Google BERT Update: Machine Reading Unlocked

Imagine you’re standing in a busy train station in London. You walk up to the ticket office, but instead of asking, “Can I get a return ticket to Manchester for next Tuesday off-peak?” you shout, “Ticket. Manchester. Tuesday. Cheap.”

For years, this is exactly how we had to type to Google. We stripped away the grammar, the politeness, and the nuance of our beautiful language. We spoke “Robot” because we knew that if we typed like humans, the search engine would get confused. We had to dumb ourselves down to be understood.

Then came BERT.

In late 2019, Google rolled out an update that fundamentally changed the digital landscape. It wasn’t just a tweak to the rules; it was understanding context, prepositions, and the subtle intent behind our messy, complex questions.

This is the story of the Google BERT update—the moment artificial intelligence finally cracked the code of human conversation. Whether you’re a business owner in Brighton, a student in Northampton, or just someone curious about how the internet actually works, this guide will take you through everything you need to know.

Please note: The content below may contain affiliate links. If you make a purchase through these links, we could earn a commission, at no additional cost to you.

1: The Dark Ages of “Keywordese”

To understand why BERT was such a revolution, we first have to look at the subtle limitations of search before it arrived.

The Old Way: The “Bag of Words” Problem

By 2019, Google was already incredibly smart. Thanks to previous updates (like Hummingbird and RankBrain), it wasn’t just matching strings of letters anymore. It knew that “chips” in Brighton meant food, not microprocessors. It understood concepts.

However, it still tended to treat a search query like a “bag of words.”

It would look at your sentence, pick out the most important keywords (the nouns and verbs), and largely ignore the rest. It struggled to understand how the small words—the “glue” words like to, for, no, and without—connected those concepts together.

The Problem with Nuance

This limitation led to a phenomenon known as “Keywordese.”

We all learned to speak it. Even though Google had evolved, our habits hadn’t. We assumed that if we added too much detail, the search engine would get confused.

If you had a backache, you might not type, “My lower back hurts specifically when I sit down but not when I stand.” You would likely type: lower back pain sitting symptoms. We stripped away the grammar because we knew Google historically struggled to give weight to the sentence structure itself.

Before BERT, if you searched “parking on a hill with no kerb,” Google might heavily weight “parking,” “hill,” and “kerb,” but miss the crucial modifier “no.” It might happily show you a page about “how to park on a hill with a kerb,” because it matched 3 out of 4 main concepts, failing to realise that the word “no” completely reversed the user’s intent.

Google was a brilliant matching engine for topics, but it wasn’t yet a reading engine that understood the flow of a sentence.

Part 2: Enter BERT (Not the Puppet)

It sounds like a friendly neighbour from Sesame Street, but the acronym is quite a mouthful:

Bidirectional Encoder Representations from Transformers.

Don’t worry if that sounds like sci-fi techno-babble. Let’s break it down into plain English.

1. Transformers (The Brain)

No, not the robots in disguise. In the world of Artificial Intelligence (AI), a Transformer is a specific way of building a model that processes language.

Before Transformers, computers read text sequentially, either from left-to-right or right-to-left. Imagine reading a book where you can only see one word at a time through a tiny hole in a piece of paper. You read “The,” then “cat,” then “sat.” By the time you get to “mat,” you might have forgotten “The.”

Transformers changed this. They allow the computer to look at all the words in a sentence at the same time. It’s like taking away the paper with the hole and letting the computer see the entire page at once. This allows the AI to see how every word relates to every other word instantly.

2. Bidirectional (The Two-Way Street)

This is the “B” in BERT, and it’s the secret sauce.

Remember the “stand” example? To understand what “stand” means, you need to look at the words that come before it and the words that come after it.

• Unidirectional models (the old way) would read: “Do aestheticians stand…” -> okay, we have the word stand. It hasn’t seen the rest of the sentence yet.
• Bidirectional models (BERT) look at the whole thing: “…aestheticians stand a lot…”

BERT looks at the text in both directions simultaneously. It uses the context of “aestheticians” (left) and “a lot” (right) to figure out that in this specific sentence, “stand” refers to the physical act of being on your feet, not a fruit stand or a music stand.

The “Fill in the Blank” Game

How did Google train BERT to be so smart? They made it play games.

They fed BERT billions of sentences from Wikipedia and digital books. But before showing BERT the text, they hid some of the words (about 15% of them). This is called Masked Language Modelling.

• Sentence: “The Queen lives in Buckingham [MASK].”
• BERT’s Job: Guess the missing word.

To guess correctly, BERT had to learn that “Buckingham” usually goes with “Palace” when talking about the Queen. It had to learn facts, grammar, and context all at once. It played this game billions of times until it had built a statistical map of the English language that was more complex than anything that existed before.

Part 3: The Launch and The Impact

Google launched BERT in the US for English queries in October 2019. By December, it was rolling out globally, including here in the UK.

Google rarely gives us specific numbers, but for this update, they were bold. They stated that BERT would impact 10% of all search queries.

That might sound small, but when you consider Google handles trillions of searches a year, 10% is a colossal number. It meant that overnight, millions of search results shifted.

Why “10%”?

Why didn’t it affect everything? Because for simple searches, the old system worked fine.

• Query: “Weather in Leeds.”
  • Old Google: Works perfect.
  • BERT: Not needed.
• Query: “Can you get medicine for someone else at the pharmacy NHS.”
  • Old Google: Might focus on “medicine” and “pharmacy” and show generic location pages.
  • BERT: Understands “for someone else” is the crucial part of the question.

BERT was designed for the long-tail—those specific, conversational, complicated questions that we actually ask when we’re worried, confused, or planning something complex.

Part 4: BERT in Action (Examples)

Let’s look at some real-world examples of how BERT changed the results we see.

The “Brazil” Visa Problem

This is the most famous example Google shared, and it illustrates perfectly the importance of prepositions (words like to and from).

The Search: “2019 brazil traveler to usa need a visa.”

• Before BERT: Google ignored the word “to.” It saw “Brazil,” “USA,” “traveler,” and “visa.” It assumed the user was a US citizen wanting to go to Brazil (because that’s a very common search). The top result was an article about US citizens travelling to Brazil. Totally wrong.
• After BERT: The algorithm understood that the word “to” is doing a lot of heavy lifting here. It realised the direction of travel was Brazil -> USA. It served a result from the US Embassy in Brazil explaining visa requirements for Brazilians.

The British Context: “Parking on a Hill”

Let’s look at a scenario that resonates with anyone who’s tried to park in a hilly British town like Bristol or Sheffield.

The Search: “Parking on a hill with no kerb.”

• Before BERT: Google placed too much weight on the word “kerb.” It might ignore the word “no.” The results would often be generic advice about parking on hills with kerbs, telling you to turn your wheels towards the kerb. This is dangerous advice if there isn’t one!
• After BERT: BERT understands that “no” modifies “kerb.” It changes the entire meaning of the sentence. The new results prioritise advice specifically for roads where the kerb is missing, possibly saving your car from rolling down the street into the path of other vehicles if the brakes fail.

The Polysemy of “Bank”?

“Polysemy” is just a fancy word for “one word, many meanings.”

Imagine you search: “Walking along the bank.” BERT uses the context of “walking” to know you mean a riverbank, not a Barclays branch.

If you search: “Opening a bank account.” BERT knows you mean a financial institution, not a river.

Before BERT, Google had to rely on the other words in the query to guess. If you just typed “bank,” it was a toss-up. BERT is much better at looking at the whole sentence to figure it out.

Part 5: BERT vs. RankBrain vs. Hummingbird

It can get confusing keeping track of all these bird names and acronyms. A common question is: “Did BERT replace RankBrain?”

The answer is no. They work together. Think of Google’s algorithm as a football team. Each player has a different job.

3. BERT (The Linguist)

BERT comes in later. It doesn’t just guess based on past data; it reads the current sentence.

• Analogy: BERT is a professor of English Literature. It parses the grammar and the structure of the sentence to understand exactly what is being asked right now.

Today, when you search for something, RankBrain might try to figure out the general topic, while BERT ensures the specific nuance of your words is respected.

Head-to-Head: The Material Difference

To really nail this down, let’s look at a specific comparison. While both use AI, they are solving different problems.

In short: RankBrain helps Google figure out what you are searching for when you are vague. BERT helps Google figure out precisely what you mean when you are specific.

Part 6: How BERT Affected British Content and SEO

When BERT landed, there was panic in the world of SEO. Marketing agencies across London, Manchester, and Leeds scrambled to figure out how to “beat” the new system.

But here’s the twist: You can’t optimise for BERT.

At least, not in the way we used to.

The End of “Robot-Speak” and Exact Match

By 2019, “keyword stuffing” was already dead (thanks to earlier updates like Panda). However, content creators were still guilty of writing in “Robot-Speak.”

We often wrote slightly awkward headings to ensure an “exact match” with a high-volume keyword. If a keyword tool told us that “trains london edinburgh” had more search volume than “trains from London to Edinburgh,” we would force the grammatically choppy version into our titles just to please the algorithm.

BERT ended this era of rigidity.

Because it understands the semantic function of “glue words” (like to, from, in), it freed writers to use natural grammar. You could finally write “Cheap flights to Spain” and rank perfectly for the query “cheap flights spain”.

The Rise of Natural Language

For British businesses, BERT was a blessing in disguise. It meant they could stop writing for robots and start writing for their customers.

If you run a bakery in Cornwall, you don’t need to worry about cramming in keywords. You just need to write a helpful, detailed page about your pastries. You can use local slang, you can write in full sentences, and you can answer specific questions like, “Do you deliver pasties to holiday cottages near St Ives?”

BERT is smart enough to match that content to a user searching for “food delivery St Ives holiday let.”

E-E-A-T and The Trust Factor

While BERT handles the understanding of the language, Google still needs to know if your content is any good. This is where E-E-A-T comes in:

• Experience
• Expertise
• Authoritativeness
• Trustworthiness

BERT helps Google understand what your page is about. E-E-A-T helps Google decide if it’s worth showing.

For example, if you’re writing about medical advice (like “paracetamol dosage for kids”), BERT ensures Google understands the question perfectly. E-E-A-T ensures the answer comes from a qualified doctor or the NHS, not a random blog.

Part 7: Practical Tips: Writing for the BERT Era

So, you have a website, a blog, or a business. How do you make sure BERT likes you? Here’s a checklist for the modern web.

1. Answer the Question Directly

BERT loves clarity. If your article title is “How to boil an egg,” don’t spend the first 500 words talking about the history of chickens.

• Bad: “Chickens have been domesticated for thousands of years…”
• Good: “To boil an egg, place it in boiling water for 6 minutes for a runny yolk.”

2. Use “Long-Tail” Keywords

People are searching using their voices more than ever. They ask Alexa or Siri full questions. Instead of targeting “running shoes,” target “best waterproof running shoes for muddy trails.” BERT excels at connecting these long, specific queries to the right content.

3. Don’t Fear Prepositions

Write naturally. Don’t cut out words like “to,” “for,” or “with” just to make a headline shorter. Those words give BERT the clues it needs to understand context.

4. Structure Matters

Use clear headings (like the ones in this article). Break your text up. This helps the AI (and the human reader) digest the information in chunks.

5. Be Specific (The “UK Nuance”)

If you’re targeting a British audience, use British terminology.

• Use “Mobile phone” not “Cell phone.”
• Use “Trousers” not “Pants.”
• Use “Autumn” not “Fall.”

It’s smart enough to know the difference, but using the language of your audience builds trust and relevance.

Part 8: Beyond BERT – The Future of Search

It was a revolutionary step, but in the fast-moving world of AI, it was just the beginning.

SMITH (The Big Brother)

Shortly after BERT, Google introduced models like SMITH. While BERT is great at understanding sentences, SMITH is designed to understand entire documents. It helps Google know if passage A at the top of a page contradicts passage B at the bottom.

MUM (The Multitasker)

In 2021, Google introduced MUM (Multitask Unified Model). Google claims MUM is 1,000 times more powerful than BERT. Where BERT reads text, MUM is multimodal. It can understand text, images, and video simultaneously.

• Example: You could snap a photo of your hiking boots and ask, “Can I use these to hike Ben Nevis?” MUM understands the image of the boots, knows the terrain of Ben Nevis, and gives you an answer.

Gemini (The New Era)

Today, we’re moving into the era of Gemini, Google’s most capable AI model yet. Gemini integrates the language understanding of BERT and the multimodal power of MUM into a seamless assistant.

But make no mistake: BERT was the foundation. It was the moment the machine learned to speak our language. Every time you ask a complex question and get a perfect answer, you have BERT to thank.

Conclusion: The Humanising of the Machine

The story of the Google BERT update is really a story about us.

For decades, we tried to teach humans to speak like computers. We taught ourselves to think in keywords, to strip away emotion, and to be robotic in our quest for information.

BERT flipped the script. It was the computer learning to speak like us.

It validated the way we naturally communicate. It acknowledged that the word “to” matters. It recognised that “no” changes everything. It allowed us to be messy, specific, and human in our curiosity.

So, the next time you type a long, rambling question into Google—something like, “where to buy vintage vinyl records in Camden open late on a Sunday”—and it gives you the perfect answer, take a moment to appreciate the technology working behind the scenes. The machine isn’t just matching words anymore; it’s listening to what you actually mean.

Further Reading

• Google Search Central Blog – Understanding searches better than ever before – The official announcement from Google.
• Search Engine Land – A deep dive into BERT – Industry expert analysis.
• Moz – The Google Algorithm History – A timeline of all major updates including BERT.