RankBrain: Google Admits a Machine Is Reading Our Queries

Posted by Sean on October 30, 2015

On Monday, Bloomberg ran a story that has had the SEO world talking all week. Google confirmed that a machine learning system it calls RankBrain has been helping to process search results for some months, and a Google research scientist, Greg Corrado, described it as the third most important of the signals that contribute to a ranking.

That’s a striking thing for Google to say out loud. It doesn’t usually rank its own signals in public, and the idea that something most of us hadn’t heard of last week sits third on the list is enough to make anyone sit up.

What we actually know

Not a lot, if I’m honest, and I’d be suspicious of anyone claiming to know much more.

From the Bloomberg piece and the comments Google has made since:

  • RankBrain is a machine learning system, meaning it learns from data rather than following rules written by engineers.
  • Its main job appears to be interpreting queries, particularly ones Google hasn’t seen before. Google has said for years that around 15% of the searches it gets each day are new to it, and those are hard to answer well with rules built on past data.
  • It seems to work by turning words and phrases into something a computer can compare mathematically, so that it can make a sensible guess at what an unfamiliar phrase means based on similar ones it does understand.
  • It has been live for a while. Nobody noticed a sudden change, which tells you something about how gradual its effect has been.

What we don’t know is how it weighs anything, what it looks at beyond the query, or how much its role will grow. “Third most important” makes a good headline but doesn’t tell us much on its own.

Why long-tail keywords are where it matters

The long tail is the collection of longer, more specific searches that each happen rarely but together make up a large share of all search traffic. Think “steel sports watch with blue dial under 40mm” rather than “watches”.

For years, a lot of SEO practice treated the long tail as a numbers game. Find the variations, write a page targeting each one, repeat. That worked when Google matched queries to pages largely by the words in them, because the page with the exact phrase had an advantage.

Hummingbird, a couple of years ago, was the first big signal that Google wanted to understand the meaning of a whole query rather than match its individual words. RankBrain looks like a much more capable version of the same idea. If a machine can work out that two quite differently worded searches mean the same thing, then the page that answers the underlying question well is going to beat the page that happens to repeat the phrasing.

What I’d change, and what I wouldn’t

I don’t think anyone needs to tear up their SEO plan this week. You can’t optimise for RankBrain directly, and I’d ignore anyone selling a service that claims to. But it does strengthen some arguments I’d already have made.

Stop writing near-duplicate pages for keyword variants. Separate pages for “luxury watches for men”, “men’s luxury watches” and “luxury men’s watches” were always a bit silly. If Google treats them as the same question, they compete with each other and none of them is especially good.

Write for the question behind the search. What is the person trying to do? Compare options, check a price, understand a term, find somewhere to buy? A page that answers that properly will hold up better than one built around a phrase.

Use natural language. Write the way people talk and ask questions. As more searches come from phones, and more of those are spoken rather than typed, queries get longer and more conversational. That’s exactly the kind of input a system like this is built to make sense of.

Keep the basics in place. Titles, headings, structure and clear on-page language still help Google understand a page. Machine learning on the query side doesn’t make the page side less important.

The bigger picture

The part of this that interests me most is the direction. For most of the history of SEO, we’ve been working out the rules of a system written by people. As more of Google’s ranking is handed to systems that learn, those rules become harder to reverse-engineer, possibly even for Google’s own engineers.

I suspect that pushes SEO, slowly, towards useful content and sites people are happy to land on, because those are harder to game. Google has been saying that for years, of course, and plenty of us have nodded along while still writing pages for keyword variants. I’ll be interested to see how much more they’re willing to say about RankBrain over the next few months.

← Previous Draft Draft Index Next Draft →