Embeddings

What are Embeddings?

Embeddings are numeric representations of meaning. A machine learning model converts a word, sentence, passage or image into a vector, a list of numbers, so that items with similar meaning sit close together in that space. Systems compare vectors to find content semantically related to a query, even when the two share no exact keywords.

Google’s machine learning course describes embeddings as lower-dimensional representations of sparse data, which fixes the lack of meaningful relations between items in one-hot encodings. OpenAI’s API docs describe embeddings as measuring the relatedness of text strings, used for search, clustering, recommendations and classification. The idea goes back to word vectors such as word2vec; modern embedding models represent whole passages.

How it works

1. Encode content. An embedding model turns each passage (often a chunk of a page) into a vector, typically hundreds or thousands of numbers long.
2. Store vectors. Vectors go into an index or vector database.
3. Encode the query. The user’s question is embedded with the same model.
4. Compare. The system finds vectors nearest to the query vector, usually by cosine similarity or a related distance measure.
5. Use results. Nearest passages are ranked, shown, or passed to a language model as context for an answer.

An illustrative example: a page section titled “How to detect lateral movement” and a query “spotting attackers moving between hosts” share few words, but a good embedding model places them close together.

From SEO to AEO and GEO

Classic SEO grew up on lexical matching: keywords in titles, headings and anchor text. Google moved beyond that years ago with RankBrain, neural matching and BERT, which interpret meaning rather than strings. AI answer engines push further. Retrieval for an AI answer often happens at the passage level, and semantic similarity is a common part of how passages are chosen.

Classic SEO habit	AEO/GEO reality
Repeat the exact keyword	Make the passage’s meaning unambiguous
One page targets many variants	One section answers one question well
Page-level relevance	Passage-level relevance

What carries over: matching search intent, topical depth and clear structure. What fades: keyword density and exact-match tactics, which do little for meaning-based retrieval.

Embeddings also explain why AI answers sometimes cite a page that never mentions the query’s exact phrasing. The page was close enough in meaning to be retrieved, and its passage happened to answer the question clearly.

What to do about it

1. One idea per section. Mixed sections produce blended vectors that match no query strongly.
2. Answer first. Put the direct answer in the first sentence under each heading.
3. Use buyer vocabulary alongside your product terms, especially in technical categories where vendors invent names.
4. Name entities explicitly. “Our platform” carries less meaning than the product and category name.
5. Test semantically. If you have engineering help, embed your key pages and buyer questions with an open model and check which sections are nearest. Treat it as a proxy, not a replica of any engine.

Common mistakes

• Assuming embeddings replaced all ranking signals. They are one component; authority, freshness and other signals still matter.
• Synonym stuffing. Lists of variants dilute meaning.
• Ignoring chunk boundaries. A key fact split across two sections may be retrieved without its context.

Related terms

• Vector Search
• Semantic Search
• Chunking
• Retrieval-Augmented Generation
• Neural Matching
• How AI engines cite sources

Sources

• Embeddings, Google Machine Learning Crash Course
• Vector embeddings guide, OpenAI developer docs
• Understanding searches better than ever before (BERT), Google