2. About me
~25%
1,300
12B 2.4B
• Run Marketing Operations &
Analytics team at Microsoft
• Responsible for RM Marketing
Ops for Windows, Surface,
Xbox, Bing, MSN, Microsoft
Rewards, Windows Store
• 16 years in Web Analytics, Big
Data & Digital Marketing
operations
3. The problem
How to maximize digital
campaign effectiveness across
multiple channels?
4.
5. Database Plan Deliver Analyze
Optimize
The traditional model (aka, the good old days)
Timescale: weeks
Ask for more budget
7. Our problems
Multichannel digital attribution is a fool’s errand
Manually optimizing over multiple channels is incredibly
time-consuming & complicated
Traditional model of plan-deliver-analyze-adjust doesn’t
scale in an omnichannel world
9. Digital marketing as an optimization problem
Inputs
User profile
Offer
Creative
Tactic/channel
Outputs
Views/clicks
Conversions
Engagement
Revenue/GM
Constraints
Budget
Inventory
Continuous optimization
10. The things you need
1 A comprehensive user profile
2 Integrated delivery systems using a common creative repository
3 A set of common objectives
4 Integrated Marketing Operations function
5 An optimization engine
11. A/B testing vs multi-armed bandits
Looking for definitive “winner” of
a number of options
“Explore” phase precedes
separate “exploit” phase
Good for relatively stable
environments (where the winner
stays the winner)
Quicker to get to statistical
significance in results
Looking for the “best” of a
number of options
“Explore” phase overlaps “exploit”
phase
Good when the conditions that
make “best” change over time
(i.e. continuous optimization)
Minimizes traffic to poorly-
performing alternatives
12. Dimensions of data
Audience Data
• Product ownership
• Product Engagement
• Marketing Engagement
• Attitudes
Offer Data
• Product info
• Price range
• Purchase model
Tactics
• Channel
• Creative
• Timing
• Format
13. Too many vs too few attributes
Too few:
Model optimizes quickly,
but with low lift
Too many:
Model optimizes too
slowly, or never
14. Single-channel, single campaign optimization
User data Matching engine Ruleset Delivery Metrics
Creative library
Data feedback
Plan
17. Games with Gold
Experiment with offers, creative and timing
within Games with Gold email series
Also extending to in-product notifications
Goals: Increase lift, reduce effort associated
with putting this email together
18. In summary…
• Think of campaign
optimization as a single
space across channel /
user / offer / creative
• Pick the crawl – walk –
run that is right for your
business
• Let me know how you
get on so I can learn
from you!
• Amplero
• Salesforce (Einstein)
• Optimove
• Kahuna
Companies in
this space
Graphic showing how people can’t work in channel-specific silos any more, and the effectiveness of their campaigns has to be understood in the context of all digital activity
+ They have to compete with others for scarce marketing budget
+ They are expected to execute more and more granular campaigns – running out of time for traditional campaign analysis
You need to think about your users, your campaigns, your offers and your objectives as four sets of related data that you need to optimize over
At any moment, you’re looking to find the optimal combination of user, offer, tactic/channel to generate the best outcomes
Everything you do (and the results it generates) should be fed back into the machine to make it better
The approach we're talking about here is basically a combination of multi-variate testing and multi-armed bandit experimentation.
The various user, campaign & offer attributes are the multiple variables that go into a series of tests; each unique combination of the attributes becomes one of the "bandits", whose potential is constantly being assessed through experimentation. Also, new "bandits" are constantly being created as new offers and creatives are put into the system (and old ones are retired) - though you want to reduce bandit churn where you can (i.e. reduce the churn in your attributes and the values of those attributes).
Key activity you need to undertake is to decide on the attributes you want to take into account.
There are three major dimensions of data you can use in this – offer info, audience data, & tactics
You can choose to optimize in one, two or all three of these areas
Which features of these dimensions you choose will be determined by a number of factors:
Offer data: If you have a large catalog of things to offer, you can experiment with lots of different offers, so long as you have the creative assets to represent them
Audience data: If you know a lot about your audience, you can experiment with audience data
Tactics: If your delivery systems are well-integrated, and you have a good creative repository, you can experiment across tactics (including channels)