Semantic search

Traditional search matched the words in your query against the words on a page. Semantic search adds a layer: it tries to read what you actually want. Google notes that people often type “keyword-ese,” unnatural strings they assume a machine prefers, so its systems build language models that map those words to intent.

How Google reads meaning

Three named systems do most of this work. RankBrain helps Google “understand how words are related to concepts.” BERT, built on transformers, reads each word in relation to the words around it, which is why it handles conversational phrasing. Neural matching works like a super-synonym system, matching “representations of concepts in queries and pages” to one another. In practice that means a search for change laptop brightness and a page that says “adjust laptop brightness” can line up, because the concepts match even when the words differ.

Word order matters too. Google’s own case was the query 2019 brazil traveler to usa need a visa. The word “to” sets the direction of travel. Older systems missed it and returned results about US citizens going to Brazil, while BERT read the whole phrase and understood a Brazilian going to the US.

For you, the shift is practical. Repeating an exact phrase does less than it once did. Covering a topic clearly, answering the real question, and writing in natural language is what actually gets matched.