Definition: what are embeddings?
An embedding translates a piece of language into coordinates: a model maps text to a vector so that semantically similar content receives similar vectors. Two texts with zero word overlap can sit almost identically in the embedding space if they express the same concept — and two texts with nearly the same wording far apart if they mean opposite things.
Typical dimensionalities of common text-embedding models: 768 (BERT base), 1,024 (BGE-M3, multilingual), 1,536 to 3,072 (current OpenAI embedding models). Similarity is computed via distance measures — most commonly cosine similarity, alongside dot product and Euclidean distance.
Why embeddings changed search
Classic search works with string matching and term weighting (BM25, TF-IDF). Modern systems transform query and document into vectors and search the embedding space (vector search) — usually hybrid, combined with lexical search. This is the retrieval foundation of every RAG pipeline and thus of every generative answer with sources.
Consequence: findability no longer hangs on exact keywords but on the semantic position of the content. In the embedding paradigm, “optimising” means anchoring your entity and your content in the right semantic region.
Consequences for content and entity work
- Synonyms strengthen the signal: paraphrases and alternative wording sharpen the semantic position — the opposite of the old keyword-consistency rule.
- Context counts: a brand's semantic neighbourhood is built through co-occurrence — which topics, people and organisations consistently appear in the same context.
- Keyword density is meaningless: density analyses measure nothing relevant in the embedding paradigm; semantic completeness and precision replace them.
- Entity consistency acts directly: uniform names, roles and attributions across sources keep the entity compact in the space — contradictions dilute the position.
Related terms
Embeddings are the foundation of semantic search, vector search and the retrieval stage in RAG. The strategic response to the embedding paradigm is topical maps and consistent entity work.
Visibility is a position in vector space
Retrieval systems find what sits semantically close to the query. Content and entity work moves that position — keyword repetition does not.
FAQ on embeddings
What is the difference between an embedding and a keyword? ▾
A keyword is a string matched exactly or in variants. An embedding is a mathematical representation of meaning: content is found via semantic proximity, even without shared words. Modern retrieval systems combine both approaches in hybrid setups.
Which distance measures are used in embedding space? ▾
Most commonly cosine similarity (the angle between vectors), alongside dot product for normalised vectors and Euclidean distance. What matters in practice is less the measure than the consequence: similarity is gradual and context-dependent, not binary.
How many dimensions do embeddings have? ▾
Common text-embedding models range from 768 to around 4,000 dimensions — roughly 768 for BERT base, 1,024 for BGE-M3, 1,536 to 3,072 for current OpenAI models. More dimensions does not automatically mean better quality for every use case.
What does the embedding space mean for SEO in concrete terms? ▾
Three things: synonyms and paraphrases strengthen the semantic signal instead of diluting it; your brand's neighbourhood is built through consistent co-mentions; and keyword density has become meaningless as an optimisation target. You optimise the semantic position, not word frequency.