Google BERT Update - Impacts 10% of Queries, Bigger Than RankBrain, Rolling Out Now

[Ryuzaki]

Let me try to summarize all of this as I consume everything I can find about it. It may seem garbled, because I'm doing this on the fly.

Official Announcement
Google dropped a post today (October 25th, 2019) on their official blog called Understanding searches better than ever before. It talks about how last year they introduced and open-sourced a neural network technique for natural language processing (NLP) pre-training. It's called Bidirectional Encoder Representations from Transformers (BERT).

It lets you train a neural network for questions and answers and is based on "transformers:" models that let you consider a word in relation to all other words in a sentence rather than one-by-one in order. It can understand the full context and is useful in understanding the intent behind search queries, in this case.

Here's an example of the improvements:

The key is that they better understand the word "to" in this context, which has previously been considered a stop-word or filler, I suppose to some degree. The link to the blog post above has several more examples if you're interested.

BERT is being applied to Search and to the Knowledge Graph and is expected to impact 1 in every 10 queries, or 10%. It's being called "the biggest leap forward in the past five years, and one of the biggest leaps forward in the history of search."​

That's the info you get out of the official blog post, but there's been more talk going on that I'll try to summarize below.

More Industry Talk
Time to check the blogosphere / Twitter, etc.

This has been rolling out all week and is almost fully live according to Google. So take a look at your traffic for this week and see if you've had any impact.

It will be used globally in all languages in Search and in the Featured Snippets. It's been particularly useful in Hindi, Portuguese, and Korean.

BERT does not replace RankBrain but improves it as an additional measure. It sounds like it's either RankBrain or BERT that a query will be routed through, but not both. How they make that decision, who knows.

Google claims you can't optimize for BERT any more than you could for RankBrain. That makes sense since it's about them understanding a query itself rather than them pulling more relevant results.

Bill Lambert apparently predicted (leaked?) this. In that link, I told the starts of the Bill Lambert saga, it's pretty humorous. It looks like the guys that were originally trying to discredit him switched gears to now trying to take advantage of his juice. Bill Lambert said a change was coming that was a game changer and even told us to look for information on Friday (today), and even told us on Monday that an update was rolling out. Fun stuff.

My own traffic on my big traffic site has lost about maybe 100-200 visitors a day, but who knows if that's related to BERT or them correcting the previous algorithm changes they released. They always go too far then back off a bit.

​

How's it looking for you guys?

 

[ToffeeLa]

Not really noticed any major changes as yet amongst the seasonal fluctuations. There's one minor site that took a rankings hit recently (before last week) but that might be anything and I haven't had a chance to take a good look at it yet.

 

[bernard]

I suppose this won't be noticeable on rankings, but only on traffic and maybe even more specifically in clicks and CTR rate.

 

[harrytwatter]

Had a few sites seemingly unaffected, one improved, and two hit, one almost halved. The odd thing is the site that was halved has by far the most hand-written personalized content, as in it is a niche I am passionate about and so I wrote about personal experiences with various products with original images etc. Comparatively the site that gained had almost entirely outsourced generic round-up type articles. Not stating I have drawn any direct causation, just some odd trends coinciding with this BERT roll out.

Was looking at the position tracker in SEMRush this morning for a dozen players I monitor, looks like total keyword movement new/lost & improved/declined has almost frozen. Usually for the sites I monitor there are hundreds of ups and downs on a daily basis but yesterday was maybe 1/4th of normal volatility and today looking more like 1/6th. idk if they slowing BERT down or if it's something else.

Anyone else notice anything interesting?

 

[tyealia]

No real changes on 5 sites I track, just the usual small ups and downs.

 

[CCarter]

Wait a minute guys - at no point did Google say BERT launched this week or when the blog post was published. It could have been a month or two beforehand or in one of the previous updates. Bill Lambert simply, AGAIN, predicted the 26-29th update at the end of each month. Literally shit me and @Grind have been talking about for years! It's in the Bill Lambert thread.

When did Google state that "BERT LAUNCHED on X date"?

 

[harrytwatter]

That is a good point. The Google post about the update doesn't list a date but I think many in the community are assuming it was rolled out around the 25th based off Barry's post (https://searchengineland.com/welcom...gence-for-understanding-search-queries-323976)

Not sure if Barry is assuming a rollout date based off the announcement alone or if he saw something on Twitter or specifically asked Gary or John.

 

[Future State]

I read somewhere, I'm pretty sure from an official Google employee (maybe not), that although BERT affects up to 10% of queries, it largely affected short tails more. I guess that makes sense since this thing was made to help better comprehend queries, and the shorter it is the more confusing it must be with less context.

I also read that New York Times took a pretty severe beating by their own admittance. I never looked at any stats though. Some people tried to say it was because they finally got included in the mobile first index, but I doubt that's the case.

 

[CCarter]

largely affected short tails more

You mean long-tail keywords right? All the examples of BERT were 5 or more words.

 

[Stones]

Posted this elsewhere but it probably belongs in here. I've noticed that Bert is filtering out pages that give a price in a currency not used in the area. So a detailed UK based page on Off Grid Solar Panels that mentions prices in any detail will now be filtered out of US based results. Obviously the more niched down you get the more examples of these dropping out of the SERPs you can see.

 

[GNews]

Posted this elsewhere but it probably belongs in here. I've noticed that Bert is filtering out pages that give a price in a currency not used in the area. So a detailed UK based page on Off Grid Solar Panels that mentions prices in any detail will now be filtered out of US based results. Obviously the more niched down you get the more examples of these dropping out of the SERPs you can see.

Google's been doing this before BERT.

Plenty of top marketers who have ecomm shops been posting longtail pages that break down the curerncy for other countries, just to have another page that's valid in search.

I did it for my store back in Jan/Feb.

I create custom imges with the currency, my product, & the country too.....

I try to think ahead of the curve. I def know Google's tech can read into imagery by pixel comparisons......... and i think it helps make page relevancy score higher by even customizing images to have distinctly 'matching' content.

 

[CCarter]

Alright so I'm not seeing any serious movement within the SERP Volatility at a macro level, so I'm starting to suspect that this update people are experience, and resulting in traffic drop has to do with long-tail keywords. I'm going to cook up something in the SERPWoo Lab that creates a metrics for "short-tail" versus "long-tail" keywords and see if I can find something off.

 

[CCarter]

Bing trolling Google: Bing is Now Utilizing BERT at a Larger Scale Than Google

Bing has been using BERT since April, which was roughly half a year ahead of Google. In a blog post, Bing details the challenges it ran into when when rolling out BERT to global search results.

Applying a deep learning model like BERT to web search on a worldwide scale can be prohibitively expensive, Bing admits. It was eventually made possible with Microsoft's 'Azure' cloud computing service.

What was initially estimated to require tens of thousands of servers to achieve was accomplished with 2000+ Azure GPU Virtual Machines. Bing went from serving a three-layer BERT model on 20 CPU cores at 77 milliseconds per inference, to serving 64 inferences in 6 milliseconds using a GPU model on an Azure Virtual Machine.​

 

[funlol]

My two biggest sites got destroyed by BERT. Almost a year later and still no improvement. I've been rewriting poorly performing content and adding brand new content at a rate of 2-3 pages per day since last November. Traffic is stalled.

Any tips?