About two years ago GRO.TEAM Founder Rorie Devine had a moment of clarity and thought “Rather than try to reverse engineer the complex and constantly changing Google search results algorithm why not train a Neural Network on its live search results data instead?”
So that is exactly what he did and after eighteen months of training and iteration the Neural Network is now producing some startlingly accurate SEO results…
What Is An Artificial Neural Network (ANN)?
An Artificial Neural Network (ANN) is an AI machine learning computer software model inspired by the neuron structure of the human brain. The Neural Network is “trained” by adjusting the weighting between the nodes to produce known outputs from specific input sets.
In the case of SEO at any time we know the outputs (search results position #1, #2 and so on) produced by a known input (the input search phrase e.g. “growth hacking london”) so modelling the Google search algorithm is achieved by training the Neural Network to produce the same search results for the same input phrase as the Google algorithm.
How Hard Can It Be?!?
Well, after 18 months of hard work I think it’s fair to say it’s quite hard actually..
The Beauty Of A Neural Network
The reason why Neural Networks are such a powerful form of Artificial Intelligence (AI) / Machine Learning is that they can “learn” and “know” things that nobody else knows, not even the creator of the Neural Network.
When training the GRO.TEAM Neural Network we can measure its output accuracy (in producing the same search results as Google) without actually knowing why it produced those results.
That is fine though, in this context we’re not trying to decode or explain the Google search algorithm, we’re attempting to be able to reliably recreate its results.
At University I had to wait days to train a simple three layer multilayer perceptron Neural Network on my x86 PC but with modern cloud compute power a Neural Network can be trained in only minutes/hours.
What Have We Learnt?
After 18 months of blood sweat and tears I think it’s fair to say that Google uses a lot of signals within its search results algorithm.
To get the kind of result accuracy we wanted we have ended up using over 1,000 input signals in the latest iteration of the GRO.TEAM Neural Network.
Can This Approach Be Used For Other Things?
My third year project at university was “The Use Of Neural Networks In Financial Time Series Prediction” and even with the limited compute power available to me at the time I was achieving good results predicting time series data such unemployment rates v time.
I also worked with an Isle Of Man based Asset Management company to try to predict intraday Cable (GBP/USD exchange rates). I achieved very limited success on this data, it was just too noisy.
In my opinion if the system producing the data is relatively deterministic (in that it will return the same output given the same inputs most of the time) a Neural Network will be able to model the system and predict future series values with surprising accuracy.
What SEO Results Have Been Achieved With The GRO.TEAM Neural Network?
The accuracy of GRO.TEAM’s Neural Network in predicting Google’s top ten search results has been hugely satisfying and very useful in helping businesses shape their search engine presence in the way Google seems to recognise, value and reward with higher search engine rankings.
I guess that’s not surprising though because the Google search system is relatively (but definitely not 100%) deterministic. Obviously Google is tweaking its results all the time these days so the GRO.TEAM Neural Network is undergoing continuous training cycles.
Can Anyone Use The GRO.TEAM Neural Network?
Sure, GRO.TEAM makes its Neural Network available to clients as part of its Guerilla SEO Package which is designed to help businesses achieve search engine results position #1 (roughly 25% clicks) or #2 (roughly 15% clicks) in a straight up SEO competition against well established and well funded incumbents.