Embeddings & semantic search
Also known as: vector search, semantic similarity
Text turned into coordinates, so a machine can find passages that mean the same thing even when they share no words with the question.
Last revised 2026-09-07. Also available as markdown — request this URL with Accept: text/markdown.
In plain English
An embedding turns a piece of text into a long list of numbers that encodes its meaning. Two passages about the same idea land near each other, even if one says "cheap" and the other says "affordable". Retrieval measures that distance.
The consequence for writing is liberating and slightly counter-intuitive: repeating a keyword twenty times does not help, while explaining the idea clearly in natural language does. Synonyms are not a trick to be gamed; they are simply understood.
How to fix / set it up
- Write the way your buyers speak, including the words they use rather than only your internal terminology.
- State names, versions, prices and specifications exactly — that is where keyword matching earns its keep.
- Cover the concept properly on one page instead of scattering thin mentions across many.
The technical detail
Most production retrieval is hybrid — a keyword index for precision on names and numbers, vectors for meaning — with the results merged. That is why exact product names and specifications still matter: vectors are weak precisely where strings are strong.
Vector similarity is topical, not evaluative. It has no notion of whether your claim is true, only of what it is about, which is why quality signals are applied separately and why an authoritative-sounding but wrong page can still be retrieved.
FAQ
Is keyword research dead?
No, but its job changed. It is now a map of what people want to know, not a list of strings to place a target number of times.