Semantic Search

What changed when search stopped needing your exact phrase?

The unit of optimisation moved from the string to the subject. Under exact matching, a page had to contain the words a searcher used, which is where keyword density came from. Under semantic retrieval a page is represented by what it means, so covering the subject completely does the work that repetition used to, and repetition itself does nothing.

Every controlled test of the older tactic has pointed the same way since: keyword stuffing measured negative in the original generative-engine-optimisation paper and in the replications that followed it. That is unusual agreement in a field where most things fail to replicate.

How is semantic search different from vector search?

Semantic search is the goal; vector search is the usual implementation. You can build semantic retrieval other ways, with entity graphs or query rewriting, and modern systems combine several. Treating the two words as synonyms is harmless in conversation and misleading in a procurement discussion, where “semantic” often describes an ambition and “vector” describes a component somebody has actually built. Google’s own systems are the clearest illustration of the difference.

Query fan-out rewrites one question into many before anything is retrieved, which is semantic work done by a language model rather than by a distance calculation, and the vector step comes afterwards. Calling the whole pipeline vector search would miss the half that decides what is being searched for. The practical upshot for a writer is that two different systems are reading you. One is deciding what the question means, and it responds to how clearly your headings state a question. The other is deciding which passage matches, and it responds to how completely the passage answers one. Both are worth writing for and only one of them was ever visible in a keyword tool.

Does semantic search mean keywords no longer matter?

No, and the overcorrection is its own mistake. Keywords still tell you what people ask and in what words, which is the input to deciding what to write about and how to phrase a heading. What has ended is the idea that the words are a lever on ranking rather than a description of demand.

The recommendation, which is a position: use keyword research to choose the questions and to phrase the headings the way people ask them, then write the answer in whatever words make it clearest. A heading matching the question is worth more than a body matching the keyword.

Related terms

• Vector SearchVector search finds documents by comparing embeddings rather than by matching words: each passage is stored as a list of numbers, and a query is answered by finding the nearest ones.
• EmbeddingAn embedding is a list of numbers representing a piece of text, positioned so that texts with similar meaning sit close together.
• Query Fan-OutQuery fan-out is Google’s own term for how AI Overviews and AI Mode answer a question: the engine breaks the question into subtopics and issues many searches at once, then writes one answer from the results.
• Search IntentSearch intent is what a person actually wants when they type a query, as distinct from the words they used.