Google BERT: Search Gets Better at Reading the Small Words

Posted by Sean on October 29, 2019

Last Friday Google announced that it has started using BERT in search, and described it as the biggest leap forward in five years. That’s a strong claim from a company that usually plays down its changes, so it’s worth a proper look.

What BERT is

BERT stands for Bidirectional Encoder Representations from Transformers, which is a mouthful, and Google’s research team published it as an open-source language model about a year ago. The short version is that it reads words in the context of the words around them, both before and after, rather than one at a time in order.

That matters most for the small words. Prepositions like “to”, “for” and “from”, and words like “no” and “not”, change the meaning of a search completely, and search engines have historically been quite bad at them. Google’s own examples make the point well. One is a search along the lines of “2019 brazil traveler to usa need a visa”. Before BERT, Google tended to return results about US citizens travelling to Brazil, because it didn’t give enough weight to the “to”. With BERT it understands the traveller is Brazilian and heading to the US.

How much does it affect?

Google says BERT affects about one in ten searches in US English, so for now this is a US English change. It’s also being used for featured snippets in two dozen countries, where it applies across languages. I’d expect it to reach UK English and other languages in time, though Google hasn’t given a date.

One in ten sounds like a lot, but it’s mainly longer, more conversational searches, the sort where the meaning depends on how the words relate to one another. Short head terms like “leather bag” or “white shirt” aren’t the target.

Can you optimise for it?

Google’s Search Liaison has been clear: there’s nothing to optimise for. And for once I think that’s simply true rather than Google being coy.

BERT is about Google understanding the searcher’s question better. It isn’t a new set of signals your page can send. What it does do is make it more likely that a page which genuinely answers a specific question gets shown for that question, even if it doesn’t contain the exact phrasing.

If anyone loses out, I’d guess it’s pages that have been ranking for long, specific searches without really answering them, helped by Google matching keywords loosely. Pages written plainly for people should do slightly better.

What it means for ecommerce copy

For an ecommerce site, I think there are a few practical points, none of them new but all a bit more important now:

  • Answer real questions in plain English. Care guides, sizing advice, delivery and returns information, explanations of materials. The searches BERT helps with are often exactly these: “can you wash a waxed cotton jacket”, “how should a shirt collar fit”. Pages that answer them clearly are in a better position.
  • Don’t write for keyword variations. The old habit of writing awkward sentences to fit in a particular phrase was always unpleasant to read. BERT makes it even less necessary.
  • Product descriptions should describe the product. When someone searches for a specific combination of features, a description that actually states those features, in normal language, is what Google can match.
  • Keep an eye on featured snippets. If BERT improves Google’s choice of snippet, the pages that answer a question directly and early are more likely to be picked.

A different kind of update

What I find interesting about BERT is that it changes how Google reads the query rather than how it judges pages. The core updates this year have all been about Google reassessing quality, and they’ve had obvious winners and losers. BERT should mostly just make results more relevant, which is harder to see in a traffic chart.

So I wouldn’t expect a dramatic shift for most UK sites in the short term, given it’s US English only for now. But I’d treat it as a good reason to look at your informational pages and ask whether they answer the questions customers actually ask, in the words they’d use.

← Previous Draft Draft Index Next Draft →