Gary Ilyes Answers How RankBrain Works

[Montresor]

Gary Illyes started his AMA on Reddit [reddit.com] a day early to answer a question about RankBrain.

Searchenginejournal [searchenginejournal.com]

“RankBrain is a PR-sexy machine learning ranking component that uses historical search data to predict what would a user most likely click on for a previously unseen query.”

Lastly, Illyes calls out those who overcomplicate things and theorize that RankBrain uses on-page signals.
“Dwell time, CTR, whatever Fishkin’s new theory is, those are generally made up crap. Search is much more simple than people think.”

[Wilburforce]

“RankBrain is a PR-sexy machine learning ranking component that uses historical search data to predict what would a user most likely click on for a previously unseen query.”

That seems to me to describe what it does, not how it works.

Sadly, predicting what users will click on is not the same thing as knowing what users want or what is best for them. The better Google predicts, the more like lab rats we become.

Sadly, also, the more someone tries to lead me by the nose, the less I wish to follow them.

For me, Google has long since become the search tool of last resort, not of first preference.

[phranque]

sick burn

[Propools]

I wonder to what extent RB will (or already does) use a specific users interactions to make such predictions on how best to serve up 'clickable' search results? -OR- Take learned responses from a group, then apply and blend those results to that specific users search behavior and serve results. Thinking out loud.

RankBrain relies more on data gathered from users’ interactions with search results, versus users’ interactions with a piece of content.

[robzilla]

Link to AMA: [reddit.com ]

[tangor]

Kinda "leading by the nose" manipulation of results. Not impressed.

[Shepherd]

...ranking component that uses historical search data to predict what would a user most likely click on...

...data about what happened on the results page itself, not on the landing page.

Using data from the SERPs to organize the the organic search results? But, but, but, I thought scores of engineers would be up in arms about this sort of thing, this sort of "optimization" of the SERPs.

Let's hear that SEO pro friendly theory about how google's increase in ad clicks is a product of new cell phones in Indonesia again, I like that one...

[MrSavage]

They could literally say anything about how something works and it couldn't be proven, nor dismissed. Nobody will ever know what is, or what isn't. These are tightly locked up secrets. Trust the Google. They have everyone's, and I mean everyone's best interest at heart.

[aristotle]

Well most people will click on the search result that LOOKS most relevant to their query, and/or LOOKS most likely to answer their question. So there's actually some logic to the idea that "most likely to be clicked" is also "most likely to provide the sought-after information".

So historical data on what tends to get clicked, if properly used, could form a valid basis for a part of the ranking algorithm.

[Shepherd]

data about what happened on the results page != data about what happened in the organic results

[EditorialGuy]

Well most people will click on the search result that LOOKS most relevant to their query, and/or LOOKS most likely to answer their question.

Also, some page titles are inherently more inviting than others. The search result that reads "Red Widgets - The Complete Guide" is more likely to be clicked than "Red Widgets Scarlet Widgets Crimson Widgets Thingamajigs Thingamabobs Whatsits."

Sure, Google could rewrite your page titles for its SERPs, but in my own experience, that hardly ever happens if you're using page titles that were created for users and not as fodder for search crawlers.

[JS_Harris]

If you can generate content(automated or using Google’s NLP API) that's indistinguishable from that of a human's, go for it.

Rand Fishkin just did after that burn, lol.

Personally I think automated content is a force in terms of scale but not in terms of quality. Sure, you can post 80 bazillion pages or snippets while you walk to the fridge and profit/rank/whatever but when you start publishing stuff in my wheelhouse you won't outrank me, I'm human, I'll step up my game as well.

The day the machine can write better content than humans is the day we can retire our sites to some kind of internet museum. If I had to take a wild guess the machines can write better than "some" humans but Google recognizes that even someone who may not "speeel rite" can offer value a machine might not be able to. I'd be surprised if Google cared for anything on our pages besides titles and the feeling we're left with... the rest is just a travel map between important points.

[cnvi]

Sorry Gary but the more Google and it’s followers mislead us, the digger the hole your shoveling. What once was a great search engine has now turned into a “we own page one” illegal monopoly. How about Google lets us decide who we will and will not link to and stop the nofollow disavow crap that the vast majority of website operators know nothing about. When will the World Wide Web be exactly that again and not the world wide evil google.

[tangor]

Roll back to the wild west days? Not sure I'd like that. Some order is required and g made a good start ... only problem is bean counters and (lately) ideologues and stepped all over it. Without some kind of oversight black hat will rule and we would all either lose, or become bandits, too, just to survive.

[NickMNS]

@JS_Harris

The day the machine can write better content than humans is the day we can retire our sites to some kind of internet museum.

I think you've got the cause and effect inverted. It's the day when humans will write content such that machines can understand it. That day has arrived.

Take some topic, the search engine bots will prefer (rank higher) an easy to understand text about that topic over some complex text that is more refined but from which the bot can only parse very basic meaning. Result, the human writers will dumb down their content for the bots. With better understanding (from dumber text), the nlp algo's will be able to produce continuously higher quality content (moving from grade 1 to grade 2) until both human and algos converge to the same basic grade 3 level of English, the least common denominator. All the while the engineers at Google will be patting themselves on the back for how accurate and how well their bots perform at analyzing text, where in reality all the bots are doing is filtering. And, that filtering is actually done by the humans that by guessing at what to feed the bot. Note the accelerating degree of the homogeneity of web content.

[RedBar]

It's the day when humans will write content such that machines can understand it. That day has arrived.

Bingo! This has been happening for quite some time and for anyone ever wondering why their site fell against another site, comparing their structured display usually gave the answer.

That's not to say in-depth pages will not succeed, quite simply one has to analyse their layouts more.

[JS_Harris]

Take some topic, the search engine bots will prefer (rank higher) an easy to understand text about that topic over some complex text that is more refined but from which the bot can only parse very basic meaning.

wut u mean? j/k

To be honest the ai understanding and leaning towards simpler copy doesn't worry me as much as humans being offended by everythign but the arbitrarily approved or socially acceptable bias fad of the moment. We can stop writing altogether if it's from a script, read the script and everything leans to it.

On the bright side I probably read only 10% of the sites I used to, bias(in any direction) is a turnoff.

[JS_Harris]

It's an interesting thread, worth reading. Google has put more engineers on the task of helping re-focus webmasters and Gary is not ultra-fond of the SEO industry in general and it's "made up terms".