How Google BERT Changes SEO

When Google releases a new algorithm they describe as “one of the biggest leaps forward in the history of search“, SEOs pay attention.

Good SEO consultants and their clients will want to know how these changes will impact SEO strategy, and if there are any actions that can be taken to improve rankings under the new algorithm.

In this article, we’ll provide more background on Google’s adoption of Bidirectional Encoder Representation from Transformers (BERT), show examples of how it changes rankings for specific queries, and provide recommendations for how you can better align your content strategy with Google’s improved language understanding capabilities.

There have already been many articles written about BERT — why one more? First, I’m writing from the perspective of a Stanford-trained machine learning engineer turned SEO. While I’m not an NLP specialist, I’ve previously built neural nets, including one that could predict the ranking of randomized Google results with 60% accuracy. I also keep a folder of “Google Fails”, which provided some interesting before and after examples demonstrating BERT’s impact. At the very least, I can promise this post won’t just be reiterating Google’s BERT launch release and the same examples you’ve seen everywhere else.

TLDR; Everything SEOs need to know about BERT in 100 words:

• Unlike previous relevancy algorithms that treated texts like a “bag of words”, BERT better understands how relationships between words change meaning and query intent.
• Since BERT understands query micro-variations that had previously been ignored, it should benefit sites with a lot of in-depth topical content that were previously outcompeted by higher authority sites with less topical coverage.
• Because CanIRank’s on-page optimization recommendations are derived from the current top-ranking results, our software will show you exactly how to optimize your content for post-BERT query intent.

What is BERT?

Bidirectional Encoder Representations from Transformers, or BERT, is a new method for pre-training neural network-based language understanding models that better accounts for the fact that a word’s meaning can change depending on the context, a concept known as polysemy.

Previous language understanding models, such as word2vec, relied on vector representations of different words called word embeddings. Word embeddings allow computers to “understand” meaning behind words, as well as their relation to one another:

While powerful, word embeddings generated by these methods had significant limitations because every occurrence of a word was given the same word embedding regardless of context. “Play” has the same word embedding whether the sentence was “my son is starring in his class play”, “did you catch that amazing play in the 7th inning?” or “enough learning about NLP; let’s go outside and play!”.

Other models need to be used to refine meaning based on context. BERT is not the first of those models, but it can do something the others can’t: learn the meaning of a word based on context both before and after the target word (“bidirectional”). BERT’s output includes representations of word meaning similar to word2vec, but adds additional neural net layers that serve to encapsulate meaning due to syntax as well as higher-level aspects of language understanding, such as dependency, negation, agreement, and anaphora.

A Giant Leap for Query Understanding?

To understand why BERT is a significant breakthrough for both Google and SEOs, it helps to look back on the evolution of search engine relevancy algorithms.

Early search engines primarily paid attention to the presence of certain keywords or keyword variations. For SEOs, this meant life was easy: do your keyword research, throw any variation with search volume into a title tag and add a few paragraphs of fluff content.

Boring to implement, terrible for users… But incredibly effective.

This approach worked so well that some started to take it too far, as content farms like eHow cranked out tens of thousands of vapid articles, many covering the same topic but targeting a slight keyword variation.

As Google worked to quash content farms and their proliferation of low-quality content, the pendulum swung far in the other direction: Google began to pay more attention to other signals like authority and site quality, and relevancy shifted to broader “meaning-based” semantic relevancy algorithms rather than keyword usage.

Semantic relevancy algorithms like Google Hummingbird work by understanding the relationships between different words and phrases. For the most part, they treat texts as a “bag of words” where changes in meaning due to the position of one word in relation to another is largely ignored.

With semantic relevancy algorithms, search results for “SEO tools”, “best SEO tools”, “top SEO tools” and even “SEO tool reviews” look nearly identical because Google assumes these searchers are all trying to accomplish basically the same thing.

David Freudenberg, Founder

“With the coming of BERT the familiar saying of “Create Great Content With User Intent In Mind” rings much more true! The game has not changed but the system has got much more advanced forcing more SEOs to play the game the way it’s made to be played.”

While largely positive, it sometimes felt like Google would ignore parts of long-tail queries, especially if they included common words. As a result, SEOs found themselves competing for the same limited subset of query intents as all their competitors. Great if you’re a high authority site like the NY Times; not so much for the niche specialist.

BERT is still very much a meaning-based relevancy algorithm. But for longer conversational queries where the relative position of individual words could change the query meaning, it should no longer feel like Google is ignoring a portion of your query.

Examples of How BERT Changes Google’s Query Understanding

QUERY: “How many small businesses have a blog?”

Before BERT, Google got so excited about the concept of “small business blogging” in this query, that they completely ignored the “how many” and failed to return any results addressing the question.

After BERT, Google now understands the query intent is to find statistics about blogging, and they even identified several closely related questions like “How many businesses have a blog?” and “What percentage of small businesses have a website?”

As impressive as that improvement is, none of the results actually answer the question. Moreover, Google still seems a little uncertain about the intent here as a majority of the results give reasons to start a blog or tips on starting a small business blog rather than statistics.

QUERY: Dolomites climbing grades

I searched this query because I wanted to find out which of the many systems used to grade the keyword difficulty of climbs was employed in the Dolomites in Italy. That way I could translate back to the grading system I was accustomed to (the Yosemite Decimal System used throughout the US), and determine which Dolomite climbs I might be able to do.

Previously, “grades” was de-emphasized and most of the results focused on “Dolomites climbing”. Before BERT, over half the results were identical for “Dolomites climbing grades” and “Dolomites climbing”.

Now, Google gives greater weight to “grades”, and understands that “grades” coming after “climbing” completely changes the intent of the query.

QUERY: Tutorial to buy an electric guitar

Prior to Google BERT, all results for this query were electric guitar tutorials, even though the query intent was actually learning how to buy an electric guitar, and not how to play one. Google picked up on “tutorial” and “electric guitar”, but failed to understand that the admittedly awkward formulation “tutorial to buy” meant that buying was a key part of the query intent.

QUERY: Thank you card from photographer to client

While the old results did have some relevant results, one of the top-ranking pages was https://www.helpscout.com/blog/how-to-write-a-killer-thank-you-note/. This is a classic example of how less nuanced meaning-based relevancy algorithms benefitted sites with high domain authority, even when they weren’t actually addressing the question.

After BERT, the HelpScout post has been pushed down to page 2, and all page 1 results are specific to photography. That’s another win for BERT and a huge relief to those of us writing in-depth content on sites that don’t always have the highest domain authority.

Can you optimize for BERT?

If I had a dollar for every time I read in an article on the BERT launch something like “you can’t optimize for BERT, because it’s machine learning” (or similar), I could take our entire team out to lunch, guacamole included.

How to Adapt Your Content Strategy

By better understanding variations between queries, BERT should reward a content approach that thoroughly covers a topic, and attempts to address lots of different query intents.

A good approach would be to start with a Customer Research Journey analysis so you have a full understanding of the questions your target customers are asking long before they begin querying actual products and services. Focus on knowledge rather than keywords.

FAQs and encyclopedic reference material are two content types that naturally tend to address a lot of different intents and should do better after this update. After all, BERT learned its understanding of language from Wikipedia.

It’s possible the shift towards BERT will slow or reverse the trend towards long form content that we’ve seen in the past few years. As with the example queries above, some long, highly-linked pieces of content are now being replaced by shorter, lower authority content that more directly addresses the query intent.

How to Change Your On-Page Optimization Process

Although Google BERT doesn’t change how Google determines the relevancy of a page, it should change how you do on-page optimization.

The most effective on-page optimization processes in 2019 have focused on understanding and satisfying query intent. It’s no longer about a keyword. It’s about what a searcher is trying to accomplish.