By: Pavel Šíma, ROIVENUE™ CEO
[LinkedIn] [Twitter]

For a long time, I struggled to conceptually grasp different attribution models via their mathematical equations. So, I decided to devise simple, real-life metaphors for them. This way it shouldn’t take more than three minutes to finally understand all the frequently used attribution models in digital marketing and recall them with ease.

1) HEURISTIC (TOTALLY USELESS) ATTRIBUTION MODELS

A) MODELS WITH ONE SOURCE OF ATTRIBUTION

LAST CLICK

All the glory goes to the striker – even if he is totally lame. Everything was fought out by the rest of the team and he merely put his foot in the path of the shot.

FIRST CLICK
Who is to blame for all the bad deeds in my life? My mother of course. For she stood at the dawn of my existence.

These above models are as far from reality as the burger you get at the counter is from the one you see on the billboard.


B) MODELS WITH MULTIPLE SOURCES OF ATTRIBUTION

LINEAR

Sure, I am grateful to all the Chinese manufacturers who put the pieces together, the freighters who brought it across the ocean, and the retailers who sold it to me. However, most of the margin will go to Apple as they put their logo on the box.

With these models, at least all the credit doesn’t go to one person. The level of inaccuracy is still humongous, though.

Do you really think a human brain can decide what touch-points should be assigned which weight? Didn’t think so.

2) DATA-DRIVEN (ALGORITHMIC) ATTRIBUTION MODELS

They try to calculate the probability a certain channel contributed to a conversion.

SHAPLEY (ALSO KNOWN AS GAME THEORY)

We all have such a friend. When he shows up at a party it’s always a blast. Then one day, he doesn’t come and you are bored to death. A hunch tells you that he is a really important part of a good party. But how can you be sure?

Imagine, you went to 10,000 parties a year instead of ten, both with and without your mate. By the end of the evening you would rate all of them from 1–10 on a scale of awesomeness.

Then you would easily be able to calculate the probability that he’s the most important ingredient of a good party.

The same way, you could evaluate any guest (= marketing channel) at every party (= conversion path).

Shapley’s big problem is that it doesn’t take the order of guests/channels into account. Some are great party starters but nevertheless fall into a coma at 9 PM. Standing on your feet at 4AM is an entirely different discipline. You need a great mixture of all types of people to throw a legendary party.

MARKOV (ALSO KNOWN AS MARKOV CHAINS)

How is gossip spread in an organization? Sometimes you hear it directly from the source, sometimes it travels through six different people, while other times it gets stuck in a loop and never reaches your ears.

Michael knows about the latest scandal first and tells Monica about it over lunch. In the bathroom, Monica passes it on to Jane. Back in the office, Jane tells Michael about it again. After that, Michael goes directly to you.

If Michael is removed, everyone is pretty much screwed and will learn nothing. Monica and Jane are not all that important in comparison, Monica added Jane and Jane didn't add anyone. That’s why Michael is the most important informant.

How do you determine who’s the most valued informant (= marketing channel) in the long run? Try to send different employees for a long vacation and observe if the level of information passed on to you dropped off or not. If it does– the person on vacation is the one who likes to talk.

MARKOV CHAIN 1ST ORDER

It measures the influence of the movement of the information (= nudging people closer to conversion) to the very next person (= channel).


MARKOV CHAIN 2ND ORDER

It measures the influence of the movement of the information between the next two people because if, for instance, you often talked to Jamie who is mute and couldn’t pass the gossip on to anyone else, the first-order Markov would blame you. Not fair!

WHICH ATTRIBUTION MODEL IS THE BEST?

Our opinion at Roivenue is that it is the Markov model. You don’t necessarily have to trust us, though. Just looking at what different models say about your data is often an eye-opening experience. We've seen examples where you can find up to a 300% (!) difference in how the different models value contributions.

Choose for yourself which attribution model you want to place your trust in. Roivenue will allow you to compare all of them. Once you decide which is the best, you can switch the whole application and order all metrics to be recalculated with one click of a button.