Signal-based Full Funnel Playbook June 2018.pdf

Signal-Based Full Funnel Playbook A better approach to full funnel campaigns on Facebook

Agenda

1. A NEW KIND OF FUNNEL

4. CLIENT SCENARIOS

2. BEST PRACTICES

5. LEARNING PATH

3. MEASUREMENT

6. LEARNING AGENDA

A New Kind of Funnel

Full Funnel Definition

Advertising with multiple messages to accomplish complementary goals across the customer lifecycle.

Full Funnel campaigns try to do 3 things: Audience

Group people based on where they are in the customer lifecycle

Creative

Craft relevant messages for each group

Optimization

Optimize for a range of events across the customer lifecycle

Traditional Full Funnel campaigns are planned as Brand + DR, which can lead to sub-optimal campaign designs T R A D I T I O N A L F U L L F U N N E L L I M I TAT I O N S •

Brand

•

•

DR

•

• Brand typically includes:

• Broad audience • Video assets • Reach optimization

DR typically includes:

• Narrow audience • Static assets • Conversion optimized

Brand optimal + DR optimal ≠ Full Funnel Optimal Sequencing brand and DR constrains delivery and can result in increased costs Phasing brand and DR also constrains delivery and can result in increased costs Different audiences and optimization leads to low overlap across campaigns “Personalization” efforts can lead to overstratification and result in delivery issues or increased costs

Signal-based Full Funnel campaigns are planned around signals, which enables high-performance campaign designs SIGNAL BASED FULL FUNNEL THEMES •

•

•

•

•

•

Audience Stratification: use intent signals to group people into different stages of the customer lifecycle Multiple Objectives: blend auction objectives to achieve a desired range of business outcomes from awareness to LTV. Content Ecosystem: craft a variety of formats and messages to enable dynamic personalization at scale Intent Signals: use the full spectrum of intent signals from across the customer lifecycle for targeting and optimization Campaign Coordination: plan campaigns together as a system and take into account interactions (e.g. overlap) Component/System Measurement: measure individual components with a variety of diagnostic KPIs, measure overall system with a strategic KPI like ROI.

Signal-based full funnel shifts from planning campaigns based on assumptions to data driven, system development

VERSUS

HUMAN ASSUMPTION

MACHINE LEARNING

Comparing Traditional and Signal-based Full Funnel Campaigns TRADITIONAL FULL FUNNEL •

•

Create a plan based on assumptions that determines who gets which message based on a limited number of factors Campaign is designed to deliver in sequence and take an individual person from awareness to consideration.

SIGNAL-BASED FULL FUNNEL •

•

Design a system that uses signals to determines who gets which message based on a wide range of factors Campaign is designed to deliver based on signals that indicate a person's position in the customer lifecycle.

•

Minimal audience stratification, typically 2 groups

•

Audience stratified across customer lifecycle

•

Brand and DR assets created in silo, not coordinated, and delivered to isolated brand and DR audiences

•

Content ecosystem with a variety of coordinated formats and messages delivered based on signals

•

Single optimization for each of Brand and DR campaigns

•

Multiple optimizations to balance desired outcomes

•

Lower-funnel signals only, mostly for DR campaign

•

Full spectrum of signals from across customer lifecycle

•

Brand and DR campaigns not coordinated

•

System of coordinated campaigns

•

Measure overall performance with DR efficiency metrics

•

Measure each component and overall system independently

Illustrative media setup of signal-based full funnel campaigns

NOTE: Stylistic only, not a formal media recommendation – individual businesses will have very different setups TYPE

CAMPAIGN

AUDIENCE

OPTIMIZATION

CREATIVE

LOWER FUNNEL ACTIONS

LOWER FUNNEL ACTION LAL SITE VISITOR RT

LOWER FUNNEL ACTION

LINK, CAROUSEL

CONVERSIONS

CONVERTER LAL LOWER FUNNEL ACTION RT

CONVERSION

LINK, LEAD GEN

REACH

BROAD INTERESTS CRM LAL 5%

REACH, VIDEO VIEWS

VIDEO, CAROUSEL

GENERATE

LANDING PAGE LAL 3% ASSET ENGAGER RT

LANDING PAGE

CAROUSEL, LINK

CAPTURE

CONVERTER LAL 1% LANDING PAGE RT

CONVERSION

LINK, LEAD GEN

AWARENESS

BROAD INTERESTS

REACH

VIDEO, STORIES

INTEREST

CRM LAL 10% CONVERTER LAL 10%

LANDING PAGE

COLLECTIONS

EVALUATION

PRODUCT PAGE LAL 5% LANDING PAGE RT

PRODUCT PAGE

CAROUSEL

INTENT

SHOPPING ACTION LAL 3% PRODUCT PAGE RT

SHOPPING ACTION

CAROUSEL, LINK

ACTION

CONVERTER LAL 1% SHOPPING TOOL RT

CONVERSION

DYNAMIC, LEAD GEN, LINK

LOYALTY

VALUE BASED LAL 2%

VALUE OPTIMIZATION

DYNAMIC, LEAD GEN, LINK

BASIC

INTERMEDIATE

ADVANCED

Best Practices

Do’s and Don’ts for Signal-based Full Funnel campaigns DO N T S - THI N G S TO AV O I D •

Don’t: just combine Brand and DR campaigns –

•

•

•

Constrains delivery and can increase costs (CPM/CPA)

Misses ad inventory and can increase costs (CPM/CPA)

•

Don’t: over-stratify your audience or ad sets –

Reduces auction efficiency and can increase costs (CPM/CPA)

Do: Use signals to stratify audiences –

Don’t: phase Brand and DR campaigns –

•

•

Lack of coordination leads to sub-par results

Don’t: sequence Brand and DR messages –

DOS - THINGS TO TRY

Do: Optimize for multiple events –

Mitigate CPA increases as budgets increase, especially for infrequent lowfunnel events

–

Promote the path to purchase by driving mid- and upper-funnel events

Do: Craft a content ecosystem –

•

Use a variety of formats/messages to enable personalization

Do: consider allowing the system to optimize –

•

Let signals determine who sees which ad

Let the system decide who receives each message vs creating distinct groups

Do: System/component measurement –

ensure each campaign is doing it’s job, and that the overall system is performing as well

Don’t: just combine Brand and DR campaigns Common expectation The ”brand” campaign (typically a R&F or awareness/video view optimized campaign) will increase a person’s propensity to buy, which will then be reflected by increased conversions or sales when compared to a solely DR cell What problems do we find? Low audience overlap – The “Brand” campaigns often optimize for efficient reach, which leads to low overlap with “DR” campaigns that optimize toward people who are predicted to convert. The overlapping audience ends up being very low. Not complementary – “Brand” and “DR” are optimized in silos to achieve different goals. Missing middle – Campaigns that only combine awareness and conversion campaigns fail to effectively prospect the mid-funnel Low sensitivity – Brand optimizations (BAO, VV, Reach) often move awareness or favorability, but may be difficult to measure with DR KPIs due to lower frequency (but higher reach)

DR

Brand

4%-7%* *See also these studies on overlap • https://fburl.com/nxypzwf7 • https://fburl.com/iyew83ym

Don’t: sequence Brand and DR messages Common expectation By delivering a series of ads in a specified order (brand then DR), we will “prime” them for better DR performance later on. 2 general “sequencing” approaches are: 1. Utilizing the sequencing function in R&F buys 2. Retargeting engagement audiences (video views, canvas opens, messenger, etc) What problems do we find? When attempting this approach there are a number issues to contend with: • Sequencing often leads to delivery constraints that drive up cost because retargeting audiences get smaller at each step • Each consumer journey is unique, so enforcing a specific sequence may not capture each person at the right moment • Different audiences – people who need to know about your brand/product and people who are ready to convert are different • Wrong signals – watching a video is a terrible signal to know if someone is in market

1

2

3

Don’t: phase Brand and DR campaigns Common expectation By adjusting delivery such that we first deliver the entirety of the brand message before starting the DR campaign, the client believes they are maximizing the effect of their brand investment. They feel it makes sense to get the audience warmed up before being prompted to take action. What problems do we find? The reality is that there are people at many places in a client’s funnel at any given time. Because of auction dynamics, this approach will most likely hurt performance as it passes on opportunities to capture people who are likely to convert during the “brand” phase. It effectively reduces the inventory that is available for both the brand and DR objectives. Always on delivery

Phased delivery Brand (2 weeks)

Missed Brand opportunity

Brand (4 weeks)

Missed DR opportunity

DR (2 weeks)

DR (4 weeks)

Time

Time

NOTE: Budget allocation across can be fluid based on business objectives (e.g. heavy up on brand to raise awareness for launch)

Don’t: over-stratify your audience or ad sets

What problems do we find? While the approach is intuitively sensible, small audience sizes result in limited delivery per audience, reducing the amount of signal that our machine learning can optimize from. The same happens when having too many ad sets. The right sized audiences stratifies broadly enough to cater to varying bidding and messaging needs, but does not cripple our algorithms. A good place to start might be with 4 -5 different campaigns.

Performance

Common expectation Since a client can identify many nuanced signals that separate the audiences they are delivering to, they are inclined to divide their strategy up granularly with only minor differences in approach to each audience.

1 group

Group Count

100+ groups

Do: Use signals to stratify audiences •

•

•

Audience stratification should be designed based on the spectrum of intent. There are multiple sources for intent signals, but the best data will typically be sourced from the advertiser. Signal volume and quality are generally inversely related ​H IGHER

​P AGE

VIEW (DEFAULT)

​V IEW

CONTENT

​A DD

TO CART

​H IGHER

•

QUALITY

​I NITIATE CHECKOUT

​P URCHASE

VOLUME

Where possible, identify people who will never buy your product (recent purchaser, partner data, etc.) and exclude them from your targeting.

AUDIENCE

DESCRIPTION

SOURCE

ACTION

NONCUSTOMERS

PEOPLE WHO WILL NEVER CONVERT

CRM, PARTNER

SUPPRESS

BROAD

AGE, GENDER, INTEREST, ETC.

FACEBOOK, PARTNER

TARGET

LOOAKLIKE

BUILT ON KNOWN INTENT

WEBSITE, APP, CRM, PARTNER

TARGET

RETARGET

KNOWN INTENT

WEBSITE, APP, CRM

TARGET

Do: Optimize for multiple events to promote the path to purchase Campaign #1 Campaign #2

Ad exposure = Reach

Explore products = Conversion

Campaign #3

Campaign #5

Purchase = Purchase

Over-optimized for Conversions Campaign Y

•

Add to cart = Conversion

Campaign #4

Campaign X

•

Visit homepage = Conversion

Complementary Campaigns

Campaign Z

Campaign A

Reach

Campaign B

Campaign C

•

Reach

Value

Engagement

Value

Engagement

•

Conversions

Traffic

Conversions

Traffic

Conversion events may be too infrequent for effective optimization at desired budget level Layering additional optimization signals can help promote the path to purchase by generating site traffic, content views, add to carts, and other mid-funnel events that may eventually lead to a purchase By having multiple campaigns we’re allowing the system to decide which campaign is relevant for each user and keep action rates high based on their level of intent. We want to leverage any insight the client has into the expected path to purchase, but need to test in order to map the most efficient multiple objective systems.

Do: Craft a content ecosystem with a variety of formats/messages

Formats

Messages

Let the system find what message best resonates with individuals across platforms. Instead of multiple messages in one creative, have multiple creatives each with one message, each formatted for all placements. Creative variations should live under a single ad set.

Do: consider allowing the system to optimize creative delivery Don’t always assume specific types of audiences will only be receptive towards certain messages

Hockey Interest

Basketball Interest

Football Interest

Consider letting the system decide which person receives each message

Do: System/component measurement to ensure each campaign is doing it’s job, and that the overall system is performing as well Individual Performance

System Performance

Measurement framework can be complicated. In general, here are 3 principles: • Measure each campaign based on campaign KPIs • Measure system on strategic business metric (incrementality measurement is preferred, e.g. iCPA, iROAS) • Determine if a new objective is adding value with a lift test (A+B vs. A+B+C)

Measurement

Measurement for complex systems

Ensure each campaign is doing it’s job, and that the overall system is performing as well

Individual Performance

System Performance

Measurement framework can be complicated. In general, here are 3 principles: • Measure each campaign based on campaign KPIs • Measure system on strategic business metric (incrementality measurement is preferred, e.g. iCPA, iROAS) • Determine if a new objective is adding value with a lift test (A+B vs. A+B+C)

Why Is It Hard to Measure Brand’s Impact on Conversions? Brand’s immediate impact on conversion from one brand campaign?

YES

Awareness

Considerat ion

Control

Control

Incremental awareness Incremental conversions

Expecting higher CPA as a result of broader audience and branding creative

Incremental conversions Conversion % trending

Tradeoff between clean measurement and business goals/flexibility

3 -6 months holdout

Brand’s long-term impact on conversion?

Is conversion% higher when exposed to brand and DR?

YES/ NO NO

Considerati on + Acquisition

Acquisition Only

Control

Control

Awaren ess

Control

DR

A+D

Contr Contro ol l

Brand and DR campaigns usually have small overlap if audience and optimizations are different. Single-exposure opportunity logging does not create proper control groups for dualexposed test group

Now Let’s Look at Different Ways to Measure

Multi-Cell Lift to Compare Systems Compare Systems Cell A

Cell B

BAU

Challenger

Traditional Full Funnel

Signal-Based Full Funnel

Control

Control

Example Question: Does the new system of full funnel campaign outperform the current approach of full funnel?

Multi-cell Lift to Determine Value of Individual component Test value of upper-funnel Cell A

Cell B

Campaign A

Campaign B

Campaign C

Cell A Campaign A

(e.g. awareness)

(e.g. consideration)

Test value of mid-funnel

(e.g. awareness)

Campaign B

Example Question: What is the incremental value of adding a mid-funnel consideration campaign?

Cell B Campaign A

What is the incremental value of adding an awareness campaign to the existing mix?

(e.g. awareness)

Campaign B

(e.g. consideration)

(e.g. consideration)

Campaign C

Campaign C

Campaign C

(e.g. acquisition)

(e.g. acquisition)

(e.g. acquisition)

(e.g. acquisition)

Control

Control

Control

Control

27

Multi-Cell Lift to Compare Component Strategies Compare Audience Strategies Cell A

Compare Audience + Creative Strategies Example Questions:

Cell B

Cell A

Stratified Audience + MultiObjective

One broad audience + MultiObjective

Manually align audience and creative

One broad audience + Auctionoptimized creative

Control

Control

Control

Control

28

Does a stratified or broad audience strategy work better?

Cell B

Does aligning creative message with specific audience work better than one broad audience with auction-optimized creative? Does optimizing for A, B, C work better than only optimizing for A?

28

How to Measure Both System and Individual Campaigns?

You Can Break Individual Campaigns and Campaign Combinations (System) to Separate Cells Cell A

Cell B

Cell C

Cell D

Campaign A (e.g. awareness)

Awareness + Campaign B

Consideration

(e.g. consideration)

+ Campaign C

Acquisition

Control

Control

(e.g. acquisition)

Control

Control

Pros - Enable incrementality measurement at total & individual campaigns levels - Able to evaluate and diagnose campaign performance - Allow multi-KPI measurement for each objective (e.g. intent lift + traffic lift + conversion lift for consideration campaign) - No cross-contamination (likely higher lift per objective) Cons - If individual campaigns targeting different audience, multi-cell may result in audience size being too small per cell

30

You Can Do a Nested Study Structure to Measure Incrementality for Both System and Campaigns Parent study – account level Children studies – campaign level Awareness Consideration Control Control

Acquisition Retargeting Control

Control

Control

Pros - Enable incrementality measurement at total & individual campaigns levels - Able to evaluate and diagnose campaign performance - Allow multi-KPI measurement for each objective (e.g. intent lift + traffic lift + conversion lift for consideration campaign) Cons - Large holdout/opportunity cost with conversion lift at both parent and children levels - Cross-contamination between campaigns (incremental by each campaign on top of all others) - Complexity for execution

You Can Measure Incrementality for System and Use Relevant Ads Reporting for Campaigns Pros - Able to measure total incremental impact - Some insights (not incrementality) for campaign performance - Smaller holdout, smaller opportunity cost

Awareness Consideration Acquisition Retargeting

Control

Cons - No incrementality measurement at campaign level. Campaign-level result is less conclusive - No brand measurement at campaign level - Limited KPIs per objective

You Can Measure Incrementality of Campaigns with Lift and Incrementality of System with MTA Data-Driven Attribution

Awareness

Pros - Enable incrementality measurement at both system and campaign level - Able to evaluate and diagnose campaign performance - Allow multi-KPI measurement for each objective

Consideration Control Control

Acquisition Retargeting Control

Control

Cons - Less feasible if advertiser doesn’t have a proper MTA set-up or the MTA model doesn’t include Facebook

-> Consider the new Facebook Attribution solution

Client Scenarios

Scenario 1: Conversion-driven advertisers testing mid/upper funnel campaigns Current Status: Running acquisition-focused campaigns, optimized for conversions, measured on ROAS Key Challenge: How to add mid and upper funnel campaigns to the mix? How to measure upper and mid funnel to prove value? Path B Path A Cell A

Cell B

Consideration

Acquisition Control • •

Acquisition Control

A system test to prove adding a consideration campaign drives incremental value Suggest measuring the systems on incremental conversions. Cell A may end up with a higher CPA. Help client focus on more incrementality to avoid saturation under a healthy overall CPA goal

• •

Static CTA ads + Video ads (for consideration)

Static CTA ads only

Control

Control

2% lookalike audience for consideration

10% lookalike audience for consideration

Control

Control

Optimizing consideration campaign for site visits

Optimizing consideration campaign for converions

Control

Control

A series of component strategy tests to figure out the best way to add a consideration campaign to the system It can be either measured by incremental conversions or iCPA, depending on primary business objective

Scenario 2: Full-funnel advertisers questioning brand’s impact on sales Current Status: Large advertisers (e.g. Telcos) already running multiple campaigns on Facebook Key Challenge: how does brand campaigns impact sales with long purchase cycle? Measure brand campaigns for brand KPIs As the full-funnel advertisers are also big spenders on TV, we should push them to measure Facebook brand campaigns as how they measure TV campaigns • If they can’t measure TV’s immediate impact on sales, then they should be okay with not measuring FB brand campaigns immediate impact on sales • Leverage TAR and cross-platform brand effect measurement to shift mentality

Use Lift tests to optimize full funnel strategy Themes Audience

Optimization

Use MTA or MMM to measure systems’ impact on sales Intent

Acquisition

Store traffic

Conversion

Credits Allocation

Is audience stratification even necessary? Can we target broad and let auction decide who to reach per objective? Does a broader lookalike audience (5%, 10%) perform better than demo/interest/behavior based audience for upper/mid funnel campaign? Does optimizing for multiple objectives drive better performance than single objective? Does bidding based on user value (e.g. from client model) result in improved performance? Does a variety of format and message outperform fewer formats/messages?

Creative

Site visits

Research Questions

Signals

Campaign Coordination

Does aligning creative messaging to audiences manually outperform allowing the auction to dynamically serve creative from a single ad set. What is the right signal for lookalike audience? Conversion or immediate objective like site visit or certain actions on site? What are the right signals for optimization? Everything for end of funnel vs. stratified signals across reach to conversion? Do multi-funnel campaigns drive more incremental and/or drive higher cost efficiency than single-funnel campaign? Should we align signals for lookalike audience and optimization at each phase of the funnel?

Scenario 3: Product launch taking phased approach Current Status: Prime audience with brand campaigns first, then switch to mid and lower funnel campaigns Key Challenge: Consumers don’t necessarily follow the planned phases Move away from sequencing or phased approach Keep campaigns always-on and adjust weights over time

Acquisition Prospecting Retention

Awareness Launch

Product Available

Sustain

Learning Path

Full Funnel Learning Agenda Paths by Client Type Known Insights

Client Scenario Advertiser Control

1

Testing the Waters

2 Planning for ideal set-up

3 Advanced Execution

Efficiency

Creative

in progress

Audience

Content Ecosystem

Broad Vs Narrow Targeting

Single vs variety of creatives

LALs vs Interest Targeting

How much personalization?

Which audience and optimization?

High vs low resolution personalization

Coordinated vs Targeted Audience & Optimization

Automating Content Ecosystem Manual vs Auction led creative execution

Research

Is audience stratification even necessary

Signals

Optimization

Co-Ordination

What is the right signal for lookalike audience? Conversion or Engagement ?

Which optimization to use?

Which FF design works best? Multi-funnel campaigns vs singlefunnel campaign

What is the right signal for optimization? End of funnel vs objective for each stage

Coordinated vs Targeted Optimization Single vs Multiple Does optimizing for multiple objectives drive better performance than single objective?

Optimal budget split across each stage Awareness vs Consideration vs Loyalty

Automating Signal , Optimization , Co-Ordination Selection Manual Vs Auction Led Execution

Manual vs Auction Led Execution Unavailable / Product Work in Progress

Learning Agenda

Full Funnel Learning Agenda These research questions often came up when advertisers across verticals tried to figure out how to do signal-based full funnel. Treat it as an exploratory starting point where you can shop around testing ideas for your full funnel tests. Don’t forget to share the learnings with us!

Themes Audience

Research Questions Is audience stratification even necessary? Can we target broad and let auction decide who to reach per objective? Does a broader lookalike audience (5%, 10%) perform better than demo/interest/behavior based audience for upper/mid funnel campaign? Does optimizing for multiple objectives drive better performance than single objective?

Optimization

Does bidding based on user value (e.g. from client model) result in improved performance? Does a variety of format and message outperform fewer formats/messages?

Creative

Signals

Campaign Coordination

Does aligning creative messaging to audiences manually outperform allowing the auction to dynamically serve creative from a single ad set. What is the right signal for lookalike audience? Conversion or immediate objective like site visit or certain actions on site? What are the right signals for optimization? Everything for end of funnel vs. stratified signals across reach to conversion? Do multi-funnel campaigns drive more incremental and/or drive higher cost efficiency than single-funnel campaign? Should we align signals for lookalike audience and optimization at each phase of the funnel?

Audience - Is audience stratification even necessary? RESEARCH QUESTION •

STUDY OBJECTIVES

Is audience stratification even necessary? Can we target broad and let auction decide who to reach per objective?

STUDY DESIGN •

Multi-cell lift study – Stratified vs Combined

Stratified Audience

Audience is stratified based on intent signals ahead of time (e.g. audience 1 optimized for web traffic, audience 2 optimized for conversion)

Control TBD%

1.

Primary: Sales or ROAS

2.

Secondary: Web conversions

3.

Also Measuring: Brand lift, attributed sales, CPA, CPM

S E T U P CO N S I D E R AT I O N S –

Constant across cells: approx. audience size, budget, media weight, flight, duration, ad formats, creative, placement, buying method, bid type

–

Variables: targeting

–

Budget: TBD

–

Audience: TBD

–

Duration: TBD

–

Timing: TBD

Combined Audience

Audience is grouped together and auction delivers based on intent (e.g. audience 1 + 2, auction determines who to reach for certain objective)

Control TBD%

42

Audience - Does a signal-based audience work for upper funnel? RESEARCH QUESTION •

STUDY OBJECTIVES

Does a broader lookalike audience (5%, 10%) perform better than demo/interest/behavior based audience for upper/mid funnel campaign?

STUDY DESIGN •

Broad lookalike audience (e.g.10% lookalike audience of existing customers)

Control TBD%

Primary: Sales or brand lift, depending on primary objective

2.

Also Measuring: attributed sales, CPA, CPM

S E T U P CO N S I D E R AT I O N S

Multi-cell lift study – Cell A vs Cell B

Cell A

1.

Cell B Demo audience (e.g. A18 -34)

Control TBD%

Cell C Interest-based audience (e.g. tech-savvy)

–

Constant across cells: audience size, budget, media weight, flight, duration, ad formats, creative, placement, buying method, bid type

–

Variables: Targeting

–

Budget: TBD

–

Audience: TBD

–

Duration: TBD

–

Timing: TBD

Control TBD%

43

Optimization - Do multiple objectives outperform single-objective campaigns? STUDY OBJECTIVES RESEARCH QUESTION •

Does optimizing for multiple objectives drive better performance than single objective?

STUDY DESIGN •

Multi-objective (e.g. 50% conversion, 30% site visit, 20% reach) Control TBD%

Primary: Sales or ROAS

2.

Secondary: Web conversions

3.

Also Measuring: Brand lift, attributed sales, CPA, CPM

S E T U P CO N S I D E R AT I O N S

Multi-cell lift study – Cell A vs Cell B

Cell A

1.

Cell B Single-objective (e.g. 100% conversion)

Control TBD%

–

Constant across cells: targeting, audience size, budget, media weight, flight, duration, ad formats, creative, placement, buying method, bid type

–

Variables: Optimization objective

–

Budget: TBD

–

Audience: TBD

–

Duration: TBD

–

Timing: TBD

44

Optimization - Does bidding based on user value (e.g. from

client model) result in improved performance? STUDY OBJECTIVES

RESEARCH QUESTION •

Does bidding based on user value (e.g. from client model) result in improved performance?

STUDY DESIGN •

Primary: Sales or ROAS

2.

Secondary: Web conversions

3.

Also Measuring: Brand lift, attributed sales, CPA, CPM

S E T U P CO N S I D E R AT I O N S

Multi-cell lift study – Cell A vs Cell B

Cell A

Cell B

Bidding on user value

Default on lowest cost

Control TBD%

1.

Control TBD%

–

Constant across cells: targeting, audience size, budget, media weight, flight, duration, ad formats, creative, placement, buying method

–

Variables: bid type

–

Budget: TBD

–

Audience: TBD

–

Duration: TBD

–

Timing: TBD

45

Creative - Does a variety of formats and messages outperform

fewer or single formats/messages? RESEARCH QUESTION •

Do we need a content system with a variety of formats and messages? Is more the better?

STUDY DESIGN •

Multi-cell lift study – Cell A vs Cell B

1.

Primary: Sales or brand lift depending on primary objective

2.

Secondary: Web conversions

3.

Also Measuring: attributed sales, CPA, CPM

S E T U P CO N S I D E R AT I O N S

Cell A

Cell B

Multiple formats/messages (e.g. static DR ads + prospecting video)

Fewer or single format/message (e.g. static DR ad only)

Control TBD%

STUDY OBJECTIVES

–

Constant across cells: targeting, audience size, budget, media weight, flight, duration, placement, buying method, bid type

–

Variables: Ad format, or creative message

–

Budget: TBD

–

Audience: TBD

–

Duration: TBD

–

Timing: TBD

Control TBD% 46

Creative – Does auction-determined creative outperform

manually set-up assets? RESEARCH QUESTION •

Does allowing auction to dynamically serve creative from a single ad set outperform aligning creative messaging to audiences manually?

STUDY DESIGN •

Auctiondetermined creative (single ad set)

Control TBD%

1.

Primary: Sales or Brand lift, depending on primary objective

2.

Secondary: Web conversions

3.

Also Measuring: Attributed sales, CPA, CPM

S E T U P CO N S I D E R AT I O N S

Multi-cell lift study – Cell A vs Cell B

Cell A

STUDY OBJECTIVES

Cell B

Manually aligned creative to audience (ad set A: creative A to audience A; ad set B: creative B to audience B)

Control TBD%

–

Constant across cells: targeting, audience size, budget, media weight, flight, duration, ad formats, placement, buying method, bid type

–

Variables: creative strategy

–

Budget: TBD

–

Audience: TBD

–

Duration: TBD

–

Timing: TBD

47

Signal – What is the right signal for lookalike audience? RESEARCH QUESTION •

STUDY OBJECTIVES

What is the right signal for lookalike audience? Conversion or immediate objective like site visit or certain actions on site?

STUDY DESIGN •

Primary: Sales or ROAS

2.

Secondary: Web conversions

3.

Also Measuring: Brand lift, attributed sales, CPA, CPM

S E T U P CO N S I D E R AT I O N S

Multi-cell lift study – Cell A vs Cell B

Cell A

1.

–

Constant across cells: audience size, budget, media weight, flight, duration, ad formats, creative, placement, buying method, bid type

–

Variables: Targeting

–

Budget: TBD

–

Audience: TBD

–

Duration: TBD

–

Timing: TBD

Cell B

e.g. Using converters as the seeds for lookalike audience for prospecting

e.g. Using site visitors as the seeds for lookalike audience for prospecting

Control TBD%

Control TBD% 48

Signal – What are the right signals for optimization? RESEARCH QUESTION •

STUDY OBJECTIVES

What are the right signals for optimization? Everything for end of funnel vs. stratified signals across reach to conversion?

Primary: Sales or ROAS

2.

Secondary: Web conversions

3.

Also Measuring: Brand lift, attributed sales, CPA, CPM

S E T U P CO N S I D E R AT I O N S

•

STUDY DESIGN

•

Multi-cell lift study – Cell A vs Cell B

Cell A

Cell B

Optimized for end of funnel activity (e.g. everything optimized for conversions

Optimized across a variety of signals (e.g. site visit for prospecting, conversion for acquisition)

Control TBD%

1.

–

Constant across cells: targeting, audience size, budget, media weight, flight, duration, ad formats, creative, placement, buying method, bid type

–

Variables: optimization objective

–

Budget: TBD

–

Audience: TBD

–

Duration: TBD

–

Timing: TBD

Control TBD%

49

Coordination - Do multi-funnel campaigns drive more incremental and/or better cost efficiency than single-funnel campaign? STUDY OBJECTIVES

RESEARCH QUESTION •

Do multi-funnel campaigns drive more incremental and/or higher cost efficiency than single-funnel campaign?

STUDY DESIGN •

e.g. Consideration + Acquisition + Retargeting Control TBD%

Primary: Sales or ROAS

2.

Secondary: Web conversions

3.

Also Measuring: Brand lift, attributed sales, CPA, CPM

S E T U P CO N S I D E R AT I O N S

Multi-cell lift study – Cell A vs Cell B

Cell A

1.

Cell B e.g. Acquisition only

–

Constant across cells: targeting, audience size, budget, media weight, flight, duration, ad formats, creative, placement, buying method, bid type

–

Variables: var1, var2

–

Budget: TBD

–

Audience: TBD

–

Duration: TBD

–

Timing: TBD

Control TBD%

50

Coordination - Should we align signals for lookalike audience and optimization at each phase of the funnel? STUDY OBJECTIVES

RESEARCH QUESTION •

Manual signal stratification or auction-determined delivery?

STUDY DESIGN •

Multi-cell lift study – Cell A vs Cell B

Cell A Manual signal stratification (e.g. 10% site visitor lookalike optimized for site visits, 2% existing customers lookalike optimized for conversions

Control TBD%

1.

Primary: Sales or ROAS

2.

Secondary: Web conversions

3.

Also Measuring: Brand lift, attributed sales, CPA, CPM

S E T U P CO N S I D E R AT I O N S

Cell B

Same audience optimized for different objectives or stratified audience optimized for same objective

–

Constant across cells: Audience size, budget, media weight, flight, duration, ad formats, creative, placement, buying method, bid type

–

Variables: Targeting and optimization

–

Budget: TBD

–

Audience: TBD

–

Duration: TBD

–

Timing: TBD

Control TBD%

51

Appendix

Tactics have evolved, but the goal is the same

Concept introduced in 1898, name coined in 1917, became integrated marketing, and continues today MASS MARKETING

I N T E G R AT E D MARKETING

PERSONALIZED MARKETING

1900 - 1950s

1960s - 2000s

2010 – Today

•

No personalizaion

•

Broadcast-level personalization

•

Individual-level personalization

•

No mechanism to differentiate

•

Channel mix used to differentiate

•

Data mix is used to differentiate

Channel mix has been the key decision for decades

But now, digital platforms can deliver mass reach, engagement, and direct response Channel

Awareness Interest Desire

Action

Awareness

Interest

Desire

Action

Audience

Message

Delivery

Format

A M O D E R N A P P R O A C H U S I N G D ATA A N D M A C H I N E L E A R N I N G

You layer in what you know about your customers through pixel, SDK or offline conversions…

People interact with content on their device

…our system identifies patterns, learning from signals to match content to the right people

….and complete actions, creating a spectrum of intent signals

55

Targeting and optimization determine who sees your ad The data mix used plays a key role TA R G E T I N G

O P T I M I Z AT I O N G O A L

Your defined target audience

The outcome you tell us is important to you

LOCATION

INTERESTS

DEMOGRAPHICS

+

CONVERSIONS

CLICKS

VIDEO VIEWS

WHO WILL SEE YOUR AD

=

Optimize your ads for our value-based delivery system Increase your ad’s “Total Value” by optimizing for a specific event For each individual, we rank eligible ads based on a “Total Value” for each ad MAXIMIZING A D V E R T I S E R VA L U E

Advertiser Bid

Your bid for the event you selected as your optimization goal – i.e. your desired result

×

Estimated Action rates

What’s the likelihood that an impression shown to this person will lead to your desired result?

OPTIMIZING CONSUMER EXPERIENCE

+

User Value

How interesting do we think this individual is going to find this ad? Is this a high-quality ad?

=

TOTAL VALUE

The ad with the highest Total Value wins the auction and shows to the individual