Digital travel summit channel attribution 2014 04-01

Post on 30-Jan-2015

323 views 0 download

description

Algorithmic marketing channel attribution: why it's important, who is doing it, how it will change your business and how to decide what method is right for you.

transcript

C O N F I D E N T I A L

Marketing Channel Attribution Modeling To Improve Conversion Rates

Jonathan Isernhagen

Digital Travel Summit Day 1

Digital Travel Summit

4/1/2014

C O N F I D E N T I A L

Agenda

1) Marketers’ mission

2) Calculating channel ROIa) spendb) VCM

3) Attributiona) algorithmic

i. data assemblyii. types of models

b) arbitrary

4) Options for action

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

CEO’s/Our duty to the (publically-traded) company

• Maximize brand awareness/sentiment?

• Maximize client loyalty?

• Maximize co-worker satisfaction?

• Drive traffic to the site?

• Maximize shopper movement down the funnel?

• Maximize transaction volumes?

• Maximize shareholder value?

• Maximize individual channel ROI?

• “Give back" to the community through charitable actions?

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

Example: Apple, Inc.

“Apple's Board of Directors oversees the Chief Executive Officer and other senior management in the competent and ethical operation of Apple on a day-to-day basis and assures that the long-term interests of shareholders are being served.

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

Source: http://investor.apple.com/governance.cfm

C O N F I D E N T I A L

Increase spending until $1 out brings $1 back…

….spend-spend-spend-spend-spend-stop

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

…which is the point at which incremental ROI = 0%

- 6 -

ROI = (VCM – Spend) Spend

Abbreviation Term Definition

ROI Return on Investment Indicator of investment profitability. Positive = good.

VCM Variable Contribution Margin The amount of profit driven by a given transaction.

Spend Channel Spend The amount spent driving traffic to the site during the period in question

Calculated over a specified time period of investment and return.

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

Agenda

1) Marketers’ mission

2) Calculating channel ROIa) spendb) VCM

3) Attributiona) algorithmic

i. data assemblyii. types of models

b) arbitrary

4) Options for action

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

ROI = (VCM – Spend) Spend

Spend

- 8 -

Which spend do you include?

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

Marketing ≈ Farming

1)Seed = impressions

2)Transactions = fruit

- 9 -

…butmarketers plant

& harvest constantly.

C O N F I D E N T I A L

spendspendspendspend

How changes in spend can mess up ROI

- 10 -

$

time

profit

impact impact

spend

impact

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

Jan Feb Mar

C O N F I D E N T I A L

Variable Contribution Margin (“VCM”)

- 11 -

ROI = (VCM – Spend) Spend

1) What is the value of each transaction?

2) Which channels deserve credit for which transactions?

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

VCM

= Transaction revenue - variable non-marketing expenses:

• Revenue: – Supplier Commissions; – GDS incentives; – Overrides (lumpy: average/booking must be assumed)– Media (not transaction-driven, but has to be modeled in

somewhere)– Attached bookings / Lifetime value: try to gauge value without

double-counting

• Expenses: – Website hosting/capacity costs– Data processing expenses– Other expenses which vary by transaction or site activity volume

- 12 -jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

Which marketing treatment(s) triggered the purchase?

"Half the money I spend on advertising is wasted; the trouble is, I don't know which half.“

-John WanamakerFather of Modern

Advertising

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

“Hard to Know” ≠ “Random”

We may not find the correct answer, but the answer does exist.

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

Google Analytics attribution options

Google offers several attribution mechanisms out of the box

Choosing the wrong model will have consequences.

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

How badly do you want to know?

Raylan: You'll pay to find that out.Boyd: What are you packing?

C O N F I D E N T I A L

Choosing your attribution strategy

Do you want/need true attribution*?

Do you havedata guys?

Do they havebandwidth for this?

Yes

No

No

Use your sitemetric solution’s

attribution

Yes

YesHire a vendor

Can you accessyour data?

Yes

No

*Do you have:1) Large enough budget?2) Multiple channels?3) Belief in ROI “knowability?”

DIY

No

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

Agenda

1) Marketers’ mission

2) Calculating channel ROIa) spendb) VCM

3) Attributiona) algorithmic

i. data assemblyii. types of models

b) arbitrary

4) Options for action

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

• Sends• Opens

Collecting the necessary data

Analysis space

• Clicks• Visitors

• Transactions• VCM• Media impressions

Site metrics tool

Back office system

• Channel-specific phone #s

Email service provider

Call Center IVR

• Impressions• Clicks• Spend

Display ad server

e.g. SAS, Revolution Analytics,SPSS, Teradata Warehouse Miner

SEM bid management tool

• Impressions• Clicks• Spend

• GRPs• Spend

Television plan

• Impressions• Clicks• Spend

Meta search feeds

Spend

Manual spend entry table

C O N F I D E N T I A L

Connecting the necessary data

Transactions Profiles Customers

Sessions

∞ 1∞ 1Clicks:TransactionsClicks ∞ 11 ∞

∞ 1∞ 1

Calls

Sends

Opens∞1

∞1

Email

IVR

GRPs

TV planImpressions

Ad server

Impressions

∞ 1

∞ 1

∞ 1

Back officeSite monitoring tool CRM system

1 ∞

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

Profile deduplication is crucial

Profiles Customers∞ 1

CRM system

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

1) App registration information tied to desktop profile

2) Logged-in customer information on multiple devices

3) E-mails sent to same address opened on multiple devices

4) Third-party services with network visibility

C O N F I D E N T I A L

Impressions are also important

This example overcredits impressions (excludes other channels) but it gives some idea of the sub-surface portion of the

impressions iceberg.

C O N F I D E N T I A L

Forrester 2012Q2 survey of attribution vendors

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

24

Algorithmic attribution per Visual IQ:Top Down (MMO) & Bottom Up (Fractional Attribution)

SUMMARY LEVEL DATA USER LEVEL DATA

CHANNELS Offline + digital cross channel Digital media channels

ROLE Strategic: Optimize spend across channelsTactical: Generates granular

media recommendations

OUTPUT Cross channel insight Full fractional attribution

CAPTURES Seasonality and external factorsInterplay between digital touch

point and channels

PREDICTABLE GRANULARITY

Conversions at aggregate level Propensity to convert at user level

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

“Given all the history we know, how likely is this shopper to convert soon?”

Datasong answers with a 2-stage model:

-Are Brand shoppers (BSEM, BSEO, DTI) likely to book? Yes

-Did your last TV spot cause their brand loyalty? Maybe

Algorithmic attribution per DataSong: Survival Modeling

Model 2Accuracy 81%

Model 1Accuracy 68%

0 0.10.20.30.40.50.60.70.80.9 10

0.5

1

NonConverters Converters

0 0.10.20.30.40.50.60.70.80.9 10

0.5

1

NonConverters Converters

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

Algorithmic attribution per DataSong: Survival Modeling

1) Axes show the time since last channel exposure2) Dots represent converters vs. nonConverters3) Orange line represents the relationship between 2

variables, e.g.:a) time since last email andb) time since last affiliate visit

4) Responders: everyone above the line. We count the folks above the line and see what our accuracy is

5) Attribution: once satisfied with a model(s), we’d:a) take a given order, b) see the time since last email and affiliate and based on

the timing, and;c) where we are on the orange line, so;d) we have a means to allocate which was more causal.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.5

1

NonConverters Converters

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

Survival Modeling: DIY

Recommended by one of our PhD statisticians.

“SAS usage is not necessary.”

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L- 28 -

Algorithmic attribution per Google: Interaction Method (“Shapley value”)

Brand SEM Email

2%

Conversion

Brand SEM Email

3%

Conversion

Every combination (“coalition”) of clicks is tested.

“How important is each player to the overall cooperation?”

http://en.wikipedia.org/wiki/Shapley_value

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

Display

C O N F I D E N T I A L

Running the process

1) For each site visitora) Assemble visit historyb) Create variables to represent:

i. Channel impressionsii. Channel clicksiii.Past purchases

2) Regress or use machine learning algorithma) Ascertain which channel touches predict bookingb) Give VCM credit to causal channels

3) Calculate ROIa) Use each channel’s VCM and spendb) Where ROI is positive, spend upc) Where ROI is negative, cut spend or change tactics

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

Output

Channel* Desktop Tablet Mobile

Brand 20%

Brand SEM 62% 51% 38%

Display -5% -12% -7%

Display - Retargeting 26% 25% 29%

Email 250%

Meta search 18% 22% 10%

Non-brand SEM -30% -18% -40%

SEO 500% 520% 390%

Social media -5% -15% 15%

- 30 -

*of impression/click, not necessarily of consumer conversion

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

Agenda

1) Marketers’ mission

2) Calculating channel ROIa) spendb) VCM

3) Attributiona) algorithmic

i. data assemblyii. types of models

b) arbitrary

4) Options for action

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

“Round up all the clicks!”

• Offline bookings• Multiple profile activity• Impressions (Display or Television/Radio/etc)• Cleared-cookie clicks

- 32 -

Leaves out:

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

55% of shoppers clear cookies every 3 months

Cookie clearing is a serious problem causing the true impact of online channels to be massively understated.

“The Consumer Side of Analytics” - Forrester/Jupiter Research 2009

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

Bracketing

1) Assume for the sake of argument that one of the site tool’s attribution schemes reflects reality, but you don’t know which one.

2) Find (or create) the scheme that is most and least generous to digital channels. a) Last click: usually the least generous to digital channelsb) 3-tier: a scheme we created to favor NBSEM and Meta/TravelSearch

3) Assume that the true attribution (& ROI) lies between these extremes

4) Adjust spend and see which scheme better predicts reality.

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

Tactical decision-making

Channel Easy decisions Hard decisions Tactics

Brand

Advertising heavily in target-rich geographies before/during periods of high demand.

• national vs. regional• constant vs. bursts• more air time vs fresh

creative

• A/B test across “identical” regions, observe rise in brand transactions.

• Get Comscore to measure brand awareness/intent after broadcast.

SEMSpending to capacity on Brand SEM

Whether to spend up on marginal NBSEM keywords

A/B test whether ↓ BSEM spend is offset by ↑ BSEO volume

Email

Emailing up to 1x/week, hitting each site visitor with the message he or she is most likely to convert from.sending messages with exceptional deals.

• Whether to upgrade to a rules-base ESP system.

• Whether to saturation-bomb your e-mail list.

• Whether to push a loss-leading promo code.

Split the list, quantify lifetime value, A/B test:• Unsubscribe impact

of saturation bombing• Uplift impact of offer

personalization.• Long-term behavior

of code recipients

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

Take-away lessons

- 36 -

1) Be intentional about your attribution decisions, and creative about how to test your assumptions.

2) For top competitors in each vertical, the question is no longer “should we do algorithmic attribution?” but rather “How can we operationalize our attribution model?”

3) If your data aren’t yet in shape to support attribution & other ad hoc analyses, get that project underway.

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

Attribution options - vendors

Provider MethodCost

Timing Comments1x monthly

Adobe Attribution model

Varies - 4-10 wksConsulting arm willing to address attribution questions

as one-off modeling projects.

Adometry Econometric model

Low High 12 wksForrester-acclaimed industry co-leader.

Highly display-focused.

Datasong

Attribution model w/Customer

scoring

Med

Med

6 wks

Comprehensive model which pulls in both online and offline spend impacts.

MedPer-customer demand curve to predict the expected

return on every campaign

Google Analytics

Interactions model

- - nowHas delivered digital-only channel interactions model

to beta clients.

Gauge LLC Simple regression

High - 10 wksGeneral consulting analysts, not on the efficient

frontier for this deliverable.

Kenshoo Unknown algorithm

- Med 4 wksPrimarily targeted towards SEM spend, but can be

expanded to comprise other digital channels

Visual IQEconometric

and Attribution models

High - 4-8 wksEconometric model. Offers cross-channel allocation

guidance, including offline spend.

Med-High

- 3-6 wksAttribution model. Offers within-channel campaign-

level spend guidance. No offline.

C O N F I D E N T I A L

Funnelism doesn’t drive investment decisions

jonathan.isernhagen@travelocity.com @jon_isernhagen www.linkedin.com/in/isernhagen

C O N F I D E N T I A L

Facebook visitor tracking

Visitor action Volume Trackable Useful

Ad link click Small Yes No

Sponsored news feed click Small Yes No

Organic click Small Yes No

Ad impression Large Yes Yes

Sponsored news impression Large No Yes

Organic impression Large No Yes