BERT Algorithm

BERT (Bidirectional Encoder Representations from Transformers) is a language model developed by Google that reads each word in the context of the words before and after it. Google released it as open source in 2018 and began using it in Search in October 2019 to better understand the meaning of queries and passages.

BERT is built on the transformer architecture introduced by Google researchers in 2017. Unlike earlier models that processed text left to right, BERT looks at a whole sentence at once, so small words that change meaning are not lost. In its October 2019 announcement, Google said BERT would help it understand about one in ten English searches in the US, with a gain on longer, conversational queries where prepositions like 'for' and 'to' matter. Its best-known example was '2019 brazil traveler to usa need a visa': before BERT, Google returned results about US citizens traveling to Brazil, and with BERT it understood that the traveler was Brazilian. Google applied BERT to featured snippets in about two dozen countries at the same time and said it would bring BERT to more languages over time.

BERT is still in use. Google's guide to its ranking systems lists BERT as an active system that helps it 'understand how combinations of words express different meanings and intent', next to RankBrain, neural matching and passage ranking. The same guide notes that MUM, a later Google model, is not used for general ranking. Generative features such as AI Overviews and AI Mode run on Gemini models, which sit on top of this ranking stack rather than replacing it.

A common misconception is that GPT, Claude and Gemini descend from BERT. They share the transformer as a common ancestor, but BERT is an encoder-only model built to understand text, while modern chat models are decoder-based models built to generate it. BERT's direct descendants are a different family: encoder models that turn text into embeddings and cross-encoders that score how well a passage answers a query. Those models are widely used for semantic search and for reranking retrieved passages in retrieval-augmented generation, which is how many AI search engines pick the passages they cite.

For content teams, BERT's lesson has held up for seven years. Search systems match meaning, not keyword strings, so a page should answer the full question in natural language, with the qualifiers that change the answer (who, where, for whom, compared with what) stated plainly. Keyword stuffing does not help a model that reads context, and vague passages are less likely to be retrieved or reranked into an answer. Writing one clear passage per sub-question serves BERT-based ranking in Google and embedding-based retrieval in AI search at the same time.