The choice of the right attribution model is a challenge for most advertisers. We all know the limitation of Last Click attribution model. We will illustrate the limitations of other, more comprehensive pre-defined attribution models such as Time Decay and Position Based or Data Driven. We will share a solution to implement a Markov Chains based approach and will introduce an approach to test a new attribution model.
3 Key Learnings
Overview of attribution models and their limitations
How to implement a Markov Chains
How to test the impact of an attribution model
Brand experience Peoria City Soccer Presentation.pdf
HeroConf London 2019 - Why your attribution model sucks how to step beyond data-driven models with a markov model approach
1. Why Your Attribution Model Sucks
How to step beyond Data-Driven Models with a Markov Model Approach
Gianluca Binelli | @ktzstyle
Most Innovative Presentation
HeroConf 2019
2. Session Outcomes
● Overview of attribution models and their limitations
● How to implement a Markov Chains based model
● How to test the impact of an attribution model
4. What do we do?
Scientific Performance
Marketing Agency
5. Who am I?
● 14y experience in online advertising
● 6.5y at Google between Dublin, NYC & London
● Managed Quarterly $XXM in advertising for
Google as part of the SEM in-house team that
promotes Google’s products (in FB, Bing, Google,
Linkedin etc)
● Advisor for Google’s own equity fund Capital G
22. Let’s assume tracking is working
● Tags/Pixels
● UTMs
● Single source of truth
○ Google Marketing Platform (AKA DoubleClick)
○ Adobe Analytics
○ Facebook Attribution
○ Google Analytics
30. What’s a chain?
● Let’s assume we have 2 keywords
● These means we have 4 states for each chains:
○ START
○ "OUR BRAND"
○ Advertise Online
○ CONVERSION
32. Once we have all the link
counts, we can compute
the transition
probability ( = the
chance to go to a certain
state starting from a
certain point )
edge Link count
Transition
probability
START > "OUR BRAND" 3 3/3
START > Advertise Online 0 0
TOT START 3
"OUR BRAND" > "OUR BRAND" 2 2/8
"OUR BRAND" > Advertise Online 3 3/8
"OUR BRAND" > CONVERSION 3 3/8
TOT A 8
Advertise Online > "OUR BRAND" 3 3/5
Advertise Online > Advertise Online 2 2/5
Advertise Online > CONVERSION 0 0
TOT B 5
What’s a chain?
33. Let’s draw something!
It is worth noticing that once
we are in A and in B we
could end up in A and B
again.
START
"OUR
BRAND"
Advertise
Online
CONVERSION
100%
25%
60%
37.5%
40%
37.5%
36. To assess the impact of a keyword we assume it is
gone
● Importance of keyword "OUR BRAND" = the change in conversion rate if
keyword "OUR BRAND" is dropped from the Graph
● or in other terms if keyword "OUR BRAND" becomes a NULL state. A NULL
state is an absorbing state so if one reaches this STATE can’t move on.
42. START
"OUR
BRAND"
NULL
CONVERSION
100%
50%
50%
Which means
As we can see, once we
removed kw Advertise
Online, the chances to
convert fell to .5.
Therefore the importance of
kw Advertise Online is the
difference between previous
conversion rate (1) and the
chances to convert if we
remove Advertise Online
(.5), i.e. 0.5.
43. ● Once we have all the importance weights for all the channels/keywords we
can finally compute the number of conversions weighted by the
importance of our channels/ keywords.
● We take the previous number of conversions (3 in our example) and we
weight them according to the importance weight:
○ 1 / (1+.5) for kw "OUR BRAND" → 3*(⅔) → 2 conversions
○ .5 / (1+.5) for kw Advertise Online → 3*(⅓) → 1 conversion
To summarize
57. Science
● Control & Treatment
○ Split target location in smaller locations
○ Split 50/50
● Dependent Variable
○ Existing attribution model in one half of locations
○ Markov in other half of locations
● Independent Variable
○ Conversions/CPA
60. Dependant Variable
Campaign Targeted Locations Attribution model
Awesome Campaign 1 Random locations group 1 Last Click
Awesome Campaign 1 Test Random locations group 2 Markov
Duplicate every campaign