An embedding is a numerical representation of a piece of text, an image or another input, produced by an embedding model. Texts with similar meanings end up with similar numbers, so the distance between two embeddings works as a measure of how related they are. Embeddings are the basis of vector search and semantic search, and they are one of the main ways AI search systems decide which passages are relevant to a question.
How embeddings work
An embedding model reads a passage and outputs a vector: a long list of numbers, often hundreds or thousands of them. Each number on its own means little to a human. What matters is the geometry: passages about the same topic sit close together, and unrelated passages sit far apart. A question can be embedded the same way, then compared against stored passage embeddings to find the closest matches.
Similarity is usually measured with cosine similarity or a related distance calculation.
similarity(a, b) = (a · b) / (|a| × |b|)
// 1.0 = same direction (very similar)
// near 0 = unrelatedWhy marketers should care
Embedding-based retrieval rewards passages that are clearly about one thing. A paragraph that mixes pricing, a customer quote and a feature announcement produces a muddy embedding that sits between several topics. A paragraph that fully answers 'how does Quillbrook handle multi-currency invoices' sits close to questions on that subject.
Embeddings also capture synonyms and paraphrases, so exact-match keyword repetition matters less in this kind of retrieval than it did in older keyword systems. Coverage of the concept, in clear language, matters more.
Where you will meet them
Beyond AI search, SEO teams use embeddings directly: clustering keywords by meaning, finding near-duplicate pages, mapping internal link opportunities and comparing a page against the passages that currently rank. Most embedding APIs are inexpensive enough for this kind of analysis on a few thousand pages.
Example
Finding cannibalizing articles
Ledgerfen, a fictional accounting app, embeds the opening section of all 600 of its blog posts and groups those with high similarity. Several clusters turn out to be three or four posts answering the same question in slightly different words. The team merges each cluster into one stronger page and redirects the rest.
Common mistakes
Treating similarity scores as universal
Scores depend on the embedding model. A threshold that works for one model can mean something different in another.
Embedding whole pages
Long, multi-topic pages blur into an average. Embed sections or passages for analysis that reflects how retrieval actually works.
Questions
Google has publicly described using neural matching and related language understanding systems, and embedding-based retrieval is standard in the industry. The specific models and how they are combined with other signals are not published.
No. Embedding models are trained to produce vectors for comparison, while generative models produce text. Some systems use both: embeddings to find passages and an LLM to write the answer.
Sources
