Quick answer: On 25 October 2019 Google announced it is using BERT, a language model, to better understand search queries. Google said it affects about one in ten English searches in the US, mostly longer, conversational queries where small words like “to”, “for” and “no” change the meaning. It is also being used for featured snippets in a couple of dozen countries. There is no BERT score to chase and nothing technical to fix. The practical takeaway is to write pages that answer specific questions clearly, in natural language, for the person actually asking.
What BERT is, in plain terms
BERT stands for Bidirectional Encoder Representations from Transformers. Google researchers published it and open sourced the approach in late 2018. The key idea is that the model reads a word in the context of the words on both sides of it, instead of reading a sentence strictly left to right.
That matters because language is full of words whose meaning depends on what surrounds them. “Bank” means different things in “river bank” and “bank account”. Small connecting words matter too. “Flights from Delhi to London” and “flights to Delhi from London” contain the same words but ask for opposite things.
Older systems were good at matching important keywords and weaker at the little words that tie them together. BERT is better at reading the whole phrase as a person would.
What Google actually announced
In its announcement, Google described BERT as the biggest leap forward for Search in the past five years. The main points:
- Queries, not pages, are the focus. BERT is being applied to understand what the searcher means. That then affects which pages Google considers the best match.
- About one in ten searches. Google estimated the change touches roughly 10 percent of English language searches in the United States at launch.
- Longer, conversational searches benefit most. Queries with prepositions and context words, the kind people type when they ask a full question, are where Google saw the biggest improvement.
- Featured snippets. Google said it is also using a BERT model to improve featured snippets in around two dozen countries where snippets are available.
- More languages later. Google said it plans to bring the improvement to more languages and locations over time.
Google gave examples. In one, a search like “2019 brazil traveler to usa need a visa” used to return results about US citizens travelling to Brazil. With BERT, Google understood that “to” mattered and showed information for Brazilians travelling to the US. In another, “parking on a hill with no curb” previously surfaced pages about parking on a hill with a curb, because the system gave too little weight to “no”.
How BERT differs from RankBrain and core updates
People are lumping BERT in with every other ranking change, so it helps to separate them.
RankBrain arrived in 2015. It was Google’s first big use of machine learning in ranking and helps the system deal with queries it has not seen before by relating them to known ones. RankBrain is still in use. BERT adds to it, it does not replace it.
Core updates, like the one widely called the Medic update in 2018, adjust how Google assesses overall page and site quality. Sites often see clear winners and losers after those. If you want background on how those work, see our guide to Google core updates.
BERT is about language understanding. It does not judge whether your site is trustworthy. It helps Google work out what the searcher wants, so the right page gets matched to the query.
This is why many site owners will see little visible movement. If your traffic comes mostly from short head terms, the change may barely register. If you get a lot of long-tail traffic, some of those queries may now go to pages that answer them more precisely, which could be yours or a competitor’s.
Can you optimize for BERT?
Not in the way people optimize for a technical factor. Google’s public liaison for Search said plainly that there is nothing specific to optimize for with BERT. You cannot add a tag, change your markup or adjust keyword density to “pass” BERT.
What you can do is make sure your pages are the kind BERT rewards: content that clearly answers the question a person typed. That has been Google’s direction for years. BERT simply makes the system better at noticing when a page does not really match what was asked.
What this means for your content
Write for the question behind the keyword
If a page targets “how to change a tyre on a hill”, it should actually address hills, not just tyres. Before BERT, a page could rank for a phrase it only partly answered because the main nouns matched. That shortcut is getting weaker.
For each important page, write down the specific question it answers. Then read the page as someone with that question. Does it answer them quickly? Does it handle the details in their wording, such as location, audience, or the “without” and “no” conditions?
Stop writing for keyword strings
Phrases like “best plumber Mumbai cheap” written awkwardly into sentences were never good copy. With better language understanding, there is even less reason to do it. Write the sentence a person would say: “If you need an affordable plumber in Mumbai, here is what to look for.”
You still need to know the terms people use. Keyword research shows you the vocabulary and the questions. Just use that research to shape your topics and headings, not to stuff exact phrases.
Cover long-tail questions properly
Long, specific queries are where BERT helps most, and those searchers are often close to a decision. A good long-tail keyword strategy now pays off even more, as long as the page truly answers each question instead of listing it.
Ways to do that:
- Add focused sections or separate pages for distinct questions, rather than one page that mentions everything once
- Use headings that match the question form (“Can I…”, “Do I need…”, “What happens if…”)
- Answer first, then explain. Put the direct answer in the opening sentence of the section
- Include the conditions and exceptions that matter to the reader
Be precise in featured snippet answers
Since BERT is also used for featured snippets, precision in short answers matters. If you want to win featured snippets, give a direct, accurate answer in two or three sentences near the top of the relevant section. Vague answers that dodge the specific condition in the question are less likely to be chosen.
Keep quality signals strong
BERT does not replace the quality systems behind E-A-T. A page that perfectly matches a query but comes from a thin, untrustworthy site still has a hard time. Clear authorship, accurate information and a good reputation still count.
What to do if your traffic changed
Rankings move for many reasons, and an update announcement makes it tempting to blame everything on it. Before reacting:
- Check timing. Google said BERT rolled out in the week before the announcement. Compare traffic before and after that window, and rule out seasonality.
- Look at queries, not just totals. In Search Console’s Performance report, compare query lists for the affected pages. Did you lose long, specific queries? Did short terms hold steady?
- Read the pages that now rank. If a competitor took a long-tail query, does their page answer it more directly than yours?
- Fix the match, not the metadata. Rewriting a title will not help if the body does not answer the question. Improve the content for the query you lost, or accept that a different page is a better fit for it.
In many cases, traffic you lost to BERT was traffic that did not convert well anyway, because those visitors wanted something your page did not offer.
Common myths about BERT
- “You need to rewrite every page.” No. Most well written pages will be fine or will benefit.
- “BERT penalizes sites.” It is not a penalty. It changes how queries are understood.
- “Use more natural language processing keywords.” There is no special vocabulary that BERT prefers.
- “Short content is dead.” Length is not the point. Some questions need a sentence, others need a full guide. Match the depth to the question.
FAQ
When did the BERT update launch?
Google announced it on 25 October 2019 and said it had rolled out over the preceding days for English searches in the US.
Does BERT affect local searches?
It can affect any English query in the US where language nuance matters, including local ones. The impact depends on the wording of the search, not the type of business.
Should I change my keyword research?
Keep doing keyword research, but pay more attention to full questions and the intent behind them. Use the findings to plan topics and answers rather than exact-match phrases.
Will BERT come to other languages?
Google said it intends to expand to more languages and countries, and it is already using a BERT model for featured snippets in a number of countries.
Write for the person, and the update takes care of itself
BERT rewards content that already does its job well: clear answers to real questions in plain language. Review your most important pages against the questions your customers ask, tighten the answers, and keep building pages for the specific needs you serve. If you want a second pair of eyes on how well your content matches what people search for, our Get found service is a good place to start.
