How we teach AI to be a marketer?

Being a marketer I love marketers. These guys cope with so many problems in a company still they stay alive effective. However, it’s getting more and more difficult lately. Some of the things are changing too fast:

 

If it is impossible to run, why not to make someone else run for you? Maybe, AI?

  • Patterns of communication and marketing trends
  • Landscape of tools and channels
  • Amount of data, business collects

Look up there again. These things transform daily. But you are supposed to be up-to-date. Some of us rush and put titanic efforts to get data-driven. But to stay data-driven in the world, where the amount of data grows exponentially is crazy. And not being a data-driven means to lose competition and money. So you actually can’t do it and still can’t NOT to do it. It’s weird. But I came up with a simple idea. If it is impossible to run, why not to make someone else run for you? Maybe, AI?

This summer we started off with an ambitious plan – to create an AI marketer, that will outperform human beings on this position within 2 years.

No more or less, than a brutal AI called Captain Growth, who can do all the dirty online marketing job for you. Much better than human, for less money. Stressless. Sleepless. Having a single aim in his artificial mind: to grow your business.

A quarter passed. And we have the first result.

The project made us ask ourselves a simple yet not obvious question: who is the digital marketer and what makes his work so important and difficult to automate.

On the one hand, he is a person who rules big ad budgets. On the other hand, he is the main guy in the team, who cares about how and where to get customers. Moreover, he is intelligent enough to understand the endless list of channels and tools to work with.

Started there, then we generalized all the marketer’s routine and ended up with such a typical loop of his functions:

  1. Analyze what’s going on, qualify the situation.
  2. Make decision, where to put efforts and what to do.
  3. Implement decision.
  4. Measure effectiveness.

Behind the scenes there is the most crucial part of this job: continious learning. You’re never too smart to be a marketer. Things keep changing every day from Google algorithm updates to new communication forms and formats.

We decided to develop our AI step-by-step, from number 1 to number 4. With learning as a general approach for every stage.

The first task: to analyze, what’s going on. Completed.

To answer the question ‘What’s going on’ and qualify a situation somehow we basically need just 3 things:

  • Actual data
  • Understanding, what’s good
  • Understanding, what’s bad

We decided to develop our AI step-by-step, from number 1 to number 4. With learning as a general approach for every stage.

Remember your last dive into Google Analytics. You spent hours going through charts and dashboards. But what were you looking for? Some call this ‘insights’. But in simple words, you were just trying to find good and bad things. Or, if nothing found, to define a situation as normal, with anything special happening. Now imagine, that a machine has to do the same.
Data is not hard to get. We set up API connections with Facebook Ads, Google Analytics and Google Adwords to retrieve data from our customers’ accounts. Some other big vendors are on our list to build integration this year. What’s next?

Understanding. Look at the chart below. Do you think the last day is good or bad?

v

You are the 1 of 20 if your answer is ‘Good’. Congrats, that is the truth. It is Sunday when sales usually go down, and it is the 10th day of an ad campaign when ctr usually goes down. But this ad showed just a slight decrease. That is not typical, that is a much better behavior than usual. And that is not obvious.

Here is how this chart looks, after we excluded a seasonality and a trend from it (the green line):

Now you see. Without understanding what’s good and bad, it is impossible to explain data. It leads to wrong decisions. And money loss.

So our marketing AI started with looking for this 2 answers: what’s good and what’s bad. To do that we currently use:

  1. Machine learning to understand what’s normal and what’s not from historical data.Looking at hundreds of businesses, we see what is usual for them and what is not. This helps a machine classify any new situation as rather normal or not, depending on what it saw previously. Pretty much like a real marketer does.
  2. Time-series prediction to fill the empty spaces in your data and bring more accurate result.Data usually has some noise, anomalies, and gaps. All three reduce an accuracy of any algorithm. So we clean data carefully before working with it. The most complex part here is to fill in the gaps. We predict missing values using all the other data we have and applying such factors as a user’s business industry, seasonality, metric and dimensions specialties.
  3. Outlier detection to find segments and datapoints, that are unusual: either good or bad.Outliers are those insights we are looking for.To define them, we need to care about statistical significance, level of unusuality, reasons, that might influence the metric. We currently can define seasonality, trends, manual actions (such as turning ads on or off) to explain outliers and find real insights behind them.
  4. Entropy-based algorithm, which defines, where exactly (on which level) the good or bad thing happened. Ok, we’ve detected a ‘good’ data point. It is about your mobile users performing 30% better on a landing page X on Tuesdays. Is it because of Tuesday? Or the reason is mobile? Maybe, all three of dimensions matter? Or just 2 of them? Our algorithm can define the exact combination of factors, that drive insight. We use entropy to come up with the optimal insight localization. Due to this you may take action exactly on the level, where problem or opportunity is.
  5. Scoring algorithm to define, how important this thing is for your business.Our algorithm is powerful. And it finds tonnes of insights every day. Seriously, we’ve detected 300 anomalies in our client’s Facebook ads for just one day. Not all of them matter. So we score insights, to show you only those things, that are significant, actionable and relevant to your business. In fact, users train Captain by liking/disliking insights in their feed. We use this reaction to continuously teach AI on what is valuable to a user.

Everything above is 100% data-driven. The approach we use is not based on hundreds of rules, created according to someone’s outdated experience. Instead, the machine learns itself by viewing data and experiencing more and more cases. Captain Growth extracts a meaning from your data and brings it to you as a feed. Want to test it on your data? Try it for free now.

( 2 votes, average: 4.50 out of 5 )
Loading...

Would like to know how
we create world's first AI marketer?

Subscribe to our blog. Mistakes, machine learning and rock-n-roll - all truth here.