Skip to content
Browse courses

Core Concepts (AI/ML)

Embedding

In shortA list of numbers (a vector) that represents the meaning of a word, text, image or entity, so similar things sit close together.

Updated 24 Oct 2026· 2 min read

An embedding is a numerical representation of meaning: a list of numbers (a vector) produced by a model so that items with similar meaning end up close together in that number space. Words, sentences, whole pages, images and entities can all be embedded. MLforSEO also calls this a semantic feature vector, and it is the foundation of most modern semantic SEO techniques.

How it works

  • An embedding model, today usually a transformer such as Sentence-BERT, reads the input and outputs a fixed-length vector (for example 384 or 768 numbers).
  • The model has learned from huge amounts of text that words used in similar contexts mean similar things, so “cheap flights” and “low-cost airfare” get similar vectors despite sharing no words.
  • Closeness is measured with cosine similarity, which gives a similarity score.
  • Older methods such as Word2Vec embed single words; modern models embed whole passages and take context into account. This is a step up from a bag of words, which only counts terms.

How it’s used in SEO and marketing

  • Keyword-to-topic mapping and keyword clustering with sBERT and BERTopic.
  • Internal linking: embed every page or paragraph and suggest links between the most similar ones.
  • Related posts, redirect mapping and duplicate or cannibalisation detection.
  • Image classification and near-duplicate detection with image embeddings such as CLIP.
  • AI search: retrieval systems embed queries and content chunks and use vector similarity as an early filter, which is why MLforSEO teaches writing self-contained chunks.

Example

For internal linking, MLforSEO’s guide embeds every paragraph on a site with sentence-transformers, then for each target page finds the paragraphs elsewhere with the highest cosine similarity. A paragraph about “redirect chains during a migration” scores high against a migration guide even if it never uses the guide’s exact title, so it becomes a link suggestion. In the information gain patent, a document’s semantic feature vector is likewise the input a model uses to estimate how much new information it adds.

Limitations and things to watch

  • Embeddings from different models are not comparable; re-embed everything when you switch.
  • Similar is not the same as relevant or correct: antonyms and competing products can sit close together.
  • Quality depends on the model’s training data, language coverage and maximum input length; long pages may need chunking.
  • Vectors are hard to interpret directly; use dimensionality reduction and real examples to inspect them.

Learn it on MLforSEO

Where Embedding shows up on MLforSEO

Learn it in a course

Read about it on the blog

Try it yourself

Sources and further reading

Other terms from Core Concepts (AI/ML)

See all 19 Core Concepts (AI/ML) terms →