1
Programmatic Advertising
from a Data Scientist’s Perspective
OMK, 17th August 2017, Bern
Gergely Kalmár | Senior Consultant Digital Analytics | Webrepublic AG
2Data Points.
Lots of Them.
Programmatic Advertising – What’s in it for a Data Scientist?
3
Agenda
1. Introduction
2. The Programmatic Data Landscape
3. A Simple Analysis: Audience Clustering
4. An Advanced Analysis: Marketing Mix Optimization
5. Summary & Outlook
4
Introduction
5
- Performance Marketing, Online Advertising, Digital Analytics
- 120 clients, national and international
- 130 employees
- 12 languages spoken in-house
- dedicated
- digital analytics team
- software engineering
- graphics department
- iab Digital Agency of The Year 2017
6
What is a “programmatic campaign”?
For the current session we define it as follows:
A programmatic campaign is a multi-channel online campaign that has
a central server that is able to track all ad impressions and clicks for
all involved channels, as well as specific online conversions.
Programmatic campaigns may (but not necessarily) as well:
- use a platform for buying display placements that are then served
programmatically,
- use a central ad server hosting the display creatives.
7
The Programmatic
Data Landscape
8
Customer Journey Tracking
9
Customer Journey Tracking
Actually
female
Cost data
unavailable
User deleted
cookies
User bought the
product because her
mother asked to
Sad data
scientist
Happened
on mobilecost data
10
The Programmatic Data Landscape
Display campaigns
Impression data
Click data
Behavioral data
Audience data
Cost data
Your website
DSP Ad server
11
The Programmatic Data Landscape
Display campaigns
Impression data
Click data
Behavioral data
Audience data
Cost data
Your website
DSP Ad server
Tracking
Media buying
12
The Programmatic Data Landscape
Display campaigns
Impression data
Click data
Behavioral data
Audience data
Cost data
Your website
DSP Ad server
Tracking
Media buying
Sync
13
The Programmatic Data Landscape
Display campaigns
Impression data
Click data
Behavioral data
Audience data
Cost data
Your website
DSP Ad server
Important!
14
Web Analytics Tool-Based Attribution
Web analytics tool
15
Web Analytics Tool-Based Attribution
Web analytics tool
Impressions
are “invisible”
Search gets all the credit
16
Web Analytics Tool-Based Attribution
Ad Server
17
Web Analytics Tool-Based Attribution
Both the display and search
channels may get credit
Ad Server
18
Fun fact: view-through
conversions can account for
up to 80% of all conversions
on display (banner) campaigns.
Source: a campaign executed in May-June 2017 for a global
brand; measured view-through conversion contribution was
80% in CH, 63% in USA, 44% in DE.
19
Section Summary
- Use an ad server to enable centralized, impression-level tracking
especially when you have display or social media campaigns.
- Do not be surprised if your display or social media campaigns do not
seem to perform well in your web analytics tool.
- Always remember: the fact that someone has clicked on your ad and
converted does not mean that the person converted because of your
ad.
20
A Simple Analysis:
Audience Clustering
21
The Programmatic Data Landscape
Display campaigns
Impression data
Click data
Behavioral data
Audience data
Cost data
Your website
DSP Ad server
22
Audience Clustering & Optimization
1. Identify. 2. Focus.
23
Audience Clustering & Optimization
1. Identify. 2. Focus.
Low-performance
niche audiences
High-performance
niche audiences
24
Audience Clustering & Optimization
1. Identify. 2. Focus.
Optimization
potential
25
Audience Clustering & Optimization
Males
Females
26
Audience Clustering & Optimization
Males
Females
News
junkies
Celebrity
news
Cooking
enthusiasts
27
Audience Clustering & Optimization
Males
Females
News
junkies
Celebrity
news
Cooking
enthusiasts
Autos &
vehicles
Land Rover
(in-market)
Luxury
travelers
28
Section Summary
- When doing audience optimization focus on high-volume, low-CPA
audiences.
- Sorting the audiences by lowest CPA may not help much, because the
ones with the highest efficiency will be typically small niches.
- It is usually helpful to analyze remarketing audiences separately from
other audiences.
29
An Advanced Analysis:
Marketing Mix Optimization
30
Marketing Mix Optimization
Digital marketing investments can be changed in a matter of
seconds.
- The investments can be shifted from less-performing channels to better
performing ones (manual).
- The overall investment can be adjusted depending on the company’s sales
performance in near real time.
Key Questions:
1. How much performance boost can we expect from our budget optimization
efforts?
2. How can we select the optimal budget for our digital marketing channels?
31
The overall number of conversions (or the conversion value if
available) needs to be maximized:
where Ci is the number of conversions for channel i, Bi is the
budget for channel i and S is the overall digital marketing budget.
Problem Definition
We need a model to predict the
number of conversions for a given
budget for each marketing channel
Assumption:
all marketing
channels are
independent
32
Budget Curves
The core problem to solve is finding the relationship between the marketing
budget and the number of conversions (including view-through) for a given
marketing channel.
An example of a possible model for Google AdWords
33
Summary & Outlook
- Use an ad server to enable centralized, impression-level tracking especially
when you have display or social media campaigns.
- High-level audience clustering can be done with relative ease, and is often
times included in Data Management Platforms (DMPs).
- Marketing mix optimization needs advanced predictive models with high
accuracy and thus specific knowledge or (typically expensive) tools.
- An appropriate system is capable of tracking campaigns on a micro-level,
however, campaign management typically happens on a macro-level. The
current optimization approaches are therefore quite rudimentary.
34
Let’s make it happen.
Thank you.
Contact: +41 44 542 45 11
gergely.kalmar@webrepublic.com
Thank you for your
attention.

Programmatic Advertising from a Data Scientist’s Perspective

  • 1.
    1 Programmatic Advertising from aData Scientist’s Perspective OMK, 17th August 2017, Bern Gergely Kalmár | Senior Consultant Digital Analytics | Webrepublic AG
  • 2.
    2Data Points. Lots ofThem. Programmatic Advertising – What’s in it for a Data Scientist?
  • 3.
    3 Agenda 1. Introduction 2. TheProgrammatic Data Landscape 3. A Simple Analysis: Audience Clustering 4. An Advanced Analysis: Marketing Mix Optimization 5. Summary & Outlook
  • 4.
  • 5.
    5 - Performance Marketing,Online Advertising, Digital Analytics - 120 clients, national and international - 130 employees - 12 languages spoken in-house - dedicated - digital analytics team - software engineering - graphics department - iab Digital Agency of The Year 2017
  • 6.
    6 What is a“programmatic campaign”? For the current session we define it as follows: A programmatic campaign is a multi-channel online campaign that has a central server that is able to track all ad impressions and clicks for all involved channels, as well as specific online conversions. Programmatic campaigns may (but not necessarily) as well: - use a platform for buying display placements that are then served programmatically, - use a central ad server hosting the display creatives.
  • 7.
  • 8.
  • 9.
    9 Customer Journey Tracking Actually female Costdata unavailable User deleted cookies User bought the product because her mother asked to Sad data scientist Happened on mobilecost data
  • 10.
    10 The Programmatic DataLandscape Display campaigns Impression data Click data Behavioral data Audience data Cost data Your website DSP Ad server
  • 11.
    11 The Programmatic DataLandscape Display campaigns Impression data Click data Behavioral data Audience data Cost data Your website DSP Ad server Tracking Media buying
  • 12.
    12 The Programmatic DataLandscape Display campaigns Impression data Click data Behavioral data Audience data Cost data Your website DSP Ad server Tracking Media buying Sync
  • 13.
    13 The Programmatic DataLandscape Display campaigns Impression data Click data Behavioral data Audience data Cost data Your website DSP Ad server Important!
  • 14.
    14 Web Analytics Tool-BasedAttribution Web analytics tool
  • 15.
    15 Web Analytics Tool-BasedAttribution Web analytics tool Impressions are “invisible” Search gets all the credit
  • 16.
    16 Web Analytics Tool-BasedAttribution Ad Server
  • 17.
    17 Web Analytics Tool-BasedAttribution Both the display and search channels may get credit Ad Server
  • 18.
    18 Fun fact: view-through conversionscan account for up to 80% of all conversions on display (banner) campaigns. Source: a campaign executed in May-June 2017 for a global brand; measured view-through conversion contribution was 80% in CH, 63% in USA, 44% in DE.
  • 19.
    19 Section Summary - Usean ad server to enable centralized, impression-level tracking especially when you have display or social media campaigns. - Do not be surprised if your display or social media campaigns do not seem to perform well in your web analytics tool. - Always remember: the fact that someone has clicked on your ad and converted does not mean that the person converted because of your ad.
  • 20.
  • 21.
    21 The Programmatic DataLandscape Display campaigns Impression data Click data Behavioral data Audience data Cost data Your website DSP Ad server
  • 22.
    22 Audience Clustering &Optimization 1. Identify. 2. Focus.
  • 23.
    23 Audience Clustering &Optimization 1. Identify. 2. Focus. Low-performance niche audiences High-performance niche audiences
  • 24.
    24 Audience Clustering &Optimization 1. Identify. 2. Focus. Optimization potential
  • 25.
    25 Audience Clustering &Optimization Males Females
  • 26.
    26 Audience Clustering &Optimization Males Females News junkies Celebrity news Cooking enthusiasts
  • 27.
    27 Audience Clustering &Optimization Males Females News junkies Celebrity news Cooking enthusiasts Autos & vehicles Land Rover (in-market) Luxury travelers
  • 28.
    28 Section Summary - Whendoing audience optimization focus on high-volume, low-CPA audiences. - Sorting the audiences by lowest CPA may not help much, because the ones with the highest efficiency will be typically small niches. - It is usually helpful to analyze remarketing audiences separately from other audiences.
  • 29.
  • 30.
    30 Marketing Mix Optimization Digitalmarketing investments can be changed in a matter of seconds. - The investments can be shifted from less-performing channels to better performing ones (manual). - The overall investment can be adjusted depending on the company’s sales performance in near real time. Key Questions: 1. How much performance boost can we expect from our budget optimization efforts? 2. How can we select the optimal budget for our digital marketing channels?
  • 31.
    31 The overall numberof conversions (or the conversion value if available) needs to be maximized: where Ci is the number of conversions for channel i, Bi is the budget for channel i and S is the overall digital marketing budget. Problem Definition We need a model to predict the number of conversions for a given budget for each marketing channel Assumption: all marketing channels are independent
  • 32.
    32 Budget Curves The coreproblem to solve is finding the relationship between the marketing budget and the number of conversions (including view-through) for a given marketing channel. An example of a possible model for Google AdWords
  • 33.
    33 Summary & Outlook -Use an ad server to enable centralized, impression-level tracking especially when you have display or social media campaigns. - High-level audience clustering can be done with relative ease, and is often times included in Data Management Platforms (DMPs). - Marketing mix optimization needs advanced predictive models with high accuracy and thus specific knowledge or (typically expensive) tools. - An appropriate system is capable of tracking campaigns on a micro-level, however, campaign management typically happens on a macro-level. The current optimization approaches are therefore quite rudimentary.
  • 34.
    34 Let’s make ithappen. Thank you. Contact: +41 44 542 45 11 gergely.kalmar@webrepublic.com Thank you for your attention.