Last Friday, Google announced that it is now using a language model called BERT in Search, and that it affects about one in ten English searches in the US. For a search update, this is a significant figure, and Google rarely quantifies such changes.
Google also did something it seldom does: it published actual example searches, before and after the change.
One of the most instructive examples concerns a pharmacy. Someone searches "can you get medicine for someone pharmacy." Previously, Google mainly matched "medicine" and "pharmacy" and returned general information about filling prescriptions. With BERT, it understands that "for someone" is the central point: the person wants to know whether they can pick up a prescription for another person.
These are two small words that search engines have largely disregarded for years.
What does BERT actually change in Google Search?
BERT was developed by Google's research team, which released it as open source last November. In plain terms, it reads words in context. It considers the words before and after each word simultaneously, instead of treating a search as an unordered collection of keywords.
This matters most for small words such as "to," "for," "no" and "without." Another example Google provided involved a Brazilian traveller needing a visa for the USA. The previous system missed the direction indicated by "to" and returned results about Americans travelling to Brazil. BERT interprets it correctly.
Google is also using BERT for featured snippets, the answer boxes at the top of the page, in the two dozen or so countries where they appear. For now, the main change applies to English in the US. Google states that more languages will follow, and we expect this to happen fairly quickly.
The announcement, Understanding searches better than ever before, is short, and its before-and-after screenshots illustrate the change clearly.
Why write for questions rather than keywords?
Much of SEO still consists of placing the exact phrase on the page: "dive resort Bali" or "plumber Vancouver," repeated in the title, the heading, the first paragraph and the image file names. This approach works reasonably well, but it also produces websites that read as if written for a machine, because they were.
We expect BERT to make this approach outdated quickly. If Google understands the question, the page that answers it clearly should outperform the page that simply repeats the words.
Google's search team states that there is nothing specific to optimise for BERT. Many in the SEO industry were sceptical, but the statement is essentially accurate. The solution is better writing.
What does this mean for a small business page?
- State clearly whom you serve and where. "Daily dive trips from Nusa Ceningan to Manta Point and Crystal Bay" says more than "best diving Bali cheap diving Bali."
- Answer the questions customers actually ask, in their own words. Can I dive if I haven't in five years? Do you accept cash? Is there a pickup from the harbour?
- Give each page one clear purpose instead of combining every service on a single page.
For businesses that do not yet have one, an FAQ built from real customer emails may be the most useful page to add this year. The questions are already in your inbox.
Why should long, natural searches now receive better answers?
People type differently on phones, and they increasingly speak to them. Searches are becoming longer and more sentence-like. Until now, the longer the search, the more likely Google was to misinterpret it and select a few keywords from the middle.
If BERT works as Google describes, long searches will begin to perform better. As they do, people will search even more naturally, creating a self-reinforcing cycle.
For a small local business, this is positive. People do not search "restaurant Lembongan"; they search "somewhere to eat on Lembongan with a view that's open late and has vegetarian food." A page that actually mentions the view, the hours and the vegetarian menu has a genuine chance of appearing for that search. A large chain's site with a generic menu page does not.
What could BERT mean for French and other languages?
For Canadian businesses, this is the key development to follow. When BERT reaches other languages, as Google has indicated it will, the same change will apply to French searches.
Many bilingual Canadian sites have a strong English page and a French page that is a quick translation, sometimes a machine translation. The French often does not resemble anything a person in Quebec would type. "Dépanneur" versus "convenience store" is a simple example, and terms such as "pourvoirie," "érablière" and "cabane à sucre" do not translate cleanly at all.
Once Google interprets meaning rather than matching strings, a natural-sounding French page should have a clear advantage over one that is technically French but reads like an appliance manual. The same applies here to Indonesian and English. Two languages require two pages, each written for how people actually search in that language. This requires more work, but it is the only way for both versions to perform.
Is your site ready for BERT?
There is no setting to change, but there is a simple test.
Read your main pages aloud. If they sound like a person explaining what you do, they are probably well positioned. If they sound like a list of search terms joined together, they are the kind of page this update is beginning to leave behind. An objective SEO audit of a small site does not take long and identifies which pages fall into each category.
We expect BERT to be the first of many such updates. Google clearly aims to understand language rather than match it, and it will improve at this each year. The pages that perform best will be those written for people.
Google has only this month begun to interpret "for someone." Many small business websites still do not state whom they serve, and no model can infer that on their behalf.