Word embeddings represent words as dense vectors so that words with similar meanings sit close together. Google researchers popularised the approach with word2vec in 2013. Google has not said word2vec itself runs in Search ranking, but it confirms machine learning systems that match concepts rather than exact words, such as RankBrain and neural matching, and US v. Google evidence describes an embedding based ranking model called RankEmbed. For sites, pages are matched on meaning, so clear coverage of the concept beats repeating one phrasing.
What are word embeddings?
Word embeddings are numeric vectors that represent words so that words used in similar contexts get similar vectors. Distance in the vector space then stands in for similarity of meaning, which lets a system match "cheap flights" with "low cost airfares" without a shared word.
The best known method is word2vec, from Tomas Mikolov, Kai Chen, Greg Corrado and Jeffrey Dean at Google, first posted to arXiv on 16 January 2013. Google released it as an open source toolkit on 14 August 2013.
Typically, the representation is a real-valued vector that encodes the meaning of the word in such a way that the words that are closer in the vector space are expected to be similar in meaning.
Wikipedia, Word embedding, Wikipedia
How do word embeddings work?
word2vec trains a shallow neural network to predict a word from its neighbours, or the neighbours from a word, across a very large text corpus. The weights it learns become the vectors. Because the only training signal is co-occurrence, relationships such as country and capital emerge without labels.
Later models such as BERT produce contextual embeddings, where the vector for a word depends on the sentence around it, and dual encoder models embed whole queries and passages. Retrieval then becomes a nearest neighbour search between a query vector and document vectors, the approach usually called dense or embedding based retrieval.
We propose two novel model architectures for computing continuous vector representations of words from very large data sets.
Tomas Mikolov, Kai Chen, Greg Corrado, Jeffrey Dean, Efficient Estimation of Word Representations in Vector Space, arXiv
What evidence is there that Google uses word embeddings in ranking?
This page is labelled General theory because Google has not said that word2vec, or any named embedding method, is used in ranking. What Google does confirm is systems that match concepts rather than strings. It describes neural matching, introduced in 2018, as a way to understand representations of concepts in queries and pages, and RankBrain, launched in 2015, as its first deep learning system in Search (Pandu Nayak, Google, February 2022).
Court evidence goes further. Judge Mehta's September 2025 remedies opinion describes RankEmbed as an AI based deep learning ranking model and cites a Google exhibit stating that embedding based retrieval is effective at semantic matching of documents and queries. That is trial evidence for embedding based retrieval at Google, though it does not tell us which embedding method is used.
Neural matching helps us understand fuzzier representations of concepts in queries and pages, and match them to one another.
Pandu Nayak, Google, How AI powers great search results, Google blog
What do word embeddings mean for your pages?
Semantic matching narrows the gap between pages that use the searcher's exact words and pages that express the same idea differently. That reduces the payoff from exact match repetition and raises the payoff from explaining the concept clearly, in self contained passages, with the entities and attributes a reader expects.
It is testable. Embed your page's passages and a set of target queries with an open embedding model, compute cosine similarity, and compare with the pages that rank. A page that scores well below the ranking set on its own target queries usually has an intent or framing gap. The model will not be Google's, so treat it as a proxy that shows direction, not a score Google computes.
Word2vec uses distributed representations of text to capture similarities among concepts.
Google, Learning the meaning behind words, Google Open Source Blog
Which Laurelin audit checks test for word embeddings?
Google's embeddings cannot be observed, so these checks use proxies. The check low semantic similarity between target queries and page passages shared with the RankEmbed page, measures embedding similarity with an open model against the ranking set.
Related checks: Content not matching search intent is the usual cause of a low similarity score. Passages not self contained matters because passages are embedded separately. Question heading with no direct answer weakens passage to query matching. No contextual links between related pages and Weak topic cluster structure affect how a site's coverage of a concept reads as a whole.
What are the key dates for word embeddings?
- 2013-01-16: word2vec paper first submitted to arXiv by Mikolov, Chen, Corrado and Dean (source)
- 2013-08-14: Google releases word2vec as open source (source)
- 2015: Google launches RankBrain, its first deep learning system in Search (source)
- 2018: Google introduces neural matching to Search (source)
- 2025-09-02: Remedies opinion describes RankEmbed as a deep learning ranking model and cites embedding based retrieval (source)
Frequently asked questions about word embeddings
Does Google use word2vec?
Google created word2vec in 2013 but has not said it runs in Search ranking. It confirms concept matching systems such as RankBrain and neural matching, and trial evidence describes an embedding based model called RankEmbed.
What is the difference between word embeddings and LSI?
Both place related words close together. LSI uses matrix factorisation on a term and document matrix, while embeddings such as word2vec are learned by neural networks from very large corpora and scale far better.
Can I measure my page's embeddings?
You can measure them with an open model and compare against ranking pages. The result is a proxy for semantic fit, not the value Google computes.