BERT is a language model that Google uses to work out what the words in a search and on a page mean from the words around them, rather than reading each word on its own. The name stands for Bidirectional Encoder Representations from Transformers, and Google began using it in Search in October 2019.
How BERT works
Older approaches to search treated a query a little like a bag of words: find pages that contain these terms, then rank them. BERT reads a sentence in both directions at once, so each word is understood in light of everything before and after it. The small words that older systems tended to skip, such as “to”, “for”, “without” and “not”, suddenly carry weight.
Take the search “can you get a mortgage without a deposit”. A system that ignores “without” might return pages about saving for a deposit, which is the opposite of what the person wants. A model that reads the whole phrase sees that the question is about zero-deposit or guarantor options. That is the shift BERT brought: better matching of a page to the search intent behind longer, conversational queries.
BERT was trained on a huge amount of text to predict missing words and the relationship between sentences. Google then applies that learned understanding to two jobs: interpreting the query and judging which passages of candidate pages answer it. It sits alongside other systems rather than replacing them. RankBrain came earlier and helps relate unfamiliar queries to known concepts; MUM came later and handles more complex, multi-part tasks. Together they are part of what people call semantic search.
Why it matters
For a UK business the practical effect is that Google is far better at spotting whether a page actually answers the question, and far less impressed by a page that simply repeats the phrase. A plumber’s page titled “emergency plumber Croydon” that never says what counts as an emergency, how quickly someone arrives or what a call-out costs is weaker than a page that answers those things in plain English.
It also rewards writing the way your customers speak. People type and say full questions: “do I need planning permission for a garden room”, “how long does a probate valuation take”. Pages that answer those questions directly, in clear sentences, are easier for a model like BERT to match. Pages written for a keyword tool rather than a person tend to fall behind.
Common mistakes
- Trying to “optimise for BERT”. There is no setting, tag or keyword density that targets it. Google has said as much; the only lever is clear content that answers the query.
- Stuffing a phrase into every heading. Repetition no longer helps a model that understands synonyms and context, and it makes the page worse to read.
- Writing one page to cover several unrelated questions, so no single passage answers any of them well.
- Dropping the small connecting words in headings to save space (“plumber cost London boiler”), which strips out the meaning a reader and a language model both rely on.
- Blaming a traffic drop on BERT years after the event, instead of checking for technical problems or a recent core update.
How to act on it
Start from the questions customers ask you on the phone and by email. Each important question deserves a direct answer near the top of the relevant page, written in full sentences, followed by the detail. Use the words your customers use, including the UK terms they actually search with, such as “quote”, “postcode”, “solicitor” or “MOT”.
Then check each page has one clear job. If a page tries to rank for both “how much does a loft conversion cost” and “loft conversion planning rules”, split or restructure it so each question has its own well-organised section or page. This is the core of the content SEO and strategy work I do: mapping each question to one page and writing that page so a person, and a language model, can tell at once what it answers.