Balancing spend and developing strategy across channels like paid and organic search, display and social is one of the biggest challenges in digital marketing.
In the real world, attribution modelling often boils down to choosing a model and seeing whether we like the results it gives. But this is hardly scientific. So what would a data-driven process to defining and assessing a cross-channel attribution model look like?
2. Harvest by numbers
13 years old
40 digital specialists
£25+ million billings
7th ranked Drum Digital 100
5th ranked digital media agency
Econsultancy Top 100
3. Which Attribution Model?
Last Click First Click Linear
Position Based Time Decay Custom Attribution
6. Attribution Study Toolset
• Web Analytics solution
Google Analytics / Omniture / Coremetrics, etc.
• Raw Clickstream data:
GA MultiChannel Funnel API / Omniture Data Warehouse, etc.
• Data Transformation / Manipulation:
Database or Statistics solution: R / SAS / SQL / even Excel
• Data Visualisation
Tableau / Visokio, etc.
• Attribution Modelling
GA built-in, third-party DMP solution, own bespoke database
7. Funnel
Stage
Offline
Reach
Returning
User
Demographics
Campaign
Targeting
Time Lag
Engagement
Time
&
Location
Click Position
Visit
Duration
Conversion
Rate
Marketing
Channel
Attribution
Model
The Attribution
Landscape
8. Arriving at an Attribution Model
Examine Marketing Channels to identify their role in
the user journey and the relationships between them.
Examine traffic signals to understand which
activities should be receive greater rewards.
Examine conversion habits to find key influencers
and inhibitors and weigh them accordingly.
9. • An astounding 69% of users have a single-visit
journey, and of the remaining returning
users 17% will visit on the same day.
• This only leaves a minority of 14% who have
multiple visits over a length of time.
• With less than a third of users visiting in
multiple sessions Last Click does not sound
like a bad idea.
• What about after we examine traffic and
conversion habits in more depth?
User Traffic Behaviour
10. Returning User Behaviour
• Users are not very likely to re-engage repeatedly
with the site as time passes.
• If they make return visits they will do so in short
intervals.
• Conversions and the CR% follow the same trend,
so users who eventually come back are not
necessarily more eager or loyal.
• The slow decline of the CR% over 5 days implies
that the Day 0 peak is Immediate Intent rather
than Direct Response to Advertising.
• Since the CR% is not higher for lagged users that
click does not seem to have worked harder…
Days Since Last Session
11. Returning Visits by Channel
• Certain channels however are effective at
getting returning traffic, which could still
convert.
• Out of the total Conversions 32% are made
on return visits, but 24% actually return and
apply on Day 0.
• The one dimensional channel behaviour
points towards Last Click, but that would
ignore the initial source of that 24%.
• First Click effectively covers 91% of
conversions (Day 0), but does no justice to
all the Reactivation channels.
12. The first step
• Combine the two and use Position Based instead.
• We already know the First Interaction is the most crucial converter, but a 3rd of conversions are on a
return visit, so the Last Interaction also deserves a good share of the credit.
• In-between interactions are not common but where they are present they should still receive a small
amount of credit among them.
13. Spread of
Returning Users
Immediate Visits :
66% Converted
11% Non Converted
Rapid successive
Visits :
7% Converted
16% Non Converted
Short Term Visits :
8% Converted
4% Non Converted
Medium Term
Visits :
10% Converted
20% Non Converted
Long Term Visits :
7% Converted
21% Non Converted
14. User Visit analysis
• The majority of conversions happen within 14 days and quickly tail off between 15 and 30 days.
Therefore a Lookback Window of 30 days will sufficiently cover the user base.
• The second biggest cluster is the Medium Term one, meaning that beyond the users that convert right away
there is still an considerable Returning audience.
• Outliers not included in the main clusters make up 2-3% of Converted users and 28% of the NonConverted.
That large percentage of users that have returned without converting could be targeted with a Reactivation
campaign and improve customer acquisition.
• Visitors in the long-tail segments are much less likely to be converters, even though they do re-engage
after a long time period.
15. • Are there any channels that often re-appear
throughout the user journey?
• Are there any common sequence groups?
• What about a channel’s role in Assisted
conversions?
Channel Paths
• 62% of conversions are Single-Path
• 15.5% are Dual-Path
• 22% are Multi-Path
• The vast majority of journeys are Single-Path
and therefore irrelevant to MCF analysis.
• The remaining 37% do not show any patterns of
magnitude.
17. Channel Paths: Touchpoint contributions
Channel Assisted
Last
Click
Assisted /
Last Click
Conversions Conversions %
Multi-Path
Contributions*
Multi-Path
Contribution %
Paid Search 2,871 5,354 0.54 6804 45.54% 3226 21.59%
Direct 1,892 3,557 0.53 4268 28.57% 2955 19.78%
Organic Search 1,750 3,022 0.58 4272 28.59% 2170 14.52%
Affiliate 958 2,283 0.42 2952 19.76% 1155 7.73%
Referral 186 316 0.59 476 3.19% 276 1.85%
Other 104 213 0.49 296 1.98% 127 0.85%
Display 73 14 5.21 85 0.57% 78 0.52%
Social Network 68 118 0.58 184 1.23% 82 0.55%
Email 42 64 0.66 98 0.66% 59 0.39%
Total 7,944 14,941 0.53 - - - -
* Multi-Path Contributions is the number of conversions for each multi-step user journey that the Channel has been present in.
18. Attribution Model Conclusions & Bonus Insights
Engagement
Base Model
Position based
Priorities:
First Click >> Last Click > Middle Click
Attribution
Window
No longer than 30 days necessary
14 days already covers 90+ % of
conversions
Time
parameters
Down-weigh Day 0 conversions as channels
only make minimal effort to win those
Add Time Lag modifiers to reward channels
that quickly generate Return traffic
User
Personae
As a more advanced option we could
assign custom credit to each of the above
Audience Segments based on their
propensity to convert
Assists
Account for Display, Email, Referral assist
contributions by up-crediting the channels
Channel
Sequences
Up-weigh Channel pairs that have consistently
displayed mutually assistive gains
Retargeting
About ~7% of users returned after 2 weeks and
within 3 months.
This is a good percentage to plan a small,
focused retargeting campaign around.
Segmentation
Can create multiple versions of a model by
applying further filters/segments on results and
adjusting weights
Or create custom reporting/target audiences
Additional credit or segments based on
channels that result in above-average user
engagement according to site metrics
19. Possible Signals
User
Behaviour
Campaign /
Source
Conversions Marketing
• Pages / Session
• Session Duration
• Time of Visit
• Sequence of Visit
• Content
Consumption
• Engagement
• Channel
• Organic vs Paid
• Prospecting /
Remarketing
• Position of Visit
• Time Lag
• Macro & Micro
Conversions
• Loyalty
• AOV
• Assists
• ROI
• Business
Objectives
• Trends
20. Methodology Suggestions
• If you try to get to know your site, its channels and its users the data will make suggestions as to which type
of Attribution Model is suitable.
This won’t be a straightforward answer, but combining that with overall marketing plans and targets will
make the next steps much more clear.
• During this process of analysing user behaviour and dissecting traffic trends you will also uncover many
other useful hints that can help fine-tune marketing activity:
Channel Role & Interplay, Engagement Thresholds & Time Window, User Clustering, Receptiveness and
(inverse) Maturity Period.
• As with everything Web Analytics this is also a continuous, ongoing process. Even if the base of it remains
the same, an Attribution Model should be re-examined and calibrated frequently to reflect the most recent
changes in user and business behaviour.
• If precise enough, you could create a set of seasonal or occasion-specific models that are recycled every
year with minor updates and modifications.
Regarding GA: Model Comparison doesn’t allow all metrics/dimensions – pull standard reports and Clickstream data / MCF API (standard reports much assign everything somewhere) together in a data analysis / visualisation tool, add your coefficients and work out a Model there, doesn’t have to be built in GA/Omniture/etc
Also combine Offline/ Third party data sets
CR is 3.12% on same-day (return visits), fluctuates around 2.7% till day 5 then sharply drops
86% first day and 69% single visit, total 7.6% within 1 week, 2.6% within 2 weeks, 4% (2w, 40days]
2.45% with 1 day lag, 1.4% with 2 days
The conversion rate does not decline very sharply, meaning that this Behaviour is not Immediate/Direct response, but Immediate Intent
Common industry behaviour, agrees with benchmarks
Channel – Count of Visit Correlation: landing pages and retarget outlying users if certain channels are more effective at pulling back (and up-weight in attribution)
Why not Time Decay?
You could, but conversions do not decline linearly. The volume of Day 0 would dwarf any long-tail interactions.
If you can adjust the model to begin TimeDecay from Day 1 (instead of just setting a half-life rate throughout), it would be feasible but would require some solid statistics.
In this case we (more or less assume) an irregular decline rate as time progresses
Excludes Day 0 – Visit 1 to avoid skewing bubbles
Green: D0-V8, D2-V3 (freebie)
Red: D0-V(8 15], D[1-3]-V[5-15]
Yellow: D2-V[3-4], D[3-5],V[2-4]
White: D[5-13]V[2-8]
Black: D[14-31]V[2-5]
Remaining outliers: Converted 2-3%, NonConverted ~28%
Red+Cyan (73) imply FirstClick, Black and White imply FirstClick (17%) and Yellow Position (8)
Life in a MultiChannel world (why we can’t realistically use Single Click)
MPC can show which Channel is present multiple times in a long journey and therefore has a lot of reach and influence over conversions.
A Channel that is quite Assist-y and also has a large Contribution% in the MCF would lose out from a position-based model (if it is working harder than other Mid channels and someone else wins/steals the FC and LC) and should be manually up-credited.
Pick a few significant ones! Use statistics confidence intervals and check residuals if possible.
Compare models to make sure there is a significant and intuitive difference from LC and alternatives – nothing extreme or unexpected.
Weights are going to be defined relatively, so all that’s necessary is to establish a baseline (most likely Direct) and make adjustments from there.