What Is Vector Search?

Practical example

A procurement manager asks Perplexity how to cut their company's energy bill. A facilities vendor's guide titled "Improving energy efficiency in commercial buildings" never uses the words "cut" or "bill", so a keyword index might miss it. A vector index places both texts close together because they express the same intent, so the guide is retrieved and can be considered as a source for the answer.

Why is vector search important for AEO?

Most AI answer engines use retrieval-augmented generation: they retrieve candidate passages first and generate an answer from them. Where that retrieval step is vector-based, a page can be found for questions that share none of its wording, and it can be passed over if its meaning is unclear or buried in promotional text. This shifts optimization from exact-match keywords toward clearly expressed topics, self-contained passages and consistent terminology.

Vector search only determines what is retrieved. Which sources are cited depends on further ranking, source quality and each engine's own selection logic, and those pipelines are not fully public.

How does vector search work?

Vector search replaces word matching with distance measurement. The process has five stages:

1. Chunking and embedding. Content is split into passages, and an embedding model converts each passage into a vector of hundreds or thousands of numbers. Passages with similar meaning receive vectors that point in similar directions.
2. Indexing. The vectors are stored in a vector index or vector database. Comparing a query against every stored vector is too slow at scale, so systems use approximate nearest neighbor (ANN) algorithms such as HNSW, which trade a small loss of accuracy for large speed gains.
3. Query embedding. The user's question is converted into a vector with the same embedding model.
4. Similarity scoring. The index measures the distance between the query vector and the stored vectors, typically with cosine similarity, and returns the closest matches.
5. Hybrid retrieval and reranking. In production systems, vector results are usually combined with keyword (lexical) results, because vector search can miss exact names, product codes and rare terms. A reranking model then orders the candidates before they are passed to the language model that writes the answer.

Because retrieval in most of these systems happens at passage level, a page is not judged as a whole. Each section competes on its own, which is why self-contained passages matter in AEO.

Related concepts

• Embeddings: the numerical representations themselves. Vector search is the operation performed on them; embeddings are what makes it possible.
• Semantic search: the goal of matching meaning rather than words. Vector search is the most common way to achieve it, but query expansion, knowledge graphs and learned sparse retrieval also contribute, so the two terms are related rather than identical.
• Retrieval-augmented generation (RAG): the pipeline in which vector search typically sits. RAG retrieves relevant passages, then prompts a language model to answer using them.
• Keyword search and hybrid search: keyword (lexical) search, often based on BM25, scores pages by term overlap. Hybrid search runs both methods and merges the results, and is the norm in production retrieval systems.