Topic-Embeddings: How Google Understands What Your Page Is About

01What are Topic-Embeddings?

Topic-Embeddings are Google's representation of what a page is about. They work not with keywords but in a high-dimensional vector space, where similar topics sit close together. The system produces two values: a focus vector (where the page's topic sits in the embedding space) and a radius (how broad the page's topic coverage is).

The Google API leak confirmed that Topic-Embedding fields exist. The focus vector is self-supervised: it is derived from the actual text of the site, from the real usage of language across all pages, without human labels. The resulting vectors flow into NSR (System 7) as focus vectors.

Topic-Embeddings are the modern replacement for keyword matching. Instead of checking whether a page contains the word 'weather,' the system checks whether the page's topic vector is close to the query's topic vector in the embedding space. So Google can match 'storm warning' with 'severe weather alert' even if the exact words don't overlap.

02Focus and radius: what they mean

The focus vector marks the center of a page's topic in the embedding space, that is, the point where the page's subject sits among all subjects. A page about 'Hamburg weather' has a focus vector close to other weather-related pages, but closer to pages about Hamburg than to pages about Munich weather. The focus vector captures both the topic (weather) and the specificity (Hamburg).

The radius says how broad a page's topic coverage is. A page that discusses only Hamburg weather has a small radius: it is tightly focused. A page that discusses weather, climate, tourism, and local history has a large radius: it covers many topics. The radius affects how Google matches the page to queries. A small-radius page matches a narrow set of queries very well. A large-radius page matches a broad set of queries less precisely.

The focus vector determines which queries a page is relevant to, and the radius determines how confident Google is in that relevance. A page with a small radius and a focus vector close to the query will rank higher than a page with a large radius and the same focus vector. The reason: the small-radius page is more clearly about the query's topic.

03Topic-Embeddings in the ranking architecture

Topic-Embeddings feed into NSR (System 7), the normalized site rank. The focus vector becomes one of the inputs that NSR uses to compute the site-level quality and authority value. This means topic relevance is not just a matching function. It is part of the quality assessment.

Topic-Embeddings map to two E-E-A-T dimensions: Expertise (Exp: indirect) and Authority (A: indirect). A page that is tightly focused on a specific topic signals expertise: it demonstrates deep knowledge of one area rather than shallow knowledge of many. The authority signal comes from the focus vector's proximity to other authoritative pages on the same topic.

Topic-Embeddings operate at the site level (Reach: Site), not the document level. siteEmbeddings, siteFocusScore and siteRadius are all site-scoped fields. The leak's own description of siteRadius names page-level embeddings only as an internal input: it measures how far individual page embeddings deviate from the site's aggregate embedding. But the value Google stores and scores is the site-level number. A site where most pages sit in the weather region of the embedding space will have a high site-level siteFocusScore. That makes it more likely to rank for weather queries, but the unit of measurement is the site, not any single page.

04siteFocusScore: the field itself

siteFocusScore is the specific leaked field that gives Topic-Embeddings its name in practice. The Content Warehouse module QualityAuthorityTopicEmbeddingsVersionedItem documents it, in Google's own wording, as a 'Number denoting how much a site is focused on one topic.' Its sibling field siteRadius is documented right next to it, as 'The measure of how far page_embeddings deviate from the site_embedding' — the two are the same measurement seen from opposite ends: a high siteFocusScore corresponds to a small siteRadius.

Mike King, who published the leak analysis, paraphrases the pair this way: siteFocusScore captures how much the site sticks to a single topic, and siteRadius captures how far the site strays from its core topic. That paraphrase is not itself leaked wording. It is King's reading of the field. But it lines up with the verbatim descriptions above.

The Quality Rater Guidelines treat Authoritativeness as topic-bound: a site can be authoritative for one subject and not for another. siteFocusScore is the clearest known mechanical counterpart to that idea. It is a single numeric field that asks, in effect, 'is this site about one thing.' The field does not reveal its weight, its exact computation, or whether it acts as a direct multiplier or as a normalization input elsewhere in ranking.

05Implications for SEO practitioners

For large publishers, the focus vector creates a strategic tension: broad coverage (many topics, large radius) vs. narrow focus (few topics, small radius). A weather site that also covers travel, cooking, and sports will have a large radius, which means its weather pages will rank lower for weather queries than a pure weather site's pages.

Topic-Embeddings are self-supervised, so the system learns from the actual text on the page. A focus vector cannot be 'injected' through meta tags or structured data. It is derived from the content itself. If the page is about weather, the focus vector will reflect that, regardless of what the meta tags claim.

For sites with programmatically generated content, the radius matters. Thousands of location pages with genuine weather data will have small-radius focus vectors, because each page is tightly focused on one location's weather. This is good for ranking: the small radius makes each page clearly relevant to its specific location query.

The embedding-based approach also explains why synonyms work: 'storm warning' and 'severe weather alert' will have focus vectors that are close in the embedding space, even though the words don't overlap. Google doesn't need the exact keyword — it needs the right topic vector.

06References