BERT, short for Bidirectional Encoder Representations from Transformers, is a natural language processing model developed at Google and applied to Search from 2019. It reads words in relation to all the other words around them, not one at a time, which helps it understand prepositions, negations and word order. Google uses it both for ranking and for selecting featured snippets, and lists it as an active ranking system.
Why small words matter
Older systems often treated words like “to”, “for” and “without” as noise. BERT made them meaningful. Google's launch examples showed queries where a single preposition reversed who was doing what, such as whether a traveler was going to or from a country. Getting those right changes which pages are relevant.
Transformers in search
BERT is built on the transformer architecture, the same family of models behind modern large language models. It was one of the first times Google described using transformers in production search. Later systems, such as MUM, build on similar foundations.
What it means for content
You cannot optimize for BERT directly, and Google has said as much. What it does is reward plain, precise writing. If a page answers “how to migrate from spreadsheets to a CRM”, it should be about that direction of migration, with the steps in the right order.
- Write complete sentences rather than keyword fragments
- Be precise about direction, scope and audience
- Answer the specific question, not a nearby one
Example
Matching direction in a migration guide
Hypothetical CRM Dealbarrow has one guide for importing data from a spreadsheet and another for exporting data back out. Both used to rank for both directions. After clarifying titles, intros and headings so each page states its direction plainly, each guide matches the queries it was written for.
Common mistakes
Treating BERT as a ranking factor to optimize
It is a language understanding system. Clear writing helps; tactics labeled “BERT optimization” usually do not.
Questions
No. BERT is an encoder model that understands text; it does not generate long answers. Chat assistants use generative models built on related transformer research.
