The other day I was doing a presentation about machine learning applied to attribution models and the issues of standard attribution models:
- They are not based on a model that tries to represent a reality, with the higher level of certainty as possible, but on a completely discretionary decision (normally from a HIPPO, Highest Paid Person Opinion).
- Since they are not trying to represent the reality but to define a criteria to standarize the measurement, all the information that flows from them (Acquisition Cost, repayment period, etc) is wrong.
So basically, you will be making your marketing decisions based on wrong premises. Meaning that you can be nickel and dimed to death without even noticing it. The attendees were opening their eyes and at a point of the presentation THE question arose “why we are using standard attribution models all this time, when they don’t represent the reality and are that risky”. The only possible answer is, because Google told you so (and I would do it too if I would be Google). People love to simplify things even at a pretty odd extremes. Like using benchmarks to define the average conversion rate of an ecommerce site against another one, as if comparing different systems were possible. Or like in this case, defining an attribution model that defined based on whatever the heart told the HIPPO that day. “Let’s use a standard attribution model so we can compare each other in a better manner”. But, you shouldn’t care about comparing yourself with others but improving your media investment comparing your previous performance with your current one so you can generate a lift. Why should you care about winning others on a specific metric at a possible cost of losing money?
The answer is, when a big player leads the industry, the people tends to convert in a “standard” whatever thing they do. But the question is, “is Google chasing the same objective as you?”. Some people believe that Google is a tech company, so they don’t it as a vendor or a competitor but “someone” that improve their life (and I wont judge that). Google is a marketing company that sell services and advertising. It goes from vendor, to competitor to partner all the time. As a matter of fact I’m Google’s vendor, client, partner and competitor all the time.
They are also the biggest media intermediator in the world, and again, good for them. Their main objective is earning money as any other company in the world, and there’s no doubt that they are accomplishing that goal, chapeaux!
The point is, all the things they develop are done with the same north in mind, making more money. So the question is, should we relax and leave Google algorithms decide what’s better for us? I don’t think so. Some years ago people got mad when they realized that Google Analytics’ attribution model favored adwords over other referring channels. Of course Facebook and any other media companies in the world would do the same. Again, this is not a communist complaint about marketing companies, believe me, I’m way far from that. It’s not their fault, come one, you was trusting your budget to your vendor’s “black box” without even asking what was inside that. Don’t call me unfair to believe you were way too naïf.
But the attribution model is just the tip of the iceberg. That was the introduction to put you us all in the same page so we can analyze a very interesting white paper called “Ad Click Prediction: a View from the Trenches”, written by a Google Team on Machine Learning algorithms focused on the “massive-scale learning problem that is central to the multi-billion dollar online advertising industry”.
Again, take a careful read to the white paper tittle and the performance variable they are focused on. Correct, their main goal is ad click, so basically they are developing algorithms that improve the ad clicks. The more clicks, the more money for Google. Makes perfect sense if you are Google, but what if you are not? Well, the advertising value is based on its capacity of converting “people” into clients, today (purchase) or tomorrow (purchase intention). None of them are related to clicks, unless you believe that more clicks will always means more sales, which goes against to one of the main economic principles that rules the entire world of material things, the “Law of Diminishing Returns”.
The problem on using a model that is based on clicks when you are interested on earning money is that you can be in the third stage of this model losing money, but with a high CTR and keep pushing budget pedal to the metal. I got tired of finding companies in that stage…and believe me when I say that at that stage you are too late.
Your media buying algorithm should be based on sales conversions instead, because that’s what you are interested in. The problem is that google can’t measure your real sales, so your algorithm must be build on your end. We’ve already talked about that we shouldn’t be interested in external but in internal benchmarks, because every system (set of things interacting together with a common objective) is unique, so you want your algorithm to be optimized on predicting in a better manner based on YOUR specific reality.
So my suggestion is reading the paper, which is really interesting, just change clicks per sales conversions. They propose the well known, simple and efficient logistic regression (Yes, supervised machine learning since we are talking about predictive analytics)
with the following process.
Visit here if you want to learn more about the model, very recommended reading for the weekend.