# Semantic Search: What It Is and What It Means for a Store

> What semantic search is, how Google reads a query by its meaning and not its words, and what changes on the product pages of an online store.

URL: https://clione.ai/glossary/semantic-search/
Lang: en

Semantic search is something the search engine does, not something you add to your site. That is why it should not be confused with [semantic SEO](/glossary/semantic-seo/), which is the work of writing and structuring content around topics and not isolated words, or with the semantic web of vocabularies and markup. It matters to a store for a concrete reason. Your customer types "running shoes for muddy trails when it rains" and your [product page](/glossary/product-page/) says "waterproof trail shoe with a membrane". **They share almost no words, and they can still find each other**, if the search engine understands they are about the same thing and if your page says what it needs to say.

## How it works

Google Cloud defines it as a technique that focuses on understanding the contextual meaning and intent of a query, instead of just matching keywords. To return more relevant results it takes into account **the relationships between words, the searcher's location, their previous searches and the context**. Its example: faced with "best laptops for graphic design students", a classic search engine looks for those words; a semantic one understands that you need a good graphics card, plenty of memory and a screen with accurate color.[^google-cloud-semantic]

At Google, this capability has been built piece by piece, and the company has told the story itself:[^google-ai-search]

| Year | System            | What it adds                                                                                                                                                    |
| ---- | ----------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 2015 | RankBrain         | Its first deep learning system in Search. It relates words to concepts and is used to rank the best results                                                     |
| 2018 | _Neural matching_ | Looks at the whole query or page, not just keywords, and matches fuzzier representations of concepts. It is a key piece for retrieving documents from the index |
| 2019 | BERT              | Understands how a combination of words expresses an idea, without dropping the small words. It plays a part in nearly every English-language query              |

Two of Google's examples explain it better than any definition. With RankBrain, a query like "what's the name for a consumer at the top of a food chain" ends up at "apex predator", even though nobody typed that phrase. With BERT, in "can you get medicine for someone pharmacy", the preposition stops being ignored and **the query shifts from how to fill a prescription to picking up medicine for someone else**.[^google-ai-search]

Google's documentation for site owners describes each in a line: RankBrain helps return relevant content **even if it does not contain all the exact words of the search**; _neural matching_ understands representations of concepts in queries and pages and matches them to each other; BERT understands how combinations of words express different meanings and intents.[^google-ranking-systems]

### Things, not strings

The other half of semantic search is the things in the world. In May 2012, Google introduced its [Knowledge Graph](/glossary/knowledge-graph/) with a motto, "things, not strings": a model of real entities and their relationships that lets it tell, for example, the Taj Mahal monument from the Taj Mahal musician. At the time it held more than 500 million objects and more than 3.5 billion facts about them and their relationships.[^google-knowledge-graph] A query is no longer just a series of words: it can refer to a specific [entity](/glossary/entity/).

### What changes for content

Google says it in its guide to generative AI features: you do not need to write in any special way, because its systems understand synonyms and the broad meaning of what someone is looking for, and connect that person with content that does not use exactly the same words. So, it says, **you do not need to worry about not having enough long-tail keywords** or about having covered every way someone might search for what you sell.[^google-ai-guide] What still depends on you is that the fact exists on the page.

## In a store

A **fictional example** with a shoe store. The queries are a shopper's; the text is from the [product page](/glossary/product-page/).

| Shopper's query                              | What the page says                    | Literal match? | Can semantic search find it?                                       |
| -------------------------------------------- | ------------------------------------- | -------------- | ------------------------------------------------------------------ |
| running shoes for muddy trails when it rains | Waterproof trail shoe with a membrane | No             | Yes: muddy trails and trail, rain and waterproof are the same idea |
| sneakers for wide feet                       | Roomy last, available in 2E width     | No             | Yes, if the page states the last in text                           |
| lightweight shoes for a marathon             | Racing shoe                           | No             | Not reliably: the page does not state the weight                   |

The third row is the lesson. Google Cloud gives a similar commerce example: a semantic search engine understands that "running shoes" relates to "sneakers", "athletic footwear" or "trainers".[^google-cloud-semantic] What no system can do is **deduce a weight the page does not give**. Semantic search fixes vocabulary, not a lack of [product attributes](/glossary/product-attributes/).


That semantic search forces you to stuff the product page with synonyms or to create a page for every way of saying the same thing. Google says the opposite: its systems understand synonyms and the broad meaning of the query, and you do not need to cover every variant.[^google-ai-guide] The real risk is the reverse: **a page that does not mention the fact**, because there is no meaning there to understand.


## How to check it

1. **Write five questions the way your customer would ask them**, without using the words on your product pages, and search for them on Google. See whether your page comes up and which pages rank above it.
2. **Open the performance report in [Search Console](/glossary/google-search-console/)** and look for queries that bring impressions to a product page without containing its words: that is semantic search working in your favor.
3. **Test your internal search with synonyms** ("sneakers", "trainers", "running shoes") and note which ones return nothing.
4. **Go over a product page looking for implicit facts**: if the weight, the last, the material or the use are not written down, no search engine will find them.
5. **Do not create pages for variants of the same query.** Google does not ask for it, because it understands the broad meaning of what is being searched.[^google-ai-guide]

## Related concept: vector search

[Vector search](/glossary/vector-search/) turns text into mathematical vectors, [_embeddings_](/glossary/embeddings/), and calculates the distance between the query's vector and each document's to find what is most similar. Google Cloud draws the line: semantic search can use vectors, but **it is a broader concept**, covering different techniques for understanding the meaning of words and their relationships.[^google-cloud-semantic] Put another way, semantics is the goal and vectors are one of the ways to get there. And [AI search](/glossary/ai-search/) goes a step further: besides finding by meaning, it writes an answer and cites its sources.

To go further, the [AI search reading path](/learn/ai-search/) orders the rest of the terms.