Not All Google Search Algorithms Are Suitable For Machine Learning

While Microsoft seems to have gone all in on machine learning in Bing, Google has seem not to. Even when we are about to skip over to 2021, John Mueller of Google said on Twitter "there are always so many algorithms in play; some are more suitable for machine learning than others."

Historically, Google has said that machine learning won't take over the algorithm. That was in 2017 and then in 2019, Google said it applies machine learning to specific problems and not to everything under the sun. They do use it in many areas from crawling and indexing, or parts of it, to featured snippets, BERT, RankBrain, and as Martin Splitt recently said maybe for passage indexing purposes and so much more.

But as John Mueller said in that tweet to Glenn Gabe, machine learning may be "more suitable" in some areas of some algorithms than others. John added that "suitability also requires room to remove bias, allow debugging, allow critical corrections, etc. -- in addition to delivering better results." Like Google said in 2017, Google doesn't want to slap ML or AI on things and not be able to know what is going on. Not to be able to debug issues. Not to be able to figure out why something does what it does.

Here are those tweets:

There are always so many algorithms in play; some are more suitable for ML than others. Suitability also requires room to remove bias, allow debugging, allow critical corrections, etc. -- in addition to delivering better results.

— 🍌 John 🍌 (@JohnMu) December 21, 2020

I appreciate the care in making this transition from the existing codebase to turning bits of it towards ML.

It almost feels like a multi-stage site migration, but instead of moving bits of your site, it's pieces of the search algorithm 😂

— Victor Pan 🇹🇼🇺🇲 (@victorpan) December 21, 2020

It seems to me that Google, at least in Google Search, uses machine learning and AI techniques in a more hands-on, selective and careful way. But for how long. I suspect Google's use of machine learning in 2017 versus 2021 is much much different.

Forum discussion at Twitter.