GEO glossary

Semantic search

Semantic search is search that matches the meaning of a query rather than its exact words. It usually works by turning queries and documents into embeddings, lists of numbers that place similar meanings close together, so a page can be found for a question it never phrases the same way.

Reviewed by Gareth Hoyle, Managing Director
Last reviewed 31 October 2026

How does semantic search work?

Keyword search looks for documents that contain the query's terms. Semantic search converts text into embeddings, which Google Cloud's documentation describes as high-dimensional numerical vectors that represent a piece of text. The closer two vectors are, the more similar their meaning, so cat, dog and lion sit near each other. The system returns the passages whose vectors are nearest to the query's.

Many systems run both methods together, which Google Cloud calls hybrid search: a semantic match for meaning and a lexical match for exact terms such as names and product codes.

How Google moved to meaning

RankBrain, launched in 2015, was Google's first deep learning system in Search and relates words to real-world concepts. Neural matching, added in 2018, matches fuzzier representations of concepts in queries and pages. BERT, launched in October 2019, affected one in 10 English searches in the US, and taught Search that the word 'to' in '2019 brazil traveler to usa need a visa' changes who is travelling where.

Why semantic search matters for content

Vector search is also how many retrieval-augmented generation systems choose the passages an AI assistant reads before answering. That changes what makes a page findable.

Covering a question fully counts for more than repeating its keyword
Each passage should make sense on its own, because passages are often scored individually
Use the words customers use, including plain descriptions of their problem
Keep exact names, model numbers and prices in the text, since hybrid systems still match them

Example

A page titled 'Fixing a dripping mixer tap' can be retrieved for 'why does my kitchen tap keep leaking after I turn it off', even though the two share few words, because both describe the same problem. A page that only names the product it sells, without explaining the cause, gives a semantic system less to match.