Adobe Summit - Data-Driven Marketing Attribution
- 1. ©2014 Adometry, Inc. All Rights Reserved. 1#AdobeSummit
March 24-28, 2014 | Salt Lake City, UT
Learn more at summit.adobe.com
1
- 2. ©2014 Adometry, Inc. All Rights Reserved. 2#AdobeSummit
DATA-DRIVEN
MARKETING
ATTRIBUTION
making the leap to
- 3. ©2014 Adometry, Inc. All Rights Reserved. 3#AdobeSummit
Paul Pellman
CEO
Lewis Broadnax
Executive Director
Web Sales & Marketing
@ppellman
A Bit About Today’s Presenters
- 4. ©2014 Adometry, Inc. All Rights Reserved. 4#AdobeSummit
marketers need help answering
TWO QUESTIONS…
“How is my
business doing?”
- 5. ©2014 Adometry, Inc. All Rights Reserved. 5#AdobeSummit
marketers need help answering
TWO QUESTIONS…
“What should we
do differently?”
- 6. ©2014 Adometry, Inc. All Rights Reserved. 6#AdobeSummit
Measured and
Managed in Silos
No Unified View
of Performance
Uses Rules-Based
Measurement
Skews Results and
is Click-Centric
Predominately a
Digital Only View
Doesn’t Consider
Offline Effects
Today’s Approach to Digital Marketing
Analytics Has Limitations
- 9. ©2014 Adometry, Inc. All Rights Reserved. 9#AdobeSummit
Mapping Your Ecosystem at the User Level for
Marketing Attribution
CRM and
Audience
Marketing
Events
Cost and
Reference
Conversion
Events
- 10. ©2014 Adometry, Inc. All Rights Reserved. 10#AdobeSummit
A Comprehensive Approach to Collecting
User-Level Attribution Data
All Major Ad Servers
• Historical data
• Simplicity
• Validation
• Greater channel flexibility
• IAB Standard viewability
• 1st and 3rd-party data
integration
• Cross-device mapping
Log Files
All Major Publishers
All Major Tag Managers
Page Tag
Conversion Pixel
Ad Tag
Log File Advantages
Tag Advantages
- 11. ©2014 Adometry, Inc. All Rights Reserved. 11#AdobeSummit
Simple vs. Data-Driven Attribution
Simple
Rules-Based
0% 0% 0% 100%
LAST
EVENT
25%
EVEN
25%25%25%
40%
AD-HOC
15%15%30%
- 12. ©2014 Adometry, Inc. All Rights Reserved. 12#AdobeSummit
Common Advanced Attribution Models
Advantages
• Well understood – “predicts the
contribution” of each touch
• Well suited for “periodic” model updates
Multiple
Regression
Algorithmic
Advantages
• More accurate – actually “counts” the
contribution of each marketing touch
• Well suited for dynamic, i.e. daily
model updates
- 13. ©2014 Adometry, Inc. All Rights Reserved. 13#AdobeSummit
Predictive Model Approaches to Attribution
Have Blind Spots
Discriminating the impact
of granular details for
each ad event
CHALLENGE 1
- 14. ©2014 Adometry, Inc. All Rights Reserved. 14#AdobeSummit
Predictive Model Approaches to Attribution
Have Blind Spots
Discriminating the impact
of granular details for
each ad event
Same marketing event in
different sequence has
same probability
CHALLENGE 1 CHALLENGE 2
- 15. ©2014 Adometry, Inc. All Rights Reserved. 15#AdobeSummit
Data-Driven Attribution Algorithmic
Methodology
Data Driven A/B Testing
3.1%
Conversion
Rate
3.0%
Conversion
Rate
- 16. ©2014 Adometry, Inc. All Rights Reserved. 16#AdobeSummit
Data-Driven Attribution Algorithmic
Methodology
Data Driven A/B Testing
3.1%
Conversion
Rate
2.5%
Conversion
Rate
- 17. ©2014 Adometry, Inc. All Rights Reserved. 17#AdobeSummit
Attribution Performance and Insights at the
Most Granular Level
Paid Search
• Campaign
• Provider
• Keyword Group
• Keyword
• Creative
Display
• Campaign
• Site
• Placement
• Format
• Creative
Email
• Campaign
• Segment
• Primary Message
• Secondary Message
- 18. ©2014 Adometry, Inc. All Rights Reserved. 18#AdobeSummit
Attribution Optimization
How did my marketing perform? How can we improve?
Past FuturePresent
Using Attribution to Fuel Predictive Optimization
- 19. ©2014 Adometry, Inc. All Rights Reserved. 19#AdobeSummit
Scenario
Timeframe Target KPI
Flexible Optimization Planning for
Real-Life Scenarios
Input Data
Set
Budget
Constraints
KPI
Constraints
Forecasted Results for Each Scenario
with Confidence Intervals
- 20. ©2014 Adometry, Inc. All Rights Reserved. 20#AdobeSummit
Overall Optimization
View optimization opportunities
across channels
Evaluating Scenarios
Reallocate spend across
and within channels
For Each Scenario –
Specific Recommendations
Granular line item changes – “I/O ready”
• Modify default recommendations
• Rerun prediction
• Evaluate revised KPIs
Optimal Investment Across and Within Channels
- 21. ©2014 Adometry, Inc. All Rights Reserved. 21#AdobeSummit
Adometry and Adobe Marketing Cloud
Integrations
Integration Point Value Provided
Log File Integration
Eliminate need for new page and conversion tags
by integrating ad server data with Adobe Analytics
data.
Dynamic Tag
Management
Speed time to value by reducing time required to
implement page and conversion tags.
Adobe Analytics
Leverage Adometry attributed KPIs in dashboards
and reports for a more complete view of cross-
channel performance.
Media Optimizer
Improve keyword performance by leveraging fully
attributed value as the target objective for
optimization.
AudienceManager
View segment and audience performance based on
fully-attributed results.
- 22. ©2014 Adometry, Inc. All Rights Reserved. 22#AdobeSummit
Achieving a Greater Return on
Marketing ROI
Value Drivers Areas of Impact
Display eCPA
20% - 30% decrease in effective CPA for
display and retargeting
Optimization Within Channels
10% - 20% improvement by optimizing
channel performance including PPC, Affiliate,
Email, and Social
Optimization Across Channels
10% - 20% improvement in performance by
optimizing spend across various channels
Reduction in Analysis and
Reporting Costs
25% - 50% decrease in effort required to pull,
aggregate, and report on marketing
performance across channels
Overall Marketing ROI
20% - 40% improvement in overall
performance of marketing investments
- 24. 24
15.3 M
17.3 M
PC Tablet + Phone
Lenovo’s Performance
Lenovo Tablet and Smartphone Volume
Exceeded PC Volume since Fiscal Q1
6.5%
8.2%
9.6%
13.1%
18.5%
2009
2010
2011
2012
2013
Lenovo WW PC Market Share
2011/12 2012/13 2013/14
- 30. 32
Our Goals for the Attribution Initiative
View Programs Holistically
Full Funnel View of Performance
Optimize Spend Across Channels
- 31. 33
Our Approach to Capturing Marketing Event Data
Channel Impressions Clicks Primary Collection Method
Brand Display ✔ ✔ DFA and Adobe Analytics
Ecommerce Display ✔ ✔ DFA and Adobe Analytics
Paid Social ✔ ✔ DFA and Adobe Analytics
Email ✔ Adobe Analytics
Direct Navigation ✔ Adobe Analytics
Affiliate ✔ Adobe Analytics
CSEs/GPLAs ✔ Adobe Analytics
Preload ✔ Adobe Analytics
Organic Search ✔ Adobe Analytics
Paid Search ✔ Adobe Analytics
Organic Social ✔ Adobe Analytics
- 33. 35
How Prevalent are Multiple Touches and Channels?
Conversion Types
New and Repeat Orders
New and Repeat Orders –
Systems
Total Orders
Total Orders – Systems
Conversions
Multi-Touch
73%
Multi-Channel
45%
- 34. 36
How Prevalent are Multiple Touches and Channels?
Conversion Types
New and Repeat Orders
New and Repeat Orders –
Systems
Total Orders
Total Orders – Systems
Conversions
Multi-Touch
73%
Multi-Channel
45%
Revenue
Multi-Touch
75%
Multi-Channel
34%
- 37. 39
What is the Predominate Role of Each Channel?
Introducer Channels
• Brand Display
• Direct Navigation
- 38. 40
What is the Predominate Role of Each Channel?
Introducer Channels
• Brand Display
• Direct Navigation
Promoter Channels
• Brand Display
• Paid Social
• Paid Search
• Organic Social
• eCommerce Display
- 39. 41
Introducer Channels
• Brand Display
• Direct Navigation
Promoter Channels
• Brand Display
• Paid Social
• Paid Search
• Organic Social
• eCommerce Display
Closers Channels
• Affiliate
• Organic Search
• Comparison Shopping
Engines (CSEs)
What is the Predominate Role of Each Channel?
- 40. 42
50%
28%
12%
3% 3% 1% 1% 0%
Direct Organic
Search
Display -
Ecommerce
Email Paid
Search
CSE Display -
Brand
Other
Revenue by Channel
What is the Revenue Performance by Channel?
- 41. 43
50%
28%
12%
3% 3% 1% 1% 0%
Direct Organic
Search
Display -
Ecommerce
Email Paid
Search
CSE Display -
Brand
Other
Revenue by Channel
What is the Revenue Performance by Channel?
Revenue/User by Channel
Organic
Search
Direct Email Display -
Ecommerce
Display -
Brand
Paid Search CSE Affiliate
- 43. 45
CSE Display -
Ecommerce
Paid Search Affiliate Display -
Brand
Return on Ad Spend (ROAS)
What is the Efficiency of Each Channel?
Conversion Rate
Paid
Search
Email Affiliate Organic
Social
Organic
Search
Paid
Social
Direct
Nav
- 44. 46
Average Lift on TOTAL order conversions to eCommerce
programs when Brand Display precedes lower funnel
(click based) programs
SEO:
+353.39%
SEM:
+25.45%
Email:
+70.68%
Affiliate:
+18.46%
CSE:
+40.72%
Average Lift on NEW order conversion to eCommerce
programs when Brand Display precedes lower funnel
(click based) programs
SEO:
+420.40%
SEM:
+10.76%
Email:
+7.85%
Affiliate:
NO LIFT
CSE:
+21.04%
Average Lift on TOTAL order conversions to eCommerce
programs when eCommerce Display precedes lower funnel
(click based) programs
SEO:
+769.15%
SEM:
+82.03%
Email:
+65.32%
Affiliate:
+49.78%
CSE:
+100.46%
Average Lift on NEW order conversion to eCommerce
programs when eCommerce Display precedes lower funnel
(click based) programs
SEO:
+616.08%
SEM:
+16.02%
Email:
+65.19%
Affiliate:
NO LIFT
CSE:
+17.98%
Brand and eCommerce Display Lift
- 45. 47
Next Areas of Focus
Applying initial insights to
drive optimization
Global rollout
Key countries and regions
Integrate call-center data
Evaluate methods for
integrating in-store
purchase data
- 47. © 2014 Adobe Systems Incorporated. All Rights Reserved. Adobe Confidential.
Tuesday
Grand PrizeEvery Session
Wednesday
Grand Prize
Thursday
Grand Prize
Take Survey & Win Sweet SWAG
(“surveys” section of mobile app)
Signed Jersey
Richard Sherman$10
Gift Card Swag Pack
Band Ski
Swag
- 48. ©2014 Adometry, Inc. All Rights Reserved. 50#AdobeSummit
Paul Pellman
paul.pellman@adometry.com
Lewis Broadnax
lewismb@lenovo.com
Editor's Notes
- Slide ObjectiveKey Talking Points
- Slide ObjectiveKey Talking Points
- Slide ObjectiveSet up key pain points with the status quoKey Talking PointsData is in silos, no one place to get to get answers to your questions, no system of record for marketing performance. Rather it is in:Web analytics toolAd serverEmail platformSocial platformsYour agencyUses long, established rule-based attributionEasy to do, supported by current tools setsPrimarily click-basedArbitrary rule setNo unified metrics across channelsPredominately a digital viewDoesn’t include digital impact on offline conversions (POS, branch, dealer, call-center)Doesn’t include impact of offline channels such as direct mailAs a result, marketers do not have the clarity and insights necessarily to excel in today’s hypercompetitive, complex multi-channel world
- Slide ObjectiveTalk about the current state of marketing attributionKey Talking PointsLot’s of blogs, articles, and presentation claiming the demise of last clickHowever, depending on which studies you read, somewhere between 45% and 65% are still using it. In an IgnitionOne study late last year, they reported:24% aren’t doing attribution – which probably means they are double-counting or using activity based metrics58% are using last click18% are doing attribution and using something other than last clickGenerally we all agree, there are better approaches that will drive improved understanding and marketing performance
- The task of assigning credit might sound simple, but in today’s omni-channel world, it is anything but – there are a lot of barriers to change. When I speak with companies, I consistently hear four main ones highlighted here on this study by AdAge and Nuestar
- Slide ObjectiveKey Talking Points
- Slide ObjectiveDiscuss advantages and disadvantages of two primary approaches to attribution data collectionKey Talking PointsViewability – Ad tags measures IAB standard of 50% in view for a minimum of 1 secData integration – foundation for extending the data to include a more comprehensive viewAudience DataCross device mappingOffline Conversion mappingDirect Mail mappingCookie Deletion1st and 3rd Party Cookies
- Slide ObjectiveSetup background for methodology conversationKey Talking PointsTypical conversion path across display, email, and searchRules based models = pre determined. These are the models that ad servers and site analytics provideTypically only looking at click-events – may include view-throughFocused on conversion paths, does not consider non-converting (order of magnitude more data about what doesn’t work)
- Slide ObjectiveKey Talking PointsKEY PLACE TO POINT OUT BENEFIT/ABILITY TO PROVIDE DAILY ATTRIBUTION – DAILY COUNTING EXERCISE, BETTER SUITED TO SCALETo understand the weight or credit due to an event in a particular sequence, we first find all examples of that sequence – both converting and non-converting. From this we can calculate the conversion rate for that sequence.Next, we find a similar sequence of events that excludes one of the events from the prior sequence – for example, the first event. Again, we find both converting and non-converting examples. We then calculate the conversion rate for this (second) sequence.Now, by comparing the conversion rates of the sequences we can determine the weight or impact of the the missing event. If the rates are identical or similar then the missing event deserves little or no credit whereas if the rates are different then that missing event had an impact and deserves credit.There are some points that you can make after this – the death is in the details stuff:Obviously simplified explanation.Number of sequence and length of sequences will varyWe normalize the weights such that they add up to one (1)Tons of events - "big data problem"If there is only one event in the sequence it will get 100% of the creditAn event is actually a complicated thing – has lots of attributes – site, campaign, placement, creative concept, tactic, etc. we calculate at the lowest level and roll the results upAlso frequency and recency.But this is the stuff that the engineers have all worked out in a principled manner. Machine learning, etc.
- Slide ObjectiveKey Talking PointsKEY PLACE TO POINT OUT BENEFIT/ABILITY TO PROVIDE DAILY ATTRIBUTION – DAILY COUNTING EXERCISE, BETTER SUITED TO SCALETo understand the weight or credit due to an event in a particular sequence, we first find all examples of that sequence – both converting and non-converting. From this we can calculate the conversion rate for that sequence.Next, we find a similar sequence of events that excludes one of the events from the prior sequence – for example, the first event. Again, we find both converting and non-converting examples. We then calculate the conversion rate for this (second) sequence.Now, by comparing the conversion rates of the sequences we can determine the weight or impact of the the missing event. If the rates are identical or similar then the missing event deserves little or no credit whereas if the rates are different then that missing event had an impact and deserves credit.There are some points that you can make after this – the death is in the details stuff:Obviously simplified explanation.Number of sequence and length of sequences will varyWe normalize the weights such that they add up to one (1)Tons of events - "big data problem"If there is only one event in the sequence it will get 100% of the creditAn event is actually a complicated thing – has lots of attributes – site, campaign, placement, creative concept, tactic, etc. we calculate at the lowest level and roll the results upAlso frequency and recency.But this is the stuff that the engineers have all worked out in a principled manner. Machine learning, etc.
- Slide ObjectiveKey Talking PointsTo understand the weight or credit due to an event in a particular sequence, we first find all examples of that sequence – both converting and non-converting. From this we can calculate the conversion rate for that sequence.Next, we find a similar sequence of events that excludes one of the events from the prior sequence – for example, the first event. Again, we find both converting and non-converting examples. We then calculate the conversion rate for this (second) sequence.Now, by comparing the conversion rates of the sequences we can determine the weight or impact of the the missing event. If the rates are identical or similar then the missing event deserves little or no credit whereas if the rates are different then that missing event had an impact and deserves credit.There are some points that you can make after this – the death is in the details stuff:Obviously simplified explanation.Number of sequence and length of sequences will varyWe normalize the weights such that they add up to one (1)Tons of events - "big data problem"If there is only one event in the sequence it will get 100% of the creditAn event is actually a complicated thing – has lots of attributes – site, campaign, placement, creative concept, tactic, etc. we calculate at the lowest level and roll the results upAlso frequency and recency.But this is the stuff that the engineers have all worked out in a principled manner. Machine learning, etc.
- Slide ObjectiveKey Talking PointsTo understand the weight or credit due to an event in a particular sequence, we first find all examples of that sequence – both converting and non-converting. From this we can calculate the conversion rate for that sequence.Next, we find a similar sequence of events that excludes one of the events from the prior sequence – for example, the first event. Again, we find both converting and non-converting examples. We then calculate the conversion rate for this (second) sequence.Now, by comparing the conversion rates of the sequences we can determine the weight or impact of the the missing event. If the rates are identical or similar then the missing event deserves little or no credit whereas if the rates are different then that missing event had an impact and deserves credit.There are some points that you can make after this – the death is in the details stuff:Obviously simplified explanation.Number of sequence and length of sequences will varyWe normalize the weights such that they add up to one (1)Tons of events - "big data problem"If there is only one event in the sequence it will get 100% of the creditAn event is actually a complicated thing – has lots of attributes – site, campaign, placement, creative concept, tactic, etc. we calculate at the lowest level and roll the results upAlso frequency and recency.But this is the stuff that the engineers have all worked out in a principled manner. Machine learning, etc.
- Slide ObjectiveKey Talking Points
- Lays the foundation for our benefits.Introduction to how we approach the problem of looking at historical data and making changes to improve performance.
- Slide ObjectiveKey Talking Points
- Slide ObjectiveKey Talking Points
- Slide ObjectiveKey Talking Points
- Each session will have a survey winner selected at the end of the conference day who will receive a $10 Starbucks electronic gift card. In addition, you'll be entered into our grand prize raffle where a grand prize winner will be drawn at the end of each conference day for an opportunity to win exciting prizes including an autographed Richard Sherman jersey, a Summit bash experience package and ski gear from Park City! Each survey responded to will give you another opportunity to win!