What is semantic search?

Semantic search is a retrieval method that matches the meaning of a query with the meaning of documents instead of their exact keywords. Google and Bing use it, and so do the AI answers built on their indexes, such as Google AI Mode and Microsoft Copilot, which can retrieve a page that answers a question in other words.

Why semantic search matters

Semantic search is why an AI engine can cite a page that shares almost no words with the prompt. Search engines moved this way years before AI answers. When Google applied BERT to ranking in 2019, it said the model would help Search better understand one in 10 searches in the U.S. in English, especially conversational queries where small words like "for" and "to" change the meaning. The same year, Microsoft said Bing had turned over 150 billion pieces of data into vectors to improve on keyword matching.

Google now tells site owners that, for its generative AI features, they need not chase every keyword variation, because its systems "can understand synonyms and general meanings of what someone is seeking."

Embeddings

Embeddings are what make semantic search work. In OpenAI's words, an embedding is "a vector (list) of floating point numbers," and "the distance between two vectors measures their relatedness." Texts with similar meanings get vectors that sit close together, even when they use different words. OpenAI's text-embedding-3-small model, for instance, turns a text into 1,536 numbers by default. Google describes the same idea in its ranking systems: neural matching helps it "understand representations of concepts in queries and pages and match them to one another."

Vector search

Vector search is the retrieval step: finding the stored vectors closest to the vector of the query. Microsoft's example from Bing is "How tall is the tower in Paris?" Bing can answer that the Eiffel Tower is 1,063 feet "even though the word 'Eiffel' never appeared in the search query and the word 'tall' never appears in the result." It is also how retrieval-augmented generation systems pick the passages they feed a model. At web scale this has to be fast: a Bing program manager described searching "100 billion-plus vectors to find the most related results in 5 milliseconds."

Example: matching meaning, not words

Take the prompt "software to chase unpaid invoices automatically". A page titled "Accounts receivable automation: scheduled payment reminders" shares almost none of its words, yet its embedding sits close to the prompt's, so semantic retrieval can return it. A keyword-only engine would favor pages that repeat "chase unpaid invoices". For the vendor, the lesson is that the page did not need the buyer's exact phrase. It needed to describe the same job clearly.

How to write for semantic search

Cover the topic, not the keyword list. Answer the questions behind the query in plain words; repeating variants adds nothing.
Name things explicitly: the product category, the use case, who it is for. Clear meaning is what gets matched, the way search engines understand a brand as an entity: a concept rather than a string of words.
Check what actually gets retrieved. Hawkscan lists the pages cited by ChatGPT, Gemini, Microsoft Copilot and Google AI Mode for your prompts, with the eight most cited domains, so you can see which pages the engines actually use for your topics.