How BERT works
BERT stands for Bidirectional Encoder Representations from Transformers, a natural language model Google published and then applied to search in 2019. The word that matters in that name is bidirectional. Earlier language processing read a sentence in one direction, word after word. BERT reads each word in the light of everything around it, on both sides.
That is what lets small words carry weight. “Visa for Nepal” and “visa from Nepal” describe opposite journeys, and the difference lives entirely in a preposition that older systems would have treated as noise. BERT keeps those words and uses them to decide which of two very similar pages actually answers the search.
It was applied to ranking and to featured snippets, in English first and then more widely. Like the query systems before it, it works on the search rather than on your page, and Google has said there is nothing to optimise for. It is part of the machinery that establishes what you were asking.
Why BERT matters
It rewards writing that says exactly what it means. A page answering “how long does a tourist visa last” is no longer interchangeable with one answering “how long does it take to get a tourist visa”, even though they share most of their words. Precision in headings and opening sentences translates directly into matching the right search.
It also helps pages written in plain, complete sentences rather than in keyword shapes. Copy that has had its small words squeezed out to fit a phrase in gives the model less to work with, not more.
Common mistakes with BERT
The first is believing there is something to buy. Services offering BERT optimisation are selling clear writing with a technical label attached. No markup, setting or ratio is involved.
The second is confusing it with topic coverage. Sprinkling synonyms and related terms across a page does nothing for a model that is reading grammar, and it usually reads badly to a person as well. What helps is answering one clear question per section, in the words somebody would actually use.
The third is heading fragments. Turning “How to register a company in Nepal” into “Company registration Nepal” removes the very words that identify which question the section answers.
What to do about it
Write headings as the questions they answer, and answer them in the first sentence underneath. That single habit does more for matching than any amount of keyword placement, because it puts a full, grammatical statement of the topic in the most prominent part of the page.
Then look for the near misses. In the Search Console query report, find searches where you collect impressions but few clicks: often the page is close to the subject but answers a slightly different question, and separating the two — or adding the missing answer — fixes it. That work sits inside semantic search generally, and it is the everyday form of content SEO.