“BERT” is a recent major Google algorithm update focusing on natural language processing (NLP). NLP, in short, is a intersection of AI, computational science, informational engineering and linguistics that aims to enable computers to understand “natural” human language. Computers have traditionally “struggled” (for want of a better word that doesn’t personify a machine…) with processing language that is inherently ambiguous without context.
BERT in context
For example, you’re at the hairdressers and the receptionist says: “the hairdresser is free”. As a human, you know this means that the hairdresser is ready for you and not that the hair cut will be free of charge or that the hairdresser has recently finished a prison term. You know this because of context. NLP models are built to enable computers (like Siri, for example) to work out the context of language. There are obviously far more complex applications of NLP but those are the underlying principles. Until fairly recently, NLP has been driven by lots of individual models – one article I read compared this to having different kitchen utensils for different tasks. BERT incorporates many different NLP models / utensils; think of it as the Swiss Army Knife of natural language processing.
How did Google build BERT??
There’s loads of interesting and terrifying information available about how this update was built which you can read if you are so inclined (TL;DR: Google has taken Wikipedia text and machine learning and a technique called "masking"). In brief, masking is when a random word within a sentence is hidden and an NLP computer must analyse the context of the sentence to predict what that masked word is. BERT stands for 'Bidirectional Encoder Representations from Transformers'. Bidirectional refers to the language processing function whereby the analysis is done on the words both before and after the hidden word. Transformers is a NLP model that processes words and phrases in relation to the other words in the sentence i.e. how they “transform” the context of the sentence. So, for the sentence “Angharad deserves a [MASKED] for this fantastic blog post”, BERT will analyse “Angharad deserves a” and “for this fantastic email” to understand the nuances of context, and therefore predict the word.
BERT has its limitations, obviously, especially around negation diagnostics, but no doubt Google will be pouring money into research to improve it.
So, the BERT update is Google’s next step towards providing better SERPs (or, if you’re cynical, it’s Google’s next steps towards world domination and reading our minds), by enabling the search engine to better understand the context of the intent of the search query.
Why should I care?
Crucially, Google has said that “BERT will only affect complicated search queries that depend on context”. What defines a complicated search is your guess, and language is inherently dependent on context anyway so surely that covers all search queries but hey, who am I to judge? 🤷♀️
Ultimately, this means you cannot really optimise for BERT specifically. Your content (as it always has) needs to be well written, with purpose and in response to a searcher’s intent.
BERT is a huge leap in terms of linguistics and information retrieval but in terms of SEO, your focus remains largely the same. So, if you're rushing to optimise your content for this latest algorithm update, STOP. What you should be doing is optimising for search intent and if you've listened to any advice from your SEO agency, you should already be doing this anyway!
There’s expectation that BERT will impact rich snippets and the meta descriptions scraped from on-page content. One example I read was about the query “can I make / receive phone calls on a plane?”. BERT was able to analyse a huge piece of text that explained why you shouldn’t use phones on planes and returned the answer “No”. Bert didn't have the nuance to understand the intent of the query was not for an explanation about why but rather a simple “can I do this?” yes/no search. Again, your content writing approach remains unchanged: write for the user, not the search engine.
And yes, I am messing with BERT’s processor with that last one.
Blog featured image photo taken by: https://www.flickr.com/photos/karineimagine/2284650110