Semantic Search

Google names three systems that work this way. Its guide to ranking systems describes neural matching as an AI system that understands representations of concepts in queries and pages and matches them to one another, RankBrain as one that helps it understand how words are related to concepts, so it can return relevant content even if it doesn’t contain all the exact words used in a search, and BERT as one that lets it understand how combinations of words express different meanings and intent^[1]. One common way to compare meaning is with embeddings.

When Google introduced BERT to Search in October 2019, it said the model would help Search better understand one in 10 searches in the U.S. in English, with more languages and locales to follow^[2]. Its example was a search for “2019 brazil traveler to usa need a visa”, where the word “to” decides who is travelling where. The figure is dated and narrow: one language and one country, at the time of launch. It is not a measure of how much of search works this way today.

Google says the same of its AI search features. Its guide to generative AI features says its AI systems can understand synonyms and general meanings of what someone is seeking, and connect them with content that might not use the same precise words, so a site doesn’t have to worry that it hasn’t captured every variation of how someone might seek its content. It adds that Google’s AI systems have improved their ability to understand the relevance of pages even when there is no exact match between the query and the page’s primary content^[3].

Words still count. Google says its ranking systems look at many factors and signals, and are designed to work on the page level^[1], and a page can only be matched if it can be read. In QueryRev’s study, OAI-SearchBot is named and blocked on 172 of the 6,294 sites whose robots.txt could be read, 2.7%, and PerplexityBot on 323, 5.1%^[4].

There is no semantic search score to read. The evidence is indirect: the queries a page earns impressions for in Search Console that its text does not contain. A page that is shown for a query it never repeats has been matched by meaning.

For a set of important queries, list the pages that contain the query’s words and the pages that answer it, and compare the two lists. The pages that answer without the words show where meaning does the work. The pages that repeat the words without answering show where the match is only on the surface.

Count the intents a page covers, not the phrasings it repeats. A page that serves the search intent of a query group fully is read as relevant to all of its phrasings.

Illustrative: an invented site, to show the method.

A cycling shop checks the query “how do I stop my bike chain slipping” against four pages. One page, titled “Bike chain slipping: how to fix it”, matches by words and by meaning. A second, “Why a drivetrain skips”, never uses the query’s words but explains the cause and the fix. A third, about a treadmill belt, uses the words and answers a different question. The fourth is about locks.

Search Console shows that the drivetrain page earns impressions for the query and its variations, without the page text containing them. The shop leaves its wording alone, and adds one plain sentence that names the problem as readers do, so the page also reads clearly to a person who searches the way the query does.