Semantic search

Type "stop AI bots reading my site" into a search engine and you get pages about robots.txt rules for GPTBot, though your query mentioned neither. That is semantic search: the engine worked out what you meant and found pages that answer it, whatever words they used.

How does semantic search differ from keyword search?

• Matches on

  Keyword (lexical) search

  Words the query and page share

  Semantic search

  What the query means

• Finds synonyms

  Keyword (lexical) search

  Only if listed

  Semantic search

  Yes

• Handles typos and paraphrase

  Keyword (lexical) search

  Poorly

  Semantic search

  Well

• Exact names, codes, prices

  Keyword (lexical) search

  Very well

  Semantic search

  Often poorly

• Typical technique

  Keyword (lexical) search

  BM25 over an inverted index

  Semantic search

  Embeddings, entities, query rewriting

No serious engine uses only one. Google combines many systems, and AI assistants retrieve with both kinds before a reranker chooses. The practical difference for a writer is that a page no longer needs to contain the searcher's exact words to be found, but it does need to answer the question they meant.

How does it work under the hood?

Several techniques stack:

• Embeddings. A model turns queries and passages into vectors, so closeness means similarity of meaning. See vector search for a worked test with real scores.
• Entities. Words are tied to known things. "Jaguar" next to "dealership" is the car maker, and an engine that knows that does not need the page to say "car".
• Intent. The engine infers what the person wants to do. See search intent.
• Query rewriting. One question becomes several searches, each more specific than the original.

What does it look like in AI answers?

You can see it from the outside. Attensira tracks 61 questions about its own category, worded the way buyers word them, and records which sources decide each answer. In the 30 days to 4 October 2026, read over the Attensira MCP, Perplexity answered "does my website show up in Perplexity" and "how to get recommended by AI" from the same two sources, trysight.ai and linkedin.com, though the questions share no word. Both are about being visible in AI answers, and that is what retrieval matched on.

The reverse holds too: shared words do not guarantee shared sources. This is consistent with meaning-based retrieval. It is an observation of outputs, not a view into any engine's index.

Where does semantic search fail?

Where meaning is thin or exactness matters:

• Exact identifiers. Model numbers, error codes, version strings and SKUs. Two codes that differ by one character look almost identical to an embedding.
• Brand names that are common words. A product called "Linear" sits close to every passage about linear things.
• Vague questions. "How to get recommended by AI" carries little meaning, and scores against relevant and irrelevant pages come out nearly level.
• Negation and numbers. "Tools without a free trial" and "tools with a free trial" embed close together.

What does it change about how you write?

Less than the hype suggests, more than keyword habits allow:

• Answer the question, not the keyword. Put a direct answer under a heading that asks the question.
• Name things precisely. Use your product name, the category name and the exact technical term, because those are where semantic matching is weakest.
• Cover the follow-ups. A page that answers the next three questions matches more of the ways people ask.
• Drop keyword stuffing and "LSI keywords". Repeating variations does not add meaning, and nothing in Google's documentation describes latent semantic indexing as part of how it ranks.

How does Attensira help?

Attensira picks the questions your buyers ask AI, in their words, and reads every engine's answers daily. When your page is never cited for questions it should answer, the crew compares the page's passages with the questions and with the pages that are cited, and drafts a rewrite that answers them directly, for your approval. The free query fan-out generator shows the specific searches one broad question can turn into.