Predictive Next Best Action for
Marketing Demand Generation and
Sales
Erik Bower & Josip Lazarevski (Palo Alto Networks)
3/11/17
Agenda
•
Current state of marketing
•
Omnichannel predictive next best action marketing concept
•
The plumbing
•
The model and how it was made
•
Results so far
First up – A better name
Omnichannel real
time predictive
next best action
marketing
“Machine
Marketing”
Marketing nirvana
!
Right message
4
Omnichannel
Right person
Right time
What are “channels”?
5 | © 2016, Palo Alto Networks, Inc. Confidential and Proprietary.
How to reach nirvana?
Get right message?
Get right timing?
Target the right people?
Email
A/B testing
Test send times, schedule, some
have automatic “throttling”
Segmentation 1-2-1 and
audience based
Web Personalization
A/B testing
Depends on the prospect
Segmentation 1-2-1 and
audience based
Display/Programmatic
Advanced optimization trained on
conversion
Day parting, frequrency capping
Segmentation 1-2-1 and
audience based
6 | © 2016, Palo Alto Networks, Inc. Confidential and Proprietary.
Example “Buyer’s Journey”
7 | © 2016, Palo Alto Networks, Inc. Confidential and Proprietary.
Traditional Marketing
Automation
•
Logic is static
•
Content gap, content age
•
Coverage is spotty
•
Customer journeys are infinite
Machine Marketing
VS.
▪
Dynamic and adaptive
▪
Greater coverage of microsegments
Predicts best content/offer
Syncs and optimizes across
channels
▪
▪
8
How would Machine Marketing work?
9 | © 2016, Palo Alto Networks, Inc. Confidential and Proprietary.
First you need the messages or “offers”
10
Then you need to be able to sync messages across channels
Campaign
Target
Audience Manager
Predictive
11
Analytics
Then deploy that message on the web….
12 | © 2016, Palo Alto Networks, Inc. Confidential and Proprietary.
….on the display channel….
..and via email channel
14 | © 2016, Palo Alto Networks, Inc. Confidential and Proprietary.
Figure out the best time to send an email
Modeling Period
Account
Contact
DnB
Product
Web
Visits,
Contacts we have
sent at least 1
email between
July 2016 and
December 2016
TRUE
Clicked on at least 1 email
Not clicked on any email
FALSE
Performance Testing Period
Contacts
we have
sent at
least 1
email on
January
2017
TRUE
FALSE
6 MONTHS
Clicked on
at least 1
email
Not
clicked on
any email
1 MONTH
|_________________________________________________|________________________|
Form Fills
July 2016 – December 2016
15 | © 2017, Palo Alto Networks, Inc. Confidential and Proprietary.
January 2017
77%
80%
Accuracy
Accuracy
But wait there’s more
Identity resolution
+
Behaviorial data
+
CRM data
+
Other data
16
A bit complicated to pull off
17
How the model was built
Understand the audience and their preferences
Lookalike modeling using similarity search to understand user behavior
B2B: Understanding account/company behavior
Lookalike modeling using similarity search (similar to household in B2C)
Hybrid of boosted trees and ensemble trees
These are the people in your neighborhood
All that joined together
forms our neighborhoods
from where we extract
knowledge –
collaborative filtering
Organizing the messages with keyword mining
Humans got
tagging wrong
70% of the time
The formula
+ Keyword Clustering (KNN,SVM)
+ Term Grading
= Keyword – Topic Mapping
Language recognition
•
We want to localize our recommendations but we don’t know their language
•
Using blend of R script and KNIME we can recognize 32 languages
library("textcat");
xls<- knime.in
result.detectedLanguage<-textcat(xls$DMText)
knime.out <-data.frame(xls$"RECSentity.id",result.detectedLanguage);
Deploy to Adobe Marketing Cloud
Understand the audience
and their preferences
Asset to person matching
Understand the content
of assets
Results so far
27 | © 2016, Palo Alto Networks, Inc. Confidential and Proprietary.
Homepage “King of the Hill” results
+1300%
Predicted
Default
Programmatic in Doubleclick results
Dynamically Populated
Click-Thru-Rate (CTR) By Audience
1.200%
1.000%
Headline
0.800%
0.600%
0.400%
Button
0.200%
1st Party Cookies Display
Ad
Populates
on 3rd Party
Sites
30 | © 2015, Palo Alto Networks. Confidential and Proprietary.
Standard Retargeting
Maestro Display
Standard Display CTR: .027%
Retargeting CTR: .080%
Maestro Display CTR: .351%
+338% Increase in Engagement
28. Jan
26. Jan
24. Jan
22. Jan
20. Jan
18. Jan
16. Jan
14. Jan
12. Jan
10. Jan
08. Jan
06. Jan
04. Jan
02. Jan
31. Dez
29. Dez
27. Dez
Exit URL
25. Dez
0.000%
Early signs are positive
32
If you are a marketing person, don’t worry……
I’m not going
to take your
job
33 | © 2016, Palo Alto Networks, Inc. Confidential and Proprietary.
No really