Vector Search

Vector search retrieves items by comparing embeddings, lists of numbers that represent meaning, instead of matching exact words. A query is turned into a vector with the same Embedding Model used for the documents, and the system returns the documents whose vectors sit closest to it. That is how a search for “cheap flights to Rome” can surface a page about “budget airfare to Italy”.

How it works

• Embed: every document or chunk is converted into a vector, often with hundreds or thousands of dimensions.
• Store and index: vectors go into a Vector Database or an index inside a search engine such as Elasticsearch, Postgres with pgvector, or a cloud service such as Vertex AI Vector Search.
• Query: the query is embedded and compared with stored vectors using a measure such as Cosine Similarity.
• Approximate nearest neighbours: comparing against every vector is slow at scale, so indexes such as HNSW find very close matches quickly, trading a little accuracy for speed.

Most production systems are hybrid: they combine vector search with lexical retrieval such as BM25 and then re-rank the merged candidates. In MLforSEO’s five-stage retrieval funnel, keyword and vector similarity together decide which chunks enter the candidate pool at all.

How it’s used in SEO and marketing

• Understanding why AI search can retrieve pages that never use the exact query wording, and why semantically diffuse paragraphs match nothing well.
• Internal-link suggestions and “read next” widgets based on page embeddings.
• On-site semantic search and RAG assistants over your own content.
• Keyword clustering and mapping keywords to the closest existing page.
• Near-duplicate detection for content and images.

Example

In 10 Ways to Automate Internal Linking with AI, each post is embedded with sentence-transformers and stored in a vector database. For any page, the nearest neighbours become link candidates, and the same index can later power semantic site search. The approach finds related pages that share no tags or keywords.

Limitations and things to watch

• Vectors are weak on exact matches such as product codes, names and numbers, which is why lexical search is kept alongside.
• Results depend on the embedding model; changing models means re-embedding everything.
• Similarity is not quality or truth. Re-ranking and filtering are still needed.
• Hard to explain: you can see that two vectors are close, not easily why.

Learn it on MLforSEO

• How modern search actually works: the five-stage retrieval funnel
• The chunk, not the page: content an AI can actually extract
• 10 Ways to Automate Internal Linking with AI
• AI Search Optimisation & Agentic SEO