Semantic search
Semantic search compares the meaning of a query with the meaning of content, not words with words. Both the query and the page are converted into a numeric representation of meaning, and the engine looks for closeness between them. That is why one page gets found through dozens of differently worded queries, and why writing content for one exact keyword stopped working.
In short
| What it is | Search by meaning rather than string matching |
| How it works | Query and content are converted into meaning vectors and compared for closeness |
| Who uses it | Google and Bing for years, AI assistants from the start. It is the foundation of RAG |
| What it means for content | Cover the topic and the intent, do not repeat a phrase. Synonyms and related concepts help, keyword density does not |
How semantic search works
Embeddings: text converted into a vector of meaning
The core is an embedding model. It turns text into a vector: hundreds to thousands of numbers describing its meaning. Texts with similar meaning get close vectors even when they share few words: "cheap mattress for a bad back" and "reasonably priced mattress when your spine hurts" sit close together, "mattress 90x200 in stock" sits farther away. Closeness is measured with cosine similarity: the higher the value, the closer the meaning. The engine converts its whole index this way in advance, per passage rather than per page. A long page covers several topics, and one vector per page would average them out.
What happens after a query is entered
- The query is processed: spelling correction, entity recognition (brand, model, place) and an intent estimate (information, a specific site or a purchase).
- The same model converts the query into a vector.
- The nearest passages are retrieved from the vector index. Classic word matching runs in parallel, because vectors capture product names, codes and model numbers less reliably. Both candidate lists are merged.
- A more precise model usually re-ranks the candidates, assessing query and passage together. Other ranking signals are added: links, source authority, freshness, and for product results price and availability from the feed.
- An AI assistant goes one step further: it passes the top passages to a language model as context; the model composes an answer and cites the source. That is RAG.
What this means for how a page is built
The comparison runs on passages, not on the whole page. So a specific paragraph decides the match: if it combines the question, the entity and the answer in one place, its vector sits close to the query. For an e-shop this means a category description that answers the questions people ask before buying, and a product description that states parameters, material and use in text, not only in an image or a context-free table. For a B2B site it means a service page that says explicitly for whom, what and how, not a generic text about quality and approach. The model connects synonyms, but it does not fill in missing information: if the page never says which problem the product solves, no embedding will match it to such a query.
Worked example: a mattress buying guide appears for 180 different queries in a month. Only 30 contain the phrase "choosing a mattress"; the other 150 are differently worded questions about firmness, back pain and materials. Out of 1,200 clicks, 360 come from the exact-phrase queries and 840 from the rest, so the page earns 70% of its traffic through meaning match, not phrase presence.
What it changes for content creation
When the engine compares meanings, what decides is whether the page answers the intent of the query, not how many times it contains a phrase. Practically that means three things: write for a topic and its questions, not for a single term; use natural language including synonyms, because the model can connect them; and structure the text so each section answers one question, because the comparison runs on passages.
From our own practice: same page, different query, different position
The clearest illustration of semantics in practice looks like this: one and the same page holds a top position for a long question-shaped query and sits deep for a short two-word keyword. The content is identical. The difference is meaning match: the page's structure mirrors how people phrase the question, not what a two-word keyword looks like.
Common mistakes
- Counting keyword density. A metric from the string-matching era. It says nothing today.
- Building a page for every phrase variant. Semantic search connects them anyway, and the pages end up competing with each other.
- Ignoring intent. A page can match the topic yet answer a different intent than the query carries. Then it fails to convert even at a good position.
Related terms
See also RAG, search intent, entity SEO, AEO and topical authority.
Frequently asked questions
Does this mean the end of keywords?
Not the end, a change of role. Keywords remain a map of demand and of the customer's language. They are just no longer a template the text gets written into.
How do I know a page works semantically?
In Search Console it appears for many differently worded queries on one topic. A narrow query profile means the page covers a keyword, not a topic.
Is this related to AI answers?
Directly. RAG runs on semantic passage retrieval, so content that works semantically also has better odds inside assistant answers.
How we can help
We build content around topics and intents as part of our AI visibility agency service.