With the end of third-party cookies rapidly approaching, it's time to look ahead. How will we handle attribution reporting in the future?
What possibilities do we, as marketing analysts, actually have?
Jente De Ridder, co-founder of Stitchd, delves into the various approaches you can adopt, from Marketing Mix Modeling to Data Clean Rooms and conversion modeling.
5. 5
Is our current attribution that good?
#SPWK
• Unmeasurable interaction points
• Not all conversions are measured
• Does not capture “buying intent”
• Cross-device usage
7. 7
What money is wasted on ads?
<Hoe
werkt
campagne
tracking
binnen
een
digitale
context?>
8. 8
Measuring marketing effectiveness
#SPWK
How attribution contribute to business goals
Customer insights
Attribution provides insights
on your customers by
uncovering at which point
in the customer journey
they are most influenced.
Marketing ROI
Attribution enables you to
understand where your
advertising budget
produces the highest
incrementality per Euro
spent.
9. 9
Attribution modeling
#SPWK
Two main types
Touch-based
• Bottom-up
• Focus on individual customer journey
• Need for detailed individual-level data
• Distributes credit for conversions among
touchpoints based on position or time
decay.
Correlation-based
• Top-down
• Macro-level overview
• Need for aggregated data
• Utilizes statistical techniques to quantify
the impact of marketing on sales or other
KPIs
11. 11
Touch-based attribution models
#SPWK
Ignores direct traffic and attributes 100% of the conversion value to the last channel that the customer clicked
through before converting.
Cross-Channel Last Click
Cross-Channel Position Based
Cross-Channel First Click
Ads-Preferred Last Click
Cross-Channel Linear
Cross-Channel Time Decay
Gives all credit for the conversion to the first channel that a customer clicked before converting.
Attributes 100% of the conversion value to the last Google Ads channel that the customer clicked through before
converting. If there is no Google Ads click in the path the attribution model falls back to Cross-channel last click.
Distributes the credit for the conversion equally across all the channels a customer clicked before converting.
Gives more credit to the touchpoints that happened closer in time to the conversion. Credit is distributed using a 7-day
half-life: a click 8 days before a conversion gets half as much credit as a click 1 day before a conversion.
Attributes 40% credit to the first and last interaction, and the remaining 20% credit is distributed evenly to the
middle interactions.
Machine learning algorithms to evaluate both converting and non-converting paths. The resulting Data-
driven model learns how different touchpoints impact conversion outcomes.
Data-Driven attribution
12. 12
Shapley (game theory)
In cooperative situations, something known as the Shapley value is used to fairly
distribute credit to each individual player/participant.
#SPWK
Image: https://clearcode.cc/blog/game-theory-attribution
13. 13
Shapley (game theory)
It takes into account the order in which each touchpoint occurs and assigns different credit
for different path positions.
#SPWK
Image: https://clearcode.cc/blog/game-theory-attribution
14. 14
Shapley (game theory)
The average value of all those possible combinations is the “Shapley value” for that
player/participant.
#SPWK
Image: https://clearcode.cc/blog/game-theory-attribution
15. 15
Markov chains
Looks at the customer journey as a sequence of states (touchpoints) that transition from
one to another with certain probabilities.
#SPWK
Image: https://adequate.digital/en/markov-chain-attribution-
modeling-complete-guide/
16. 16
Markov chains
It calculates the probability of conversion given the removal of a particular channel.
#SPWK
Image: https://adequate.digital/en/markov-chain-attribution-
modeling-complete-guide/
17. 17
Markov chains
Removal effect: the difference in conversion probabilities with and without a channel
provides an estimate of that channel's importance.
#SPWK
Image: https://adequate.digital/en/markov-chain-attribution-
modeling-complete-guide/
18. 18
Touch-based attribution
#SPWK
Typical outputs
Channel First touch Last touch Shapley Markov
Organic Search 100 90 105.0 95
Paid Search 150 160 155.0 152
Email 120 110 115.0 118
Paid Social 130 140 135.0 132
Organic Social 110 105 107.5 108
Absolute impact. The
conversions obtained
via this channel.
Estimated impact.
Higher value = greater
impact on driving
conversions.
19. 19
Touch-based attribution
#SPWK
Pros and cons
Pros
• Tracking on the “user level” and insights
into the customer journey
• Understand the role that each channel
plays within the customer journey
(awareness vs closing)
Cons
• Ignores person’s history with the brand
• Rules don’t fully reflect complexity
• Hard to track every touchpoint
• Cross-device
• Cookie lifetime
20. 20
Favors “directly measurable” channels
Compared to MMM, touch-based
attribution models overestimate
the impact of paid search by 2x.
And they underestimate brand
advertising by 10x.
#SPWK
23. 23
Interaction effects
#SPWK
Macro-economic factors
Sales
Media investments
(marketing channels)
Top of mind
(brand awareness)
MMM is much more than correlation between “ad budget” and “sales”. Many possible
interaction effects need to be considered when looking for true causation.
24. 24
Point of diminishing returns
After a certain point, also called the saturation point, each dollar you spend has a relatively
lower impact.
#SPWK
25. 25
Adstock effect
Or carryover effect. A past ad could influence present sales. Not all people that see your ad will buy
the same day, some will buy your product the day after, or the week after.
#SPWK
26. 26
Marketing Mix Modeling
#SPWK
Typical outputs
53%
7%
6%
4%
2%
-28%
Base + Distribution
TV
Promotions
Digital spend
Print
Price/
Competition
Contribution chart Response curves
Linear = constant returns
Concave = diminishing returns
Convex = exponential growth
S-shaped = low impact in beginning, nice
returns once traction, with an eventual saturation
point
27. 27
Marketing Mix modeling
#SPWK
Pros and cons
Pros
• No need for individual user level data
• Can incorporate complexity of multiple
variables influencing sales.
• Customizable to specific business context.
Cons
• Need for a lot of historical data
• Need for variation in the data
• Actively experiment!
• Less granularity
• Slower
• More complex to interpret
28. 28
Post Purchase Surveys help improve MMM
Surveys help to get an idea of
”unmeasurable channels” like
word-of-mouth.
#SPWK
29. 29
Split tests help improve MMM
A typical example is to run
different creatives / campaigns /
channels per region.
Example: would there have
been a difference in how
#SPWK was promoted in the
Netherlands vs Belgium?
#SPWK
37. 37
Data Clean rooms
#SPWK
1st
Party data
2nd
Party data
Data Clean Room
Stitch and share data in a privacy
compliant way with double-blind joins
Your customer data Customer data
from another organisation
Audience insights Ad Targeting
38. 38
Data Clean room
#SPWK
1st
Party data
Publisher
data
Example Use Case
Samsonite wants to understand how
many sales can be attributed to
impressions of its recent banner
campaign.
39. What I hope you will remember
39
Key Take Aways
42. “
” 42
Adopt an experimentation mindset and
dare to take (calculated) risks.
43. “
” 43
Before you start: think about what you want
to achieve. What is the business result
you’re aiming for?
44. Let’s get Stitchd!
Any questions? Get in touch.
www.stitchd.be | jente.deridder@stitchd.be
Notes de l'éditeur
Bad data => bad decisions.
How to handle attribution reporting?
Are cookies the only problem with attribution?
Skepticism around the value of attribution reports, regardless of the depreciation of cookies.
Most known from Google Analytics.
What do all these heuristics have in common? Well… we made them up
Rule based instead of measuring the true intent of the user. Oversimplification: Position / time are the only factors taken into account.
The data-driven model used within Google Analytics.
You first start by identifying each player’s contribution when they play individually, when 2 play together, and when all 3 play together.
Then, you need to consider all possible orders and calculate their marginal value – e.g. what value does each player add when player A enters the game first, followed by player B, and then player C.
Now that we have calculated each player’s marginal value across all 6 possible order combinations, we now need to add them up and work out the Shapley value (i.e. the average) for each player
It calculates the probability of conversion given the removal of a particular channel, allowing marketers to see how likely it is that any given channel influences the path to conversion. The difference in conversion probabilities with and without a channel provides an estimate of that channel's importance.
It calculates the probability of conversion given the removal of a particular channel, allowing marketers to see how likely it is that any given channel influences the path to conversion. The difference in conversion probabilities with and without a channel provides an estimate of that channel's importance.
It calculates the probability of conversion given the removal of a particular channel, allowing marketers to see how likely it is that any given channel influences the path to conversion. The difference in conversion probabilities with and without a channel provides an estimate of that channel's importance.
Shapley: The number is a measure of its contribution to achieving conversions. A higher Shapley Value indicates a greater impact on driving conversions, while a lower value suggests a lesser impact.
MMM is designed to measure the impact of advertising and promotions across channels while controlling for external factors outside a brand’s control.
Better suited to grasp the complexity that comes with attribution than the touch-based models.
The highest-intent audience (the low hanging fruit) has already been converted in the first few runs of the ad
Expansion requires reaching less relevant or more expensive audiences to reach, harming average performance
Ad fatigue sets in for audiences that are repeatedly exposed to your ads, hence they stop paying attention
Background (mathematical): https://www.toppr.com/guides/business-economics/theory-of-production-and-cost/the-law-of-diminishing-returns/
Up until when will people buy from you because of the impression/interaction they had with an ad?
Out of the 100 units sold, 53 units would be sold even if the marketer doesn’t invest in any form of advertisement. Basically, these 53 units are sold because of the brand’s equity in the market and the awareness it had created in the customer’s mind in the past. Similarly, 7 units are sold due to TV advertisements and 3 units are sold due to Consumer promotions and BTL promotions each.
Interpreting price correctly is the key to understand the MMM contributions. Many times, people are misled due to the negative sign present when it comes to representing price. Notice that, when we sum up the contributions in the above chart with the negative sign on price, the sum is 44 not 100.
If we ignore the negative sign on the price, the contributions would sum up to 100. Since sales and price have a negative correlation for most of the brands (damn you Apple!), the price contribution is represented with a negative sign to show the quantity of loss in sales it can cause.
Here, negative sign of the price indicates that 28 units of sales was lost, due to increase in price. This is a notional concept which depicts that 28 additional units of sales could have been gained, had there been no increment in price.
The complexity of econometric models can also be a challenge, particularly for marketers who may not have a strong background in statistics or programming. It's essential to work with experienced analysts or data scientists who can help to develop and interpret these models correctly.
So, should our conclusion be to forget about touch-based attribution and go back to the correlation based analysis methods?
Offcourse, we have Google and Meta looking out for us…
Nothing that AI can’t solve.
Ad interactions and conversions are grouped in two groups. One group contains ad interactions that have a clear, observable link to a conversion. The other group contains ad interactions that don't have a clear, observable link to a conversion.
The observed conversions are divided into subgroups based on shared characteristics (a variety of dimensions, including location, time, and browser), and key metrics are calculated for each.
For example, conversions observed in the morning in France are found to have a certain conversion rate, whereas this rate may be different in the evening.
Those subgroups are used for sorting unobserved ad interactions and conversions.
Using the known conversion rates and other characteristics from the observed subgroups, machine learning links unobserved ad interactions and conversions, where appropriate. The observed and modeled conversions are then integrated into your conversion data to help you make informed decisions about ad performance reporting and fed into bidding.
A few year ago, Facebook and Google made available conversion lift tests. This feature makes it possible to conduct a controlled experiment and to measure the incremental value of marketing campaigns.
Conversion lift test creates a control group out of the pool of users that would have seen our ad (because we have won the auction) but instead of showing the ad, they show the next ad in the ranking.
With the security and access controls that data clean rooms provide, media companies and publishers can provide detailed reporting, and advertisers can track attribution more effectively.
Ads Data Hub from Google is a good example. But note that Big Query also offers DCR capabilities now.
Snapshot: we take a snapshot and start drawing conclusions from it. However, the context in which we operate is constantly changing. So that snapshot is always only partially useful to explain the future.
The end of the third party cookies does not change this. It will make our data even less trustworthy, but it also forces us to rethink what we were relying on.
Each model and approach has their own advantages and disadvanteges.
The value lays in understanding how different models compare and forming your conclusions based on that.
Lot’s of innovations happening. Try things out and learn on what works and doesn’t.
As a scorekeeper, to show the overall incremental impact of marketing investments on the business
As a forecaster, to predict the outcome that raising or lowering marketing budgets will have on marketing’s contribution to the overall budget
As a coach, to suggest shifts to current marketing investments that improve performance