Top Banner
Copyright©2012MarketingAssociatesLLC.Allrightsreserved. DecisionSciences Quantifyingthe"Buzz"Effect:IntegratingSocial MediawithLoyalty&DefectionModels MarketingAssociates: KeithShields,Director,DecisionSciences RoniLeibovitch,SeniorConsultant,DigitalIntelligence MindyDeatrick,SeniorConsultant,QuantitativeSolutions FordMotorCompany: MargaretKishore,PerformanceandMetricsManager CreateaBusinessBlueprint WithDataDrivenCustomerInsights
32

AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Jan 20, 2015

Download

Sports

dmadetroit

Keith Shields Marketing Associates Managing Director of the Decisions Science Group
“Quantifying the Buzz Effect: Integrating Social Media with Loyalty & Defection Models”
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page 1: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Quantifying�the�"Buzz"�Effect:�Integrating�Social�Media�with�Loyalty�&�Defection�Models

Marketing�Associates:�Keith�Shields,�Director,�Decision�Sciences

Roni�Leibovitch,�Senior�Consultant,�Digital�IntelligenceMindy�Deatrick,�Senior�Consultant,�Quantitative�Solutions

Ford�Motor�Company:Margaret�Kishore,�Performance�and�Metrics�Manager

Create�a�Business�BlueprintWith�Data�Driven�Customer�Insights

Page 2: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

About�the�Title…

� “Buzz”�refers�to�the�amount�of,�and�sentiment�of,�the�Ford�related�comments�available�through�social�media�outlets.

� The�“Buzz�Effect”�refers�to�the�increase�or�decrease�in�brand�loyalty�/�defection�(measured�by�repurchase)�that�occurs�as�a�result�of�a�change�in�the�Buzz.

� “Quantifying�the�Buzz�Effect”�means�we�want�to�put�a�number�on�the�amount�of�that�increase�or�decrease.

� The�advantage�of�this�is�that�we�can�begin�to�put�a�dollar�value�on�salient,�publicly�known�events…such�as�refusing�to�take�government�bailouts.

� “Integrating�Social�Media�With�Loyalty�/�Defection�Models”�means�that�we�will:�� Extract�signals�of�future�vehicle�purchase�decisions�from�customer�comments found�

through�social�media�outlets AND� Capture�those�signals�in�the�form�of�predictive�variables�to�put�into�loyalty�models.� Those�variables�and�their�associated�model�coefficients�will�quantify�the�buzz�effect.� For�the�purpose�of�this�analysis�we�focus�our�efforts�on�Twitter.�

2

Page 3: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Warnings�and�Disclaimers

� We�will�do�our�best�to�reveal�trends,�patterns,�and�findings�without�showing�actual�numbers�(but�for�some�cases).��Hiding�/�changing�of�numbers�is�done�to�protect�the�innocent�(Ford�Motor�Company�especially).

� In�the�course�of�the�presentation�we�will�share�many�Ford�related�“tweets”.��These�will�be�actual�tweets.��They�will�not�be�censored�because�their�informal�nature�highlights�a�point�we�want�to�make�about�text�mining.��Please�try�not�to�be�offended.

� We�use�“off�the�shelf”�techniques�when�it�comes�to�categorizing�sentiment.� Our�expertise�is�in�modeling�and�predicting�customer�behavior�based�on�all�

available�and�relevant�customer�data.�� We�see�social�media�as�a�potentially�rich�source�of�customer�data,�and�those�data�

just�happen�to�be�free�form�text.

3

Page 4: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

One�More�Item�of�Note…

4

Page 5: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Background�on�Ford’s�Social�Media�Efforts…

� Measuring�the�“Consumer�Experience”� Alan�Mulally and�Apple…� The�Dealership�Experience:�Sales�and�Service� The�Ownership�Experience� How�do�people�share�experiences?��Traditionally�by�talking�to�each�other.��But�how�

much�today�is�done�through�Twitter,�Facebook,�Blogs?

� By�analyzing�the�comments�and�sentiment�expressed�through�Social�Media�outlets�can�we�glean�meaningful�insights�about�the�Ford�Consumer�Experience?��

� Can�we�make�inference�about�a�consumer’s�affinity�for�Ford…or�an�existing�customer’s�loyalty�to�Ford?

� If�no,�then�we’re�probably�not�trying�hard�enough.� Examples�next�2�slides.

5

Page 6: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Google�Twitter�Search:�Ford�Comments

� Search:�“My�Ford�Focus�is�great.”

� I�love�my�Ford�Focus,�but�not�so�much�Ford�Service�in�Northampton�Mass.�Thieves.

� Got�my�new�computer�yesterday�and�can't�wait�to�get�my�new�2012�Ford�Focus SEL�in�4�6�weeks!�23�Apr

� Am�test�driving�Hondas�and�Fords 7�Apr

� We’d�like�to�have�a�mechanism�for�intervening�here.��On�April�7�this�person�indicated�he�was�facing�a�choice�between�buying�a�Honda�and�buying�a�Ford.

� Does�this�mean�we�can�simply�scrape�Twitter�for�the�words�“test�drive”?��Seems�like�it�would�be�predictive�of�future�behavior…

6

Page 7: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Google�Twitter�Search:�Ford�Comments

� Search:�“I�don’t�like�my�Ford Escort.”

� The�ford�escort�texting�and�driving,�I�really�likemy�life�and�my�car,�please�don't�try�and�drive�into�us,�twice.�Close�call!

� My�old�'93�ford�escort�is�running�130k�and�runs�like�a�charm....�And�my�2003�ford ranger�truck�has�80k�without�problems.

� Again�this�seems�like�something�that,�if�captured�and�quantified�in�the�form�of�a�variable,�would�be�predictive�in�the�context�of�a�loyalty�model.

7

Page 8: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Google�Twitter�Search:�Ford�Comments

� Search:�“Ford,�government�bailout”� This�weekend�my�wife�and�I�purchased�a�FORD.�Why?�Because�they�chose�not�to�accept�the�government funded�bailout.

� Ford didn't�accept�the�government�bailout � that's�pretty�awesome.� Wait�#Ford pulled�the�ad�that�was�critical�of�the�#Obama bailout�but�is�now�running�one�that�jokes�about�drinking�and�driving?�

� GM�CEO�wants�higher�gas�tax.�Buy�a�Ford car�or�truck.�Please�RT

� This�is�an�example�of�how�capturing�“influencers”�could�be�very�important.��This�person�happens�to�have�340�followers�and�routinely�tweets�about�auto�related�topics.�

� So�the�effort�to�mine�Twitter�for�Ford�sentiment�extends�beyond�improving�the�loyalty�and�defection�models…but�the�title�of�this�presentation�does�not.��That�said,�we�will�discuss�how�we�are�affecting�marketing�programs�with�our�existing�knowledge�of�influencers.��

8

Page 9: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Background�on�Ford’s�Social�Media�Efforts…

� Measuring�the�“Consumer�Experience”� Alan�Mulally and�Apple…� The�Dealership�Experience:�Sales�and�Service� The�Ownership�Experience� How�do�people�share�experiences?��Traditionally�by�talking�to�each�other.��But�how�much�today�is�done�through�Twitter,�Facebook,�Blogs?

� By�analyzing�the�comments�and�sentiment�expressed�through�Social�Media�outlets�can�we�glean�meaningful�insights�about�the�Ford�Consumer�Experience?��

� Can�we�make�inference�about�a�consumer’s�affinity�for�Ford…or�an�existing�customer’s�loyalty�to�Ford?

� Yes!��So�what�can�we�do�capitalize�upon�good�sentiment�and�reverse�bad�sentiment?

9

Page 10: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Start�With�the�Current�Infrastructure� The�Ford�Motor�Company�has�a�customer�data�warehouse�that�collects�relevant�data�from�all�customer�

touchpoints,�“customerizes”�it,�and�applies�a�suite�of�predictive�models�that�are�used�for�targeted�campaigns.

� More�importantly�the�warehouse�is�connected�to�many�customer�facing�and�dealer�facing�operational�systems,�and�it�passes�important�information�about�customer�behavior,�both�past�behavior�and�predicted�behavior,�to�operational�systems�when�decisions�regarding�the�customer�have�to�be�made�in�real�time.

10

Page 11: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Fitting�in�to�the�Current�Infrastructure…

� Social�media�is�just�another�customer�touch�point.� The�text�we�mine�from�social�media�outlets�is�another�set�of�data�about�the�customer,�just�

like�the�call�center,�website,�or�the�customer�surveys.� We’d�like�to�use�that�data�just�like�the�rest�of�the�customer�data:�to�help�us�predict�customer�

purchase�behavior.

� In�some�sense�social�media�provides�a�source�of�unsolicited�surveys.

11

Page 12: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

A�Source�of�“Unsolicited�Surveys”…

� Why�do�we�survey�customers?��From�the�narrow�perspective�of�someone�who�predicts�customer�behavior,�the�graph�below�is�a�big�reason�why.

� How�much�do�we�spend�on�surveys?� Whatever�it�is,�our�feeling�is�that�if�we�can�establish�the�above�relationship�with�social�media�

sentiment�(use�it�as�your�X�axis),�and�cover�more�customers�for�less�than�what�we�currently�spend�on�surveys,�then�we�have�the�beginning�of�a�business�case�for�extracting�sentiment�from�social�media.

� What’s�more�compelling�is�that�relationship�between�a�customer’s�opinion�and�loyalty�holds�up�when�we�control�for�predicted�loyalty.��

� The�“Loyalty=1”�group�is�the�group�that�scores�in�the�lowest�20%�of�a�loyalty�model…a�model�built�without�survey�data.�

� The�results�remain�consistent�within�each�loyalty�tranche�so�much�so�that�customers�within�group�5�can�have�lower�repurchase�rates�than�those�in�group�3,�depending�on�survey�response.�

12

Page 13: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

� We�believe�that�there�are�three�“pillars”�for�the�business�case�to�actively�engage�consumers�through�social�media�outlets�(specifically�Twitter):1. Conquesting�new�customers2. Concern�resolution3. Voice�of�Customer

Introduce�Social�Media�as�an�additional�consumer�touch�point

1. Conquest$XX�mils�per�year

STRATEGY:

SUPPORTS�THESTRATEGY:

2. Concern�Res$XX�mils�per�year

3. VOC$XX�mils�per�year

A�Quick�Digression�on�Business�Cases…

13

Page 14: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

How�We�Use�Survey�Data�In�the�Models…

� Let�P�=�probability�a�Ford�customer�will�repurchase�another�Ford�upon�disposing�of�any�one�of�his�current�Fords.

� Logistic�regression�is�a�very�popular�way�to�model�and�predict�P.� ln[p�/�(1�p)]�=�b0 +�b1*x1 +�b2*x2 +�…�bn*xn� b0,�b1,�b2 are�parameter�estimates.��They�quantify�the�extent�to�which�x1,�x2,�…,�xn affect�the�

probability�of�repurchase.� x1,�x2,�…,�xn are�explanatory�variables,�e.g.�#�of�previous�Ford�purchase,�time�since�most�

recent�Ford�purchase,�miles�from�nearest�Ford�dealership,�etc…�

� Now�let�s1 =�1�if�“very�likely”,�0�otherwise� Let�s2 =�1�if�“likely”,�0�otherwise�…� Let�s5 =�1�if�“not�at�all�very�unlikely”,�0�otherwise� Refit�the�logistic�regression:

� ln[p�/�(1�p)]�=�b0 +�b1*x1 +�…�bn*xn +�bn+1*s1�+�bn+2*s2 +�bn+3*s3�+�bn+4*s4 +�bn+5*s5

Can�be�thought�of�as�the�VOC�(Voice�of�Customer)�Index,�but�it’s�based�on�just�survey�data,�which�may�only�be�available�on�10%�(roughly)�of�the�customers.This�is�a�nice�metric,�because�it,�by�design,�predicts�loyalty.

14

Page 15: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

How�We�Use�Twitter�Data�In�the�Models…

� Let’s�treat�the�Ford�customer’s�“tweets”�the�same�way�we�treat�survey�data.� Go�back�to�our�logistic�regression:

� ln[p�/�(1�p)]�=�b0 +�b1*x1 +�…�bn*xn +�bn+1*s1�+�bn+2*s2 +�bn+3*s3�+�bn+4*s4 +�bn+5*s5� And�let�t1 =�1�if�we�can�identify�a�“Ford�positive”�tweet�for�the�customer,�0�otherwise.� Let�t2 =�1�if�we�can�identify�a�“Ford�neutral”�tweet�for�the�customer,�0�otherwise.� Let�t3 =�1�if�we�can�identify�a�“Ford�negative”�tweet�for�the�customer,�0�otherwise.� Refit�the�model:

� ln[p�/�(1�p)]�=�b0 +�b1*x1 +�…�bn*xn +�bn+1*s1�+�bn+2*s2 +�bn+3*s3�+�bn+4*s4 +�bn+5*s5+�bn+6*t1�+�bn+7*t2 +�bn+8*t3

Can�be�thought�of�as�the�BUZZ�INDEX,�and�it�comes�directly�from�what�ford�customers�are�saying�on�Twitter.��This�metric�also,�by�design,�predicts�loyalty.��So�this�is�a�quantification�of�the�BUZZ�EFFECT.

15

Page 16: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Interpreting�the�“Twitter�Enhanced”�Model…

� ln[p�/�(1�p)]�=�b0 +�b1*x1 +�…�bn*xn +�bn+1*s1�+�bn+2*s2 +�bn+3*s3�+�bn+4*s4 +�bn+5*s5+�bn+6*t1�+�bn+7*t2 +�bn+8*t3

� When�we�fit�this�model,�we�get�an�intuitive�result:�� bn+6�>�bn+7 > bn+8 =>�good�tweets�lead�to�higher�loyalty�than�do�neutral�tweets,�

neutral�tweets�lead�to�higher�loyalty�than�bad�tweets.

� Not�as�intuitive�(but�interesting�nonetheless):�� All�three�parameters�are�greater�than�0��(implying�ANY�tweeting�is�better�than�no�

tweeting).� bn+8 (the�parameter�for�bad�tweets)�is�NOT�SIGNIFICANT.��There�is�not�a�sufficient�

volume�of�bad�tweets�to�support�a�significant�result.��The�large�majority�of�FLM�tweets�are�good.�

� We�think�that�the�upshot�of�all�of�this�is�that�tweeting�about�Ford,�irrespective�of�sentiment,�signifies�a�high�level�of�customer�engagement.��This�has�implications�beyond�our�efforts�to�better�predict�loyalty.

16

Page 17: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

The�Buzz�Variables�Improve�the�Loyalty�Model:�So�What?

� More�data�and�better�data�yield�models�that�do�a�better�job�of�“separating”�loyalists�and�non�loyalists.��

� One�way�this�manifests�itself:�ranking�the�population�of�customers�with�a�better�model�will�yield�higher�repurchase�rates�in�the�top�decile (or�demi�deciles…depending�on�how�many�groups�you�want�to�establish),�and�lower�repurchase�rates�in�the�bottom�decile.

� So�a�marketing�campaign�that�increases�everyone’s�likelihood�of�repurchase�by�15%�(not�an�uncommon�number),�does�so�on�a�larger�base�of�loyalists�within�the�top�decile,�and�thus�creates�more�incremental�sales�for�the�same�amount�of�mailings.

1 2 3 4 5 6 7 8 9 10

Repu

rcha

se�Rate

Model�DecileLow�Loyalty�to�High

Old�ModelModel�w/VOC�&�Buzz

� Say�the�difference�between�these�two�bars�is�200�bps.�� Some�of�the�incremental�sales�from�the�campaign�

noted�in�the�bullet�above�(top�decile only),�are�attributable�to�having�a�better�model.��

� How�many?�.02�*�.15�*�top�decile population� If�the�population�of�interest�is�250,000�customers,�then�

the�impact�of�the�better�model�is�750�incremental�repurchases.

� If�a�repurchase�is�worth�$5,000�profit,�then�the�case�for�the�“buzz�variables”�is�substantial:�$3.75�million.

17

Page 18: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Why�We�Will�Regularly�Re�Fit�the�Buzz�Index� We�have�several�reasons�to�believe�that�results�may�change�when�we�re�fit�our�

enhanced�loyalty�model�(twitter�sentiment�data�being�the�enhancement):1. The�number�of�Tweeters�is�increasing�all�the�time.��Ford’s�customer�email�capture�isn’t�great�

but�it�is�improving,�and�there�is�evidence�that�Ford�customers�are,�relatively�speaking,�very�active�on�Twitter.

2. Attribution�of�tweets�to�customers�is�difficult�and�unsure;�finding�the�Twitter�names�of�Ford�customers�is�difficult�and�painstaking.

3. Classifying�the�sentiment�of�Tweets�is�an�imprecise�exercise,�especially�when�using�off�the�shelf�tools�and�software.

4. The�content�of�the�“Ford�tweeting”�population�leads�to�potentially�biased�results;�it�is�biased�toward�a�demographic�that�naturally�tends�to�be�less�Ford�loyal:

� Females� Young� Used�vehicle�owners� Appear�to�be�more�service�loyal,�which�is�a�good�thing

5. The�tweeting�population�also�happens�to�be�geographically�biased,�but�this�does�not�concern�us�as�much�as�#4.

18

Page 19: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Re�Fit�the�Buzz�Index:�The�Increasing�Number�of�Tweeters�

� According�to�GIGAOM�(http://gigaom.com/),�Twitter�had�175�million�users�in�December�2010,�and�was�growing�by�370,000�new�users�every�day.��Also�as�of�12/2010:

� 65%�of�those�users�lived�outside�the�US.� Roughly�6%�of�all�Americans�were�active�on�Twitter.��More�recent�studies�indicate�the�number�is�9%�10%�(or�

13%�of�internet�users).

� Of�the�Ford�customers�active�as�of�12/2010�(who�had�a�valid�email),�we�were�able�to�find�9%�of�them�active�on�Twitter.

� Not�surprisingly�we�have�found�the�number�of�Ford�related�tweets�to�be�increasing�over�time.��� Good�and�neutral�tweets�have�increased�whereas�bad�tweets�have�stayed�flat.��Seems�like�a�good�thing,�but�we�

have�some�thoughts�on�comment�classification.�

Frequency�of�Ford�Related�Comments�Found�on�Twitter

0

10,000

20,000

30,000

40,000

50,000

60,000

May�08

Jul�0

8

Sep�08

Nov�08

Jan�09

Mar�09

May�09

Jul�0

9

Sep�09

Nov�09

Jan�10

Mar�10

May�10

Jul�1

0

Sep�10

Nov�10

Jan�11

Mar�11

Ford�Related�Twitter�Comments�"Binned"

0

1000

2000

3000

4000

5000

6000

May�08

Jun�08

Jul�0

8Au

g�08

Sep�08

Oct�08

Nov�08

Dec�08

Jan�09

Feb�09

Mar�09

Apr�09

May�09

Jun�09

Jul�0

9Au

g�09

Sep�09

Oct�09

Nov�09

Dec�09

Jan�10

Feb�10

Mar�10

Apr�10

May�10

Jun�10

Jul�1

0Au

g�10

Sep�10

Oct�10

Nov�10

Dec�10

Jan�11

Feb�11

Mar�11

Postive,�Negative

0

10000

20000

30000

40000

50000

60000

Neu

tral

Negative

Positive

Neutral

19

Page 20: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Re�Fit�the�Buzz�Index:�Attributing�the�Tweets�� Twitter�names�can�be�found�if�you�can�supply�an�email.

� With�an�email�address�you�can�find,�through�the�Twitter�API,�a�Twitter�name,�a�first�name,�and�a�last�name�associated�with�that�email�address.��It�will�not�return�the�email.

� Ford�has�11�million�bought�new�still�retained�customers.��We�have�emails�on�several��million�of�them.

� We�cannot�run�several�million�emails�through�the�Twitter�API.��Even�if�they�could�be�processed,�we�would�not�be�able�to�get�back�the�email.��We�would�only�get�back�the�thousands�of�Twitter�names�associated�with,�but�not�matched�to�the�emails.�

� The�only�sure�way�to�attribute�emails�to�Twitter�names�is�to�go�through�the�API�one�email�at�a�time. Can�we�off�shore�this?��We�can,�but�we�still�will�not�be�done�until�2013�if�we�go�this�route.��So�we�had�to�be�more�clever…

� Whatever�method�we�choose,�we�need�to�recognize�the�Twitter�name�as�useful�customer�data,�and�as�such,�store�it�in�the�data�warehouse.��Next�slide…

20

Page 21: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Re�Fit�the�Buzz�Index:�Attributing�the�Tweets�– You�Must�Retain�Data�� What�comes�out�of�the�attribution�process�is�a�table�that�looks�like�this:

� We�have�another�process�(using�RADIAN6)�that�scours�Twitter�for�comments�that�contain�words�in�our�“start�list”�(e.g.�Ford,�Lincoln,�Mercury,�Taurus,�Mustang,�Fusion,�etc…).���It�produces�a�table�that�looks�like�this:

EMAIL TWITTER_NAME FIRST_NAME [email protected] Keith Shields

[email protected] ronedog Roni [email protected] Mindy Deatrick

TWITTER_NAME COMMENT SENTIMENTronedog My�new�Ford�Focus�is�also�imported�from�Detroit. Good

21

Page 22: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Re�Fit�the�Buzz�Index:�Attributing�the�Tweets�– You�Must�Retain�Data�

Data�Creation

Customer�Data�Warehouse�

Twitter

� We�have�to�integrate�this�into�existing�data�warehouse�processes�(which�should�be�easy�enough,�if�we’re�treating�this�like�just�another�source�of�customer�data):

EMAIL TWITTER_NAME FIRST_NAME [email protected] Keith Shields

[email protected] ronedog Roni [email protected] Mindy Deatrick

Customer�Touchpoint

TWITTER_NAME COMMENT SENTIMENTronedog My�new�Ford�Focus�is�also�imported�from�Detroit. Good

22

Page 23: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Re�Fit�the�Buzz�Index:�Classifying�the�Tweets…�� Prepackaged�comment�binning�algorithms�are�not�as�accurate�as�we’d�like…they�result�in�

a�high�instance�of�inappropriate�comment�binning�(“Positive”,�“Neutral”,�or�“Negative”).��Here�are�some�actual�examples�of�inappropriately�binned�comments:

� Positive:�“Classic�Car�For�Sale�2001�FORD�EXPLORER�� Mt.�Royal�NJ:�Runs�and�Looks�Great!!!”� Negative:�“You�have�insulted�my�Ford�Fiesta,�shame�on�you.”�AND�“Just�drove�a�Ford�Fiesta�

getting�30�mph.�Not�bad!”�AND�“Just�dropped�my�car�off�at�the�FORD�Dealership�.�I�want�a�FORD�Fusion�soooooo BAD.”

� Neutral: “Ford�Escape!!!”�AND�“Smart,�easy,�&�fun�ride�with�the�new�Ford�Focus�with�Ford�Sync's�help!”

� In�order�for�us�to�get�the�intuitive�results�we�showed�on�slide�12�we�had�to�depart�from�sentiment�classification�algorithms�and�do�a�“brute�force”�classification.� Interns,�Cornerstone�Schools,�Detroit,�Mi.�(http://www.cornerstoneschools.org/)� There�are�other�inexpensive�ways…all�of�which�we�believe�to�be�more�accurate�than�existing�

“machine�intelligence”…albeit�not�as�scalable:• Existing�call�center�personnel• Mechanical�Turks�(https://www.mturk.com/mturk/welcome)

23

Page 24: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Re�Fit�the�Buzz�Index:�The�Biased,�But�Changing,�Population�of�Tweeters…�

� See�the�graph�on�the�right.��The�fastest�growing�population�of�tweeters�is�18�34�year�olds.��About�56%�of�tweeters�are�34�years�or�younger.

� The�challenge�for�Ford:�the�median�age�for�Ford�customers�is�well�above�34,�despite�some�recent�strong�entries�in�the�small�car�market.��

� The�most�“tweeted�about”�Ford�is,�not�surprisingly,�the�Focus.

24

Page 25: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Re�Fit�the�Buzz�Index:�Geographically�Biased�Population�of�Tweeters…�

� The�numbers�represent�how�much�higher�the�Twitter�use�per�capita�is�in�that�state�versus�the�nation�as�a�whole.�

� For�example:�if�the�national�usage�rate�is�10%,�then�Michigan�is�11%�lower�than�that:�.10�� .11(.10)�=�8.9%.�

25

Page 26: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Given�the�Worries�About�Attribution,�Classification…�

� The�real�opportunity�could�lie�in�the�business�experts�using�intuition�and�common�sense�to�tailor�campaigns�and�programs�to�tweeters�based�on�their�most�recent�comments.

� This�would�only�rely�upon�a�good�mechanism�for�scraping�relative�comments�from�Twitter�and�reacting�procedurally�and�appropriately.

� If�we�look�at�the�comments�as�unsolicited�survey�responses�we�see�opportunities�for�customized�offers�and�programs�(no�models�needed�– the�comment�reveals�the�customer’s�intent):

� Private�Sales�Offer�and/or�Pre�approval:�“Just�dropped�my�car�off�at�the�FORD�Dealership.� I�want�a�FORD�Fusion�soooooo BAD.”

� Rewards�Program�Offer:��“My�moms�taking�me�to�get�this�Ford�explorer�in�the�mornin tho i�should�have�a�new�whip�before�July�then�im haulin ass�to�the�A”

� Offer�for�trade�in:�“Ford�focus�sucks.�Very�uncomfortable�vehicle.”� Offer�for�service�discount�/�extended�warranty:�“My�car�is�running�rough�and�

keeps�blowing�the�injector�and�on�plug�coil�fuses,�its�a�2006�3.0�V6�Ford�Fusion.�HELP!!”�

26

Page 27: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

� We�believe�that�there�are�three�“pillars”�for�the�business�case�to�actively�engage�consumers�through�social�media�outlets�(specifically�Twitter):1. Conquesting�new�customers:�pay�attention�to�influencers2. Concern�resolution3. Voice�of�Customer

Introduce�Social�Media�as�an�additional�consumer�touch�point

1. Conquest$XX�mils�per�year

STRATEGY:

SUPPORTS�THESTRATEGY:

2. Concern�Res$XX�mils�per�year

3. VOC$XX�mils�per�year

So�We�Revisit�Our�Three�Pillars…

27

Page 28: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Conquest�New�Customers:�Influencers

Fact: 93.6%�of�Twitter�users�have�less�than�100�followers,�while�98%�of�users�have�less�than�400�followers.�Meanwhile,�1.35%�of�users�have�more�500�followers,�and�only�0.68%�of�more�than�1,000�followers.�

28

Page 29: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Conquest�New�Customers:�Influencers

Fact: As�Twitter�users�attract�more�followers,�they�tend�to�Tweet�more�often.�This�is�particularly�evident�once�someone�has�1,000�followers�the�average�number�of�Tweets/day�climb�from�three�to�six.�When�someone�has�more�than�1,750�followers,�the�number�of�Tweets/day�rises�to�10.�

Fact: A�small�group�of�Twitter�users�account�for�the�bulk�of�activity.�Sysomos discovered�that�5%�of�users�account�for�75%�of�all�activity,�10%�account�for�86%�of�activity,�and�the�top�30�account�for�97.4%�of�activity.

29

Page 30: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Conquest�New�Customers:�The�Opportunity

� Find�out�who�is�expressing�in�market�sentiment�and�send�them�a�targeted�offer.� We�estimate�that�through�Twitter�alone,�roughly�35,000�customers�per�year�express�inclination�

to�buy�Ford.�� Marketing�Associates�built�the�process�to�find�the�tweets�and�measure�the�back�end�results.��

Here�are�some�great�“Focus�tweets”…just�from�the�last�couple�of�weeks:� “I�think�I�want�a�2012�Ford�Focus.” 3/12/2012� “I�want�a�2012�ford�focus...just�because�it�parks�itself.�:\”��3/20/2012� “2012�Ford�Focus�ST�or�2013�Dodge�Dart?�I�dunno,�the�Dodge�Dart�it�is�just�a�Neon�that�

mated�with�an�Alpha�Romeo�but�the�Focus�ST�looks�pretty�promising�and�I�always�loved�my�buddies�SVT�Focus..�Decisions..�Decisions..”��3/1/2012

� Applying�a�result�from�an�analysis�of�"handraiser campaigns",�we�assume�15%�of�the�35,000�will�purchase�FLM.��This�is�35,000�*�15%�=�5,250�sales.

� Assuming�20%�lift�from�a�targeted�offer�to�in�market�customers�(not�an�uncommon�number),�we�estimate�that�a�conquesting campaign�directed�at�in�market�"social�media�leads“.��This�is�5,250�*�.2�=�1,050�incremental�sales.

� INTEGRATION�will�be�through�the�customer�data�warehouse�and�EXECUTION�through�the�concern�resolution�center.

� At�this�point�we�won’t�trouble�you�with�another�infrastructure�diagram.

30

Page 31: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Some�Recommendations…

� Treat�social�media�as�another�source�of�relevant�customer�data.��� Comments�about�your�product�are,�in�some�sense,�unsolicited�surveys.��They�can,�like�surveys�

do,�improve�your�ability�to�predict�the�behavior�of�your�own�customers.

� Pay�careful�attention�to�the�integration�of�social�media�data.��Integration�requires�“customerization”,�so�subsequent�customer�behavior�can�be�tracked.

� Attributing�comments�to�customers�is�tricky.��It�can�also�be�painstaking.��The�good�news�is�that�it�can�be�done�cheaply.

� Correct�classification�of�comments�is�essential�to�understanding�the�true�signal�in�the�comments.��

� The�most�accurate�means�of�classification�may�also�be�the�least�scientific:�have�an�English�speaker�(who�preferable�understands�colloquialisms)�read�the�comments,�and�bin�them.

� Categories�can�be�“good”,�“bad”,�“in�market”,�“service�issue”,�or�whatever�aligns�with�the�differentiated�treatments�and�offers.

� Retain�data,�and�integrate�intelligently�into�a�data�warehouse.��Test�and�measure�several�tactical�approaches�to�customers�and�prospects�who�are�commenting�about�your�products.

31

Page 32: AIMS2012 Marketing Associates Quantifying the Buzz Effect.

Copyright�©2012�Marketing�Associates�LLC.�All�rights�reserved.

Decision�Sciences

Quantifying�the�"Buzz"�Effect:�Integrating�Social�Media�with�Loyalty�&�Defection�Models

Thanks�for�your�time�and�attention.

Questions?��

Want�a�copy�of�this�presentation?Text�KEITH to�30241

32

Marketing�Associates�Alerts:�Receive�up�to�2�msgs per�month.�Msg&Data rates�may�apply.�Text�STOP�to�stop.�For�more�info,�email�[email protected].