
Embeddings and vectorization for SEO: what they are and how to use them
Understand what embeddings are, how Google uses vectors to measure relevance and which practical applications exist for SEO and GEO. With sources and real examples.
highlights
- Google has operated with semantic models since 2013 (Word2Vec/Hummingbird) and 2019 (BERT). Most SEO tools still work on the lexical model.
- An embedding is a numeric vector that represents meaning. Proximity in vector space equals semantic similarity, measured by cosine similarity.
- Chunking is mandatory: models have a token limit per input, and long, unstructured pages produce less precise embeddings.
- There are five direct applications of embeddings in SEO: internal linking, semantic cannibalization, competitive analysis, contextual clustering and Site Focus Score.
- For GEO/AI Search, the implication is even more direct: RAG systems split content into chunks for retrieval, and paragraph structure becomes relevance infrastructure.
what google understands when it reads a page
In 2013, Google researchers published the Word2Vec paper. The model learned to represent words as vectors in a high-dimensional space, capturing semantic relationships no lexical system could. The classic example: king - man + woman ≈ queen. It isn’t string matching. It’s arithmetic on meaning.
That same year, Google launched Hummingbird, its first algorithm update to incorporate semantic understanding at scale. In 2019 came BERT, which brought bidirectional language models into ranking. Since then, Google’s retrieval system has worked on two planes at once: lexical (term matching) and semantic (approximation by meaning).
Marc Najork of Google Research detailed these hybrid approaches in a presentation referenced by iPullRank in its survey of vectorization and SEO. The key point: the two layers coexist, but the semantic layer carries growing weight, especially for ambiguous and long-tail queries.
The implication for SEO is straightforward. Optimizing only for lexical matching means optimizing for a version of Google that hasn’t existed for more than a decade.
what an embedding really is
A vector is an ordered list of numbers. An embedding is a vector generated to represent meaning, not just raw data.
When a language model processes a text, it turns it into a point in a high-dimensional space. The mxbai-embed-large model, for example, generates 1024-dimensional vectors. Each dimension captures some implicit semantic property learned during training.
The geometric consequence is that pages about related topics sit close together in that space. A page about “adult dog food” and another about “healthy nutrition for Labradors” are a short distance apart. A page about “DSLR cameras” is far from both. This proximity is measured by cosine similarity: the angle between the vectors. The smaller the angle, the greater the similarity. Screaming Frog explains the mechanics precisely when describing how to map related pages at scale.
The crucial distinction is between dense embeddings (BERT, mxbai-embed-large) and sparse embeddings (TF-IDF). TF-IDF creates a vector in which each dimension represents a word in the vocabulary, lexical by definition. Dense embeddings compress meaning into a fixed vector of hundreds of dimensions, able to capture synonyms, contextual relationships and implicit concepts no word count would ever capture.
the chunking problem
Embedding models have a token limit per input. mxbai-embed-large accepts up to 512 tokens. Typical content pages have far more than that.
The solution is chunking: splitting the text into smaller windows that overlap, so you don’t lose context at the edges of each segment. A 260-token chunk (measured with tiktoken cl100k_base) with a 40-token overlap is a reasonable starting point. Note that tiktoken cl100k_base underestimates mxbai-embed-large’s real token count by roughly 30–50% for Portuguese content: a 260-token cl100k chunk results in at most ~380 real tokens, within a safe margin below the 512 limit.
iPullRank documents that short paragraphs focused on a single concept generate more specific embeddings with a better chance of matching queries. A dense paragraph covering multiple subtopics produces an “average” embedding that represents everything imprecisely.
For GEO and AI Search, the implication is even more direct. RAG (Retrieval-Augmented Generation) systems, used by the leading AI assistants, retrieve chunks, not pages. A poorly formed chunk, with multiple concepts mixed together or abrupt transitions between ideas, is less likely to be retrieved to answer a specific query. Paragraph structure becomes relevance infrastructure.
five practical applications of embeddings in seo
Each of these applications has its own technical and editorial implications. The articles in this series cover each one in detail. Here is the overview.
Internal linking by semantic similarity
Instead of linking pages by keyword matching, calculate cosine similarity between every pair of crawled URLs and suggest links based on real proximity of meaning. Screaming Frog details the workflow.
Semantic cannibalization detection
Two articles can compete for the same territory of meaning without sharing a single keyword. A cosine similarity matrix with alert thresholds (0.85–0.93 for warning, above 0.93 for critical) identifies these pairs before they cause ranking confusion. Screaming Frog v22.0 introduced this kind of analysis natively.
Vector-based competitive analysis
Mapping semantic overlap between your own site and competitors reveals what keyword gap misses: two sites can compete for the same thematic territory without sharing a single keyword. iPullRank details paragraph-level analysis.
Contextual clustering
Grouping pages by embeddings reveals the content architecture algorithms see, not the category structure the editorial team defined. The two rarely match perfectly, and the gap is actionable.
Site Focus Score and thematic coherence
A site’s semantic centroid is the vector average of all its embeddings. Each page’s distance from the centroid indicates how far that page strays from the core topic. The 2024 Google documentation leak produced signs that the site focus concept is monitored internally, but this isn’t conclusively documented yet: it’s an inference based on unofficial sources. Screaming Frog v22.0 uses the concept of a "most representative page".
what changes in practice
The shift from the lexical to the semantic model changes three things day to day.
First, content production. The question stops being “is this keyword in the text?” and becomes “does this text cover the semantic cluster of the search intent?” An article can leave out the exact keyword and still rank for it if its embedding is semantically close enough.
Second, auditing. Keyword cannibalization is visible to any tool. Semantic cannibalization, where two pages compete for the same meaning with different vocabulary, only shows up in vector analysis.
Third, GEO. For generative AI systems, paragraph structure is what determines whether a specific chunk of your content will be retrieved to compose an answer. This isn’t an abstraction: it’s the mechanics of RAG systems described in the public documentation of most AI assistants available today.
Most SEO tools still operate on the lexical model. Embeddings are no longer a technical curiosity. For audits that aim to capture what Google actually evaluates, they are infrastructure.
final thoughts
The argument isn’t that keyword research or TF-IDF stopped working. It’s that they capture one layer of the problem, and there’s a second layer that only shows up in vector space.
The good news: the tools for working with embeddings in SEO are accessible. Screaming Frog v22.0 integrates semantic analysis natively. Python with sentence-transformers handles the basic workflow for smaller sites without additional infrastructure.
What still requires human judgment is interpreting the results: a cosine similarity score identifies the problem, but the editorial decision to consolidate, differentiate or discard pages still belongs to the analyst.

Lucas Cassapula
Partner & Head of SEO at Wesearch and Co-founder of Mentionflow
I am a partner at Wesearch and co-founder of Mentionflow. I have worked with SEO for almost 10 years. I am a data-driven geek who is always testing hypotheses, looking for patterns and turning ideas into products. I share studies, experiments and automations focused on SEO and GEO.