Optimize for the Human: Understanding Google BERT

Businesses of all kinds rely on the internet to connect with clients. And customers rely on Google search to find what they are looking for. But as Google says, “search is not a solved problem.” Algorithms cannot perfectly model human language.

Google’s latest attempt to better understand search queries is its most important algorithm update in years. Will it affect how you rank in search results? How can your business respond to the update — and is that even possible?

Introducing BERT

Google’s search engine often does an amazing job of understanding vague queries. But its understanding of language is far from perfect. In some cases, even when your website and firm are just what someone is looking for, the nuances of language can keep you from connecting.

Google is constantly working to build models to help computers better understand language the way people naturally use it. This field is called natural language processing, or NLP.

In November 2018, Google introduced a new NLP technique they called Bidirectional Encoder Representations from Transformers, or BERT. A year later, they made it a part of their English-language search algorithm (expanding to other languages in the future).

A transformer is Google’s name for a model that considers the meaning of each word in the full context of all other words in a search query. This is crucial for modeling language because the meanings of words can change dramatically depending on what precedes and follows them.

Understanding nuance

BERT allows the search engine to understand searches it had previously failed to interpret correctly, and Google provided some specific examples of these improvements.

The query “2019 brazil traveler to usa need a visa” is, to the reader, obvious in its intent: to find out how Brazilians can obtain a U.S. travel visa. But pre-BERT, Google would prioritize results concerning U.S. citizens traveling to Brazil. Now, the query returns a link to the U.S. Embassy in Brazil.

Another example illustrated a better understanding of the query, “do estheticians stand a lot at work.” Previously, the top result was a page that did, in fact, talk about working as an esthetician, but matched the term “stand-alone esthetics schools” for the “stand” in the query — a clear error. Now, the top result directly concerns the physical demands of the job.

Other examples Google provided include:

• Returned a book for “young adults” on a query about a book for “adults,” not understanding that “young adult” did not match the query.
• “Parking on a hill with no curb” ignored the all-important word “no.”
• “Can you get medicine for someone pharmacy” — meaning, pick up a prescription for someone else — failed to understand the phrase “for someone.”

Incremental progress

Google updates its search algorithms frequently, with major, minor, confirmed and unconfirmed changes.

Panda was first launched in 2011. It penalizes thin content, duplicate content “farmed” from other sources, and pages with large numbers of advertisements.

Penguin came about a year after Panda, in 2012. It focused on recognizing high-quality backlinks (links to your site from other sites). It penalized sites with backlinks that were artificially generated and rewarded those with high-quality backlinks.

Hummingbird arrived in 2013, and with it, Google took serious steps to better understand the context and intent of search terms. Searchers could start to move away from queries that were simply lists of keywords and instead use natural language and even ask questions in the search bar.

And now, with BERT, Google is raising the bar on understanding natural language in a big way.

Staying ahead

Note how the ultimate goal — and the constant challenge — has always been to deliver the best possible search results to the user. At first, the focus was on weeding out low quality results — sites which never intended to be useful, or which failed to rise to the challenge. Increasingly, Google is now focusing on understanding the intent behind search queries.

The rising prevalence of voice search is a key driver in this effort. When people dictate searches into their phones, this yields longer, more conversational queries. People are less likely to use “keyword-ese” when speaking, meaning the search engine needs to adapt.

So as with previous search algorithm updates from Google, as well as those yet to come, there is not a great deal to be done to optimize for the specific update. Instead, you must continue to optimize for the human. This means well-written content, regularly updated, which is useful to clients. You have heard all this before.

However, here are a few pointers to understand optimization in this context.

Continue focusing on “longtail” search queries. These are the queries that show up less frequently because they are specific or wordy. They often include complete sentences and questions. Potential clients use these queries because they know only the particulars of their case, without knowing the terminology for such cases and the attorneys who handle them.

Think about the search queries that you target, particularly the longer and more specific ones. Try searching for variations with plain-English phrasing and see how Google interprets them — especially cases where a preposition (to, on, for, etc.) is key to the meaning. Can you identify searches that are not returning good results? What part of the query is being left out or misinterpreted?

Use clear and plain language. This is particularly true in regards to articles which must include technical terminology. Educate the reader by defining legal terms, but also phrase things the way the reader might do.

Do not become complacent about monitoring your performance on key searches. Keep thinking about ways to make the meaning and intent behind your content clear. As good as Google’s search is, it still needs a little help.

To illustrate this, Google pointed out how its search responds to the query, “what state is south of nebraska.” Amusingly, the best guess is a community called South Nebraska. The day may come when computers can truly understand human language; until then, optimize for the reader, but keep an eye on the algorithm as well.