According to Google, the BERT Update is one of the biggest advances to-date in the evolution of search and the biggest update in the past five years. Understandably, the the announcement of the update ruffled a lot feathers and raised a lot of questions for online marketers and SEOs: Are my rankings in danger? What should I optimize the website for in the future? What does BERT mean for voice search? Detailed answers to the most important questions about the Google BERT update can be found in this article.
If you are looking for support with the search engine optimization of your website, you can get more information and analysis from our experts:
Video: All you need to know about BERT
In a webinar with Jan Pötzscher, CMO at semcona, we tried to answer all your questions about Google’s BERT Update. For anyone who couldn’t attend the BERT webinar on November 22, 2019, here’s the video. The webinar covers the following topics:
- Why Google’s new BERT Update affects your daily content marketing and SEO routines and how your understanding of relevance has to change.
- How to identify relevant and trending topics, as well as keywords which fit into your content marketing strategy.
- How to write and enrich texts that fit BERT’s demand for relevance, holism and user centricity.
We also received many additional questions from our webinar audience, who wanted to know about the BERT Update, the cause of ranking drops and how to optimize their websites. In this post, I will try to answer the most relavant questions. Here we go:
There was massive turbulence on our site (and some others) on the 13th of September. Could this have been an early rollout of BERT? We lost brand search terms and our keyword visibility dropped by 30%. There was no manual action taken on our site, but it definitely looks like a penalty.
The time you mention coincides with the Google September 2019 Core Update, which was officially announced by Google and which has nothing to do with BERT. Many websites saw wins and/or losses in their rankings. Due to these rankings shifts, many websites saw an increase or decrease in their SEO Visibility (and traffic). It is important to make clear that this is not a penalty. It means that Google considers your website for the terms you lost rankings for less satisfying for the users than the pages that have replaced it in the rankings. There can be various reasons for this.
When do you think the new Google algorithm will be rolled out in Europe? Did BERT already affect Germany? Will BERT affect English searches in Germany as well?
BERT is an update to how Google interprets search queries. Initally, this advanced form of ML-trained interpretation only applied to organic searches in the English language for Google.com’s US index. The worldwide rollout of BERT, affecting the organic search results for more than 70 languages, began in early December 2019. Regarding Featured Snippets, BERT was already in use worldwide from the beginning, so from the 24th of October in all languages.
“Bert is a logical development for Google, following in the footsteps of Panda, Hummingbird and RankBrain. However, this time we’re not looking at a change in the way data is indexed or ranked. Instead, Google is trying to identify the context of a search query and provide results accordingly. This is an exciting addition to what context-free models like Word2Vec and GloVe are able to offer. For Voice Search and Conversational Search, I would expect to see significant leaps forward in the quality of results in the near future.” – Malte Landwehr, VP Product, Searchmetrics
We operate on the Dutch market, so we haven’t felt the impact of BERT so far. Does that mean we should remain more ”old school” keyword oriented? Or should we start focusing more on long-tail keywords and topic that are interesting for users?
Even without BERT, Google has already been trying to do this – understand the user’s input and deliver the best answers. For many years now, Google has been able to list pages as good results even if the searched for terms are not included – because Google is already quite good at delivering content that matches a user’s intent. With BERT they will only get better at this.
To cut a long story short: Please do not stick to old-fashioned SEO techniques by stuffing your content with keywords you want to rank for. Build content for the user! This should already be the case and this will not change with BERT. If you want a data-based approach to how this content should be written, you could also use our tool “Content Experience”, and get access to our AI-driven approach to content creation.
Our website has been experiencing a tremendous visibility increase during the last two weeks in Germany. Since you said that BERT hadn’t hit Germany yet – does that mean there must be another reason?
It is most likely that your website has been affected by one or some of the many smaller updates that haven’t been confirmed by Google but that have happened in recent weeks, so not related to BERT. Google constantly optimizes its algorithms. In this respect, it is quite possible that, for example, the last media update or cleaning the SERPs had an impact on your site.
A website we help with great content has lost 80% visibility – can this be BERT or another reason? No manual action by Google was taken.
Is it SEO Visibility or is it GSC Impressions/Traffic? I personally think it is quite unlikely that BERT affects many searches with high search volume. Google understands quite well whether they meet the user’s expectations by analyzing the CTRs, Back-to-SERP rates and final-clicks. I rather believe that they apply BERT to more long-tail search queries, which are most likely searched for the first time ever in Google’s search history (the human language is so complex – about 15% of all searches done on Google have never been searched before). It is hard for Google to understand such unknown phrases and here AI, ML and with that BERT come into play. As a consequence, I think BERT mainly hits phrases that SERP-listening services like Searchmetrics SEO Visibility are not even tracking. So you might be more likely to see an increase or decrease as a result from BERT in your Google Search Console search analysis data.
Would you recommend only focusing on long-tail keywords? Or should we use a mix of short-head keywords and long-tail keywords? Would you also recommend using more long-tail keywords for category pages?
I would never recommend focusing on long-tail keywords. The long tail covers those search terms that are almost never searched for. Only huge masses of long-tail searches can add up to masses of traffic. But how would you even want to focus on masses? My recommendation is still to satisfy your users and your target audience by providing content that targets their needs and user intent. If you do your research on this and provide the right content then you will automatically rank for the long tail. And with BERT coming into play, Google will learn to understand the long tail even better and deliver the best search results that match the user intent – and (my guess is that) more and more pages will rank for those long tail keywords without even containing the words the long tail is made up of.
Do you think that headless CMS might be a solution in the future for landing pages that are optimized for long-tail keywords?
The headless CMS will not help or hinder your (long-tail) optimized pages from the keyword coverage point of view. They will help in speeding up your websites, which is good for your SEO, but this is not related to BERT or keywords (neither long-tail nor short-head).
Considering a user intent is to purchase, does creating SEO content still provide a competitive advantage?
Firstly, please do not – for any page – create SEO content. If there is content that you think is only there “for the search engine” and your visitors are not interested in it – delete it (or better: don’t even write it to start with).
Let me rephrase your question: “Considering a user intent is to purchase, do you think some users might need further assistance that could be delivered through a paragraph of text or two?”
Yes, I do think that not all of the visitors to your eCommerce pages – although they already do have the intent to buy something – already know which product to buy and are only looking to click on it and complete the transaction. Some of them might need assistance from you, helping them to understand how to decide which is the perfect product for them. Your content could help them with their decision. Write it for those users. Design it for those users. Help those users to see this content easily so they can benefit from it. This will also be recognized by Google and you might get more traffic from people needing this advice, and have them purchasing on your website.
Should I create content that is thematically similar to the competition and which alreaddy ranks on page 1?
Yes and no. Just thinking logically, the content already ranking on page 1 is considered relevant by Google for a given topic and search intent. This means that content semantically similar to the top 10 ranking pieces might also be considered relevant.
However, if you just semantically copy what is already ranking on page 1, Google will have no reason to exchange one of the current rankings for your page – because your page would not provide any insight, information or word that is not already being covered by the others.
Your aim needs to be to semantically include what is already considered relevant and become better, more complete, more satisfying, more engaging, more multi-facetted – simply: to become the best page on this topic and search intent that can be found on the whole internet. Create a page that Google would be ashamed not to rank on first position for a topic – then you will have reached your goal. You will then have set a new standard for what other pages need to semantically contain to rank on page 1.
Can you summarize BERT’s most significant change in one sentence?
BERTs most significant change (for users of Google) is that it understands complex search queries much better than before. (one sentence 😉 )
Can you explain what “bi-directional encoder representation from transformers” mean?
I am not a Machine Learning expert, but from what I read and understand it is the following: BERT is a Machine Learning procedure that consumes a lot of content in order to “experience” human language. It then takes this input and deletes random words from it, leaving them blank to then “guess” what the best-fitting words for these blanks would be. This bi-directional usage of input helps the ML algorithms to use the training data a lot more often (and unsupervised) to continually learn, test, and learn from it. In this way, the algorithms, for example, can understand that in some sentences small prepositions (which might have mistakenly been interpreted as “STOP-words” by algorithms of the past that were only focused on keyword queries) are actually important for the meaning of a phrase. Google has therefore taken a big step further in the direction of actually “understanding” what a phrase means.
Some links I would like to recommend for readers who are more interested in the technical aspects of BERT:
- The illustrated BERT, ELMo and co from Jay Alammar written in 2018: http://jalammar.github.io/illustrated-bert/
- Understanding BERT: Is it a Game Changer in NLP? by Bharat S Raj: https://towardsdatascience.com/understanding-bert-is-it-a-game-changer-in-nlp-7cca943cf3ad
- What are the main differences between the word embeddings of ELMo, BERT, Word2Vec and GloVe? Answer from Ajit Rajasekharan: https://www.quora.com/What-are-the-main-differences-between-the-word-embeddings-of-ELMo-BERT-Word2vec-and-GloVe
What conclusion should I make based on the update, how should I react and optimize my website based on BERT? Are there any changes due to BERT on a more technical side like naming of pictures, H1 headers etc.?
No. You should optimize your website to give the best answers to your users’ intent. The better your content serves this purpose, the more likely it iis to rank for search phrases interpreted by BERT-trained algorithms.
How do you see Voice Search queries in relation to BERT?
The spoken language is so much more complex than the written language. This means that in spoken language/voice search, Google will encounter even more search phrases that have never previously been heard than in classical desktop-based search. As BERT is improving the understand of complex sentences, it will also become extremely important for Voice Search.
How can I optimise my content for Voice Search?
For the first iterations of Google Voice Search (different from Amazon Voice Search), you could produce answers to informational search queries that answer questions with short and concrete paragraphs. As much of search query interpretation is still based on terms, I would recommend collecting some ways that people could ask for this special information that you intend to give an answer to, choose one of those questions as a heading, and implement the words used in the other questions to craft your answer. This way, multiple word combinations that might be used in the (spoken) question would be found in your question and answer. Your answer should sound gentle and concrete when read out loud. You might want to use “speakable” structured data markup to show Google that your answer could be a good match for answers that are read out aloud to the users.
In the future, a step could be to actually provide spoken content on the landing page that Google (one day) could simply cut off a part of it to “play” the answer and recommend users to listen to the whole spoken text.
Are podcasts better for Google than blogposts? Or is it a mix of content that boosts your Visibility in Google search?
Currently, I would say neither podcasts nor blogposts are good for increasing your SEO Visbility on Google. The spoken text of podcasts is not yet being analyzed and used to give answers on Google Search. Blogs are usually meant as some kind of magazine, like a newspaper which could rank well for the time it was published but, as time passes, it is going to get lost in the archives, poteentially retaining some long-tail rankings and attracting some visitors. Writing a blog for SEO is not a strategy I would suggest. Crafting evergreen content however, information that gets updated every time there is something to add to (or to delete because information also can get outdated) would be the best way of building content that I can suggest.
How good is Google right now in understanding audio, video, or image files such as podcasts, YouTube, Instagram etc.?
I think they are pretty good at understanding spoken language (just look at the captions Youtube automatically creates for spoken footage of videos uploaded, or look at how well Google Home understands your questions). They are also quite good at understanding what is shown on images – they even sell their capabilities in this area to whoever wants to use it (https://cloud.google.com/vision/). Movies are just a combination of spoken language and images. But if you want to know more about my personal opinion on how much of these capabilities they are actually using for understanding image, sound and movie content: I don’t think they use their capabilities much. My guess is that this would be too time and energy consuming. And currently, with Google being the (in my opinion) best web search engine in the world, they don’t need to use all their powers yet … it is enough to know that they are ready to be implemented once they are needed in order to stay ahead of everyone else.