What is semantic search? Definition and examples
On this page
Semantic search finds results by what a query means rather than the exact words it contains, so different wording can still lead to the right page.
How it works
Classic keyword search matches words: a page ranks if it contains the terms in the query, weighted by how rare and frequent they are. Semantic search uses language models and vector search to find results "based on meaning, not just keywords", as Elastic's documentation puts it (Elastic).
The steps:
- Every passage of content is turned into a vector embeddingA vector embedding is one fixed-length list of numbers that a model produces for a piece of content, placing it near other content with a similar meaning. by an embedding model and stored, usually in a vector databaseA vector database stores embeddings and quickly finds the ones closest to a query, which lets AI systems search content by meaning instead of exact words..
- The query is embedded with the same model.
- The search returns the passages whose vectors sit closest to the query's vector. In Google's words, the distance between two items in embedding space "can be interpreted as a measure of relative similarity" (Google).
Semantic search has a known blind spot: exact identifiers. A model may place "error E-4012" near other error codes rather than on the one page that mentions it. Keyword ranking functions such as BM25 catch those exact matches, which is why many systems run both and merge the results, often called hybrid search (Anthropic). A rerankingReranking is a second search step in which a model reorders a shortlist of retrieved passages by how well each one actually answers the user's question. step can then sort the merged list.
Why it matters for your website
Visitors do not use your vocabulary. They type "cancel", "stop paying" or "end contract" when your help article is titled "Managing your subscription". A keyword search box returns nothing, and a visitor who gets zero results often leaves or opens a support ticket.
Semantic search narrows that gap, and it is the same technology that lets an AI assistant find the right passage before answering. It does not fix missing content, though: if no page explains your refund policy, the closest match will be the wrong one. Search logs, especially queries that return nothing useful, are a reliable list of pages to write. Our article on AI site search covers how to put this on a site.
Example
| Visitor's query | Keyword search returns | Semantic search returns |
|---|---|---|
| "can I pay by bank transfer" | Nothing (the page says "wire") | "Payment methods: we accept wire transfers on annual plans" |
| "stop my subscription" | Nothing | "Managing your subscription: how to cancel" |
| "E-4012" | The troubleshooting page for E-4012 | Several pages about errors, ranked loosely |
The first two rows show where meaning wins; the third shows why hybrid search keeps a keyword index alongside.
Related terms
- Vector embedding: the numbers that represent meaning.
- Embeddings: the wider idea behind them.
- Retrieval: the step that uses semantic search before an AI answers.
- Vector database: where embeddings are stored and searched.
- Reranking: reordering results for relevance.
- Search intent: what the searcher is trying to do.
Sources
- Semantic search (Elastic documentation) — accessed October 21, 2026
- Introducing Contextual Retrieval (Anthropic, September 2024) — accessed October 21, 2026
- Embeddings: Embedding space and static embeddings (Google Machine Learning Crash Course) — accessed October 21, 2026