: The Spread of Your Content in Semantic Space
siteRadius measures how far individual pages spread from the site's central embedding vector. A site where all pages cluster closely around one topic has a small radius. A site covering many loosely related topics has a large radius.
Google associates small siteRadius with deep expertise and specialization. Large siteRadius signals breadth without depth — the opposite of topical authority. The goal is to maximize depth (many pages on closely related subtopics) while minimizing breadth (avoiding topics distant from the central embedding).
Cosine Similarity: The Relevance Test
Cosine similarity is the mathematical measure of how similar two vectors are in direction (regardless of magnitude). A cosine similarity of 1.0 means two vectors point in exactly the same direction — perfectly related topics. A similarity of 0.0 means completely unrelated. Negative values indicate opposing concepts.
For content decisions, use cosine similarity to validate whether a proposed new topic belongs in your topical schema:
- Similarity 0.7+ — core topic cluster; will reinforce siteEmbedding
- Similarity 0.5–0.7 — adjacent topic; acceptable if the SILO is clearly bounded
- Similarity below 0.5 — increases siteRadius; should not be published unless as a distinct sub-brand or subdomain
Detecting and Fixing TopicalDrift
TopicalDrift occurs when a page's content vector drifts away from its intended SILO's centroid. Drift is detected when a page starts ranking for queries outside its intended topic scope — a sign that its content has accumulated enough off-topic material to shift its embedding.
Fix TopicalDrift by: removing or condensing off-topic sections, strengthening the page's on-topic EAV coverage, and updating internal links so that most linking context comes from semantically close pages in the same SILO.
Key Takeaways
- siteEmbedding is the average semantic vector of all pages — it represents what Google thinks your site is about
- Off-topic content dilutes siteEmbedding and reduces topical authority for your main topic cluster
- siteRadius measures content spread — small radius signals deep expertise; large radius signals generalism
- Cosine similarity below 0.5 is a signal that a proposed topic should not be added to the site
- TopicalDrift is when a page's embedding shifts away from its SILO — fixed by removing off-topic content and updating linking context