Semantic search

How does semantic search work?#

Semantic search works by converting the query and every indexed passage into vector embeddings, then returning the passages whose vectors sit closest to the query’s — so results are ranked by how close their meaning is, not by how many words they share.

1. At index time, each page is split into passages and every passage is run through an embedding model. See vector embeddings and content chunking.
2. At query time, the question is embedded with the same model.
3. Nearest-neighbor search finds the passages whose vectors are closest, usually by cosine similarity, using indexes built to search billions of vectors quickly.
4. Reranking has a slower, more precise model re-score the short list against the query. See reranking.

The approach beat keyword search on its home turf early. In the 2020 Dense Passage Retrieval paper, an embedding-based retriever outperformed BM25 by 9 to 19 points in top-20 passage retrieval accuracy across open-domain question-answering benchmarks (Karpukhin et al.). A pre-trained version of that retriever powered the original retrieval-augmented generation paper the same year.

Semantic search vs keyword search: what's the difference?#

Keyword search finds pages containing the query’s words and ranks them by term statistics, while semantic search finds pages that mean the same thing even with different words — and modern engines run both together, because each catches what the other misses.

That’s why answer engines are hybrid. Perplexity runs lexical and embedding retrieval in parallel and merges the results (Perplexity Research). Google has layered meaning-based systems onto its ranking for a decade: RankBrain in 2015, neural matching in 2018, and BERT in 2019, which Google said would help it understand “one in 10 searches in the U.S. in English” (Google). Meaning-based matching widens the set of questions a page can be found for; exact terms still decide the precise lookups — product names, prices, error codes.

Do AI search engines use semantic search?#

AI search engines use semantic search at the retrieval step: Perplexity documents embedding retrieval alongside keyword matching, and Google tells site owners its AI systems “can understand synonyms and general meanings” — so pages compete on meaning as well as on exact terms.

• Perplexity: hybrid retrieval

  Official

  Lexical and embedding retrieval run together over its own 200-billion-URL index, then cross-encoders rerank the passages.

• Google AI Overviews and AI Mode

  Official

  Grounded in core Search ranking, which has used meaning-based systems since RankBrain. Google’s guidance: “You don’t need to write in a specific way just for generative AI search.”

• ChatGPT

  Observed

  Titles of pages ChatGPT cited were semantically closer to the prompt than uncited titles (0.602 vs 0.484) and closer still to its fan-out queries, in Ahrefs’ April 2026 study — meaning-level matching, measured from outside.

• Claude

  Observed

  Retrieves from Brave Search, whose top 10 held 79.2% of the URLs Claude cited in Profound’s 2026 study. How Brave matches meaning isn’t documented.

Semantic retrieval also explains why a page can be cited for a question it never phrases exactly. Engines split a prompt into sub-queries — see query fan-out — and each sub-query is matched by meaning against the passages in the index.

How do you optimize for semantic search?#

Optimize for semantic search by covering the questions behind a topic in plain language — one clear idea per section, headings phrased the way buyers ask, and the exact names buyers use — rather than repeating keyword variants.

• Start from questions, not keywords. Map what a buyer asks before, during and after choosing. See search intent and long-tail keywords.
• Give each question its own section, with a heading that reads like the question and a first sentence that answers it.
• Cover the topic, not the phrase. A topic cluster of pages answering related questions gives the engine more passages that mean what buyers ask.
• Keep the exact terms too. Product names, category words, prices and units still win the lexical half of hybrid retrieval.
• Skip synonym stuffing. Google says its AI systems understand synonyms, and the original GEO study found keyword stuffing did little or hurt visibility in generative answers (Aggarwal et al.).

Related terms#

• How LLMs answerVector embeddingsNumerical representations of text — lists of hundreds or thousands of numbers — arranged so that passages with similar meaning sit close together, which lets search systems match a query to content by meaning rather than by shared keywords.Read the entry
• How LLMs answerRerankingA second retrieval stage in which a more precise model re-scores the top results of a first, faster search against the query and reorders them — the step that decides which few passages an AI answer engine actually reads and cites.Read the entry
• Content & relevanceSearch intentThe goal behind a query — to learn something, reach a specific site, compare options or buy — and matching it decides which kind of page, from a guide to a product page to a comparison, search engines and AI answers treat as relevant.Read the entry
• How LLMs answerQuery fan-outAn AI search technique in which the engine rewrites one user question into several narrower sub-queries, runs them in parallel and builds its answer from the combined results — which is why a page can be cited for a prompt it doesn’t rank for.Read the entry
• RAGHow LLMs answerRetrieval-augmented generationA technique in which an AI system first retrieves relevant documents from an index and then gives them to a large language model to write its answer, so the response can cite current sources instead of relying only on what the model memorized in training.Read the entry
• Content & relevanceTopic clusterA group of interlinked pages on one subject — a broad pillar page linked to and from narrower pages on each subtopic — built to show search engines and AI retrieval systems complete coverage of the topic rather than a single article.Read the entry