Semantic search
又称: vector search, meaning-based search, embedding search
一句话解释
Semantic search matches queries to content by meaning rather than by matching words. Text is converted into vectors — numerical representations where similar meanings sit close together — so a query can retrieve a passage that answers it without sharing any of its vocabulary.
What changed for content
Semantic retrieval ended keyword density as a strategy and replaced it with something harder to fake: the passage has to actually mean what the query is asking about.
Practically, three habits stopped working and one started mattering more.
Stopped working: repeating the target phrase, listing synonym variations, and writing separate near-duplicate pages for near-duplicate queries — those now compete with each other for the same vector neighbourhood.
Started mattering: clarity. A passage that states one idea plainly produces a cleaner embedding than one hedging across three.
Chunking is the practical consequence
Semantic search does not embed whole pages. It splits them into chunks and embeds each. A page is therefore not one point in vector space but many, and each chunk competes independently.
This is why a self-contained section beats a well-argued page: the chunk is what gets retrieved, and a chunk that only makes sense in context embeds as something muddier than what it actually says. See passage retrieval.
Where it sits in an AI answer
Semantic search is usually the retrieval step inside retrieval-augmented generation. It decides which passages the model gets to see. Everything downstream — whether you are cited, whether you are named — depends on clearing that step first.
常见问题
- How is semantic search different from keyword search?
- Keyword search matches strings; semantic search matches meaning. A page about 'reducing customer churn' can be retrieved for 'how do I stop people cancelling' without containing any of those words, because the two are close in vector space.
- Does semantic search make keywords useless?
- It makes exact-match optimization useless while leaving keywords useful as a description of demand. You still need to know what people ask; you no longer need to repeat their phrasing on the page.
- What is an embedding?
- A list of numbers representing a piece of text's meaning, produced by a model. Texts with similar meanings produce similar lists, which lets a system find related content by mathematical proximity rather than by matching words.
相关术语
- Retrieval-augmented generation (RAG)
Retrieval-augmented generation is a technique where a system retrieves relevant documents and places them in a language model's context, so the model answers from those sources rather than from memory alone. It is how AI answers come to have citations, and how live information reaches a frozen model.
- Passage retrieval
Passage retrieval is the practice of splitting documents into sections and retrieving and ranking those sections independently, rather than treating a page as one unit. It is why a long page can be pulled into an answer on the strength of one paragraph, and ignored despite being strong overall.
- Conversational search
Conversational search is querying a search system in natural language — full questions rather than keyword fragments — often across several turns where each question builds on the last. It is the input style answer engines are designed for, and it makes the question, not the keyword, the unit of demand.
- Topical authority
Topical authority is the degree to which a site covers a subject comprehensively and credibly, rather than touching it once. It is built by covering a topic's real question space in depth, with internal connections between the pieces, and it makes a site a more reliable source to retrieve from.
相关指南
- Generative Engine Optimization (GEO): the complete guide
GEO is the practice of getting your brand named and cited inside AI answers. What it is, how it differs from SEO, and the levers that measurably work.