Upload
dinhkien
View
213
Download
0
Embed Size (px)
Citation preview
© 2011 MIT Center for Digital Business. All rights Reserved.
Ad Morphing May 19th, 2011
Proprietary & Confidential
Glen Urban, George Pappachen, Gui Liberali MIT, WPP/Kantar, Erasmus University
Google and WPP Supported Research
Purpose
• Test web site morphing methodology on advertising
• Serve ad to match cognitive style
• Increase CTR, purchase propensity?
Morphing Site Designs
Click descriptor Website
Bayesian update
Priors on distribution of decision styles
Updated user decision-style descriptor On-line
Optimization
Optimal morph for current user
Pgm: probability of purchase for individual in group g given morph m
Source: Hauser et al., 2009
Side by Side Comparison Morph 1 Morph 8
Visual
Technical
Content
More
Content
Audio
Less
Content
General
Content
Source: Hauser et al., 2009
Data Collection
Consumer Profiling
Targeting by Cognitive
Style
Analysis of Results
• Track clickstreams on CNET.com • Survey small segment to determine cognitive style clicking patterns
• Using MIT engine, determine cognitive style of cookies from clickstream
• Remessage cookies on CBS ad network
• Morph AT&T banner messaging to match cognitive style
• Learn optimal matching dynamically
• Measure lift in clickthrough rate behavior for morphed vs. non- morphed AT&T ads
Project Plan
Research Design • Priming Study – clicks to cognitive
style – Bayesian model
• 8 AT&T banners, 4 styles
• 15 CNET pages monitored, up to 5 links monitored per page
• Once ad shown, stays with same ad
• 30% of respondents in control condition (random ad)
• 70% of respondents in test condition (morphing ad)
• 250 thousand + impressions
• 20 thousand impressions to reach convergence – “burn in”
• Each monitored click updates priors on cognitive style
• After 5 updates serves optimal ad
Morph Design – 8 Alternatives
• Same product, vary amount of content – features vs. deal
• “Learn more” versus “Buy it now”
• Font
• Color
• Web-like banners and traditional banners
• Different format
MIT-CNET/CBS Collaboration System coded into www.CNET.com across multiple divisions
1. Content engineering team
• Infer cognitive style from clicks
• Compute optimal morph using current assignment rules
2. Ad serving team
• attempted ads, served ads (depends on inventory, sales)
3. BI/reporting team
• daily report of successes and failures to MIT update assignment rules
MIT-CNET/CBS Flow
CNET Site
Generate G matrix:
Bayesian Engine:
Ad Server:
Check for ad
priority, serve ad if
appropriate
Ad Selection:
30% (control): Show random ad
from 8 options
70% (test):
Site Visitor
Update r
Update G (8x4) on
CNET side
Make table of day’s
activity, send to MIT
On 5th click,
select ad to
show.
G
Make
attempt
Sho
w Ad
Impression
Tracker:
Record ad
impression,
timestamp,
click-
through,
current r
CNET
MIT
Update data accumulator
matrices α and β.
Summary
Table
α, β
Load data conversion
matrix D.
D
• Improves knowledge of optimal assignments as more and more users go through the system
• Burn-in phase
Dynamics of Morphing
Traffic
Reviews: 48.6% News: 49.8% Downloads: 0.94% Video: 0.06%
Experimental Ad
impressions on MIT-
monitored CNET
pages between April
12th, 2011 and May
3rd, 2011.
Home: 0.5%
Current Results Click-Through Rates
Overall: 182k impressions, 60.8k cookies
TEST (Morphing Ads) CONTROL (Random) DIFFERENCES
CTR test group
n test clicks test
CTR control group
n control clicks
control diff p
Lift (%) Impression Level
Impression 0.211% 12,801 27 0.118% 6,783 8 0.093% 0.066 79%
Cookie 0.248% 8,873 22 0.116% 4,296 5 0.132% 0.054 113%
Contextual Morphing: 19.5k impressions, 13.2k cookies Limited to all visited CNET webpages with "phone” or "htc” in its URL between April
12th and May 5th. Most are reviews and news.
TEST (Morphing Ads) CONTROL (Random) DIFFERENCES
CTR test group
n test clicks test
CTR control group
n control clicks
control diff p
Lift (%) Impression Level
Impression 0.168% 170,107 285 0.151% 97,038 147 0.016% 0.284 11%
Cookie 0.183% 52,014 95 0.163% 23,280 38 0.019% 0.533 12%
Cognitive Style of Users (divided by Ad seen)
Test (Morphing) Average 1 2 3 4 5 6 7 8
Deliberative-Holistic 0.09 0.05 0.07 0.03 0.14 0.20 0.03 0.04 0.15
Deliberative-Analytical 0.42 0.74 0.27 0.59 0.28 0.24 0.45 0.34 0.41
Impulsive-Holistic 0.23 0.12 0.11 0.16 0.40 0.34 0.20 0.30 0.22
Impulsive-Analytical 0.27 0.09 0.55 0.23 0.18 0.22 0.32 0.33 0.22
Control (Random) Average 1 2 3 4 5 6 7 8
Deliberative-Holistic 0.08 0.07 0.08 0.08 0.08 0.09 0.06 0.08 0.06
Deliberative-Analytical 0.45 0.46 0.45 0.48 0.47 0.43 0.44 0.44 0.46
Impulsive-Holistic 0.22 0.23 0.19 0.22 0.22 0.18 0.25 0.23 0.24
Impulsive-Analytical 0.25 0.25 0.29 0.21 0.23 0.30 0.25 0.25 0.24
Ads Assigned by Cognitive Style
Test (Morphing) 1 2 3 4 5 6 7 8
Deliberative-Holistic 0.07 0.10 0.04 0.20 0.28 0.04 0.06 0.21 Deliberative-Analytical 0.22 0.08 0.18 0.08 0.07 0.14 0.10 0.12
Impulsive-Holistic 0.06 0.06 0.09 0.22 0.18 0.11 0.16 0.12 Impulsive-Analytical 0.04 0.26 0.11 0.08 0.10 0.15 0.15 0.10
Control (Random) 1 2 3 4 5 6 7 8
Deliberative-Holistic 0.12 0.13 0.13 0.13 0.15 0.10 0.13 0.10 Deliberative-Analytical 0.13 0.12 0.13 0.13 0.12 0.12 0.12 0.13
Impulsive-Holistic 0.13 0.11 0.13 0.13 0.10 0.14 0.13 0.14 Impulsive-Analytical 0.12 0.14 0.10 0.11 0.15 0.12 0.12 0.12
Some Implications
– Adaptation of on-line ads to match cognitive styles has a clear impact on CTR
– Cognitive styles can provide insight for on-line strategy
– Morphing works best for contextual advertising
Relevant Challenges, Opportunities
– Adapt banners and entire browsing experience
– Adaptation based on Geo-targeting?
– More live experiments:
• Large-scale, multiple-platform experiment, test of campaigns across different (competing?) sites
• Controlled experiment to explore budget allocation among overall awareness and different stages along the funnel