opticon2017
Decisions at Scale:
Empower your Team to do
More with Experiment Data
Giannis Psaroudakis
Product Manager, Optimizely
Bing had 8 years
of consecutive
market share
growth
Source: Comscore
“The growth of experimentation is the major reason Bing is
profitable and its share of U.S desktop searches nearly tripled.”
Ronny Kohavi, GM of Analysis and Experimentation, Microsoft
Source: Comscore
Old Reality Culture of
Top-Down Innovation
Embrace Success
Make Decisions
Follow Orders
Bottom-Up Innovation
Embrace Failure
Validate Decisions
Follow Data
Experimentation Hero
10’s experiments / year
Experimentation Program
100’s experiments / year
Culture of Experimentation
1000’s experiments / year
Scaling the Experimentation Program
Optimizely Data Platform
Mission
Make Customer Data Useful Through Experimentation
ResultsEvents Stats EngineAudiences Metrics
Process signals
across customer
experiences
Target user groups
with common
attributes
Measure the
impact of
hypotheses
Understand the
impact of
hypotheses
Validate hypotheses
with statistical
confidence
6.5+Billion Events
Daily
1.2
+
Billion
Experiences
Daily
50+Million Unique
Users Daily
Meet Kristen
Director of
Experimentation
Empower every other team in her organization
to make validated decisions with high velocity.
Center of Excellence
Support
Sales Marketing
E-commerce
Engineering
Center of Excellence
Support
Sales Marketing
E-commerce
?
? ?
?
?
Engineering
Can't leverage the results Experimentation is not useful
Can’t measure what matters to them Reduced adoption of the program
Don't have confidence in the results Losing trust in experimentation
“I want to support the measurement
needs of all teams across the
organization”
1. Can’t measure what matters
3. No confidence in the results
2. Can't leverage the results
E-commerce
Product Manager
“I want to measure the impact of
changing our e-commerce experience
on our revenue metrics”
Sales Leads
Manager
“I want to measure the impact of our
online customer outreach campaigns on
offline conversions, such as phone calls”
App Performance
Engineer
“I want to measure the impact of our
code changes on the application’s
page load time”
Social Media
Marketing Manager
“I want to increase the impact of our
social media campaigns by reducing
the bounce rate on landing pages”
Sources
E-Commerce
Sales Leads
Engineering
Social Media
Metrics
Web/Mobile
Internal System
Server
Web/Mobile
Revenue
Offline Conversions
Page Load Time
Bounce Rate
Sources
X Event API
Bring your own signals into
experimentation no matter where
they live.
Web, Mobile
& Full-Stack
Offline
Conversions
Business
Intelligence
Secure
Environments
Event APISnippet / SDKs
Optimizely X
Optimizely
Business
Intelligence
System
RPV: $145
LTV: $2145
ticket_count: +1
Call
Center
Sources
Metrics Builder
Measure the metrics that matter most to your
business with greater precision and flexibility.
Metrics
! Intuitive interface
! New metric types
! Metrics calculated in real-time
Bounce & Exit Rage
Metrics
Conversions
Revenue
Custom Value
Measure distinct user actions
Measure any numerical goal
checkouts per visitor,
page views per visitor
Measure transaction amounts
revenue per visitor (RPV),
average order value (AOV)
Session Duration Measure engagement duration avg. time spent on news article pages,
avg. time spent on site
Measure user abandonment
page load time per visitor,
purchased items per order
bounce rate on homepage,
exit rate on checkout page
Metric Description Examples
(2018)
(in beta November 2017)
Teams can measure what matters to them
• Identify the signals and metrics teams need
• Incorporate them into experimentation
• Scale by enabling everyone to measure what
matters to them
1. Can’t measure what matters
3. No confidence in the results
2. Can't leverage the results
“I want teams to have flexibility in
the way they analyze the results”
1. Can’t measure what matters
3. No confidence in the results
2. Can't leverage the results
Analysis Method
E-Commerce
Sales Leads
Engineering
Social Media
Results Page
Analytics Software
Raw Data
Spreadsheet
Analysis Methods
Results Export
Expand your analysis workflow with a
variety of new and improved methods
to export your results.
Results API
CSV Export
Raw Data Export
Analytics Integration
Analysis Methods
GET /v2/experiments/{experiment_id}/results
...
"metrics": [
{
"name":"items in cart"
"results": {
"123456": {
"is_baseline":false,
"lift": {
"confidence_interval": [
1.152505,
2.612566
],
"is_significant":true,
"significance":0.9885,
"value":0.44
"visitors_remaining":2601
}
}
...
}
Your Application
Custo
m
Alerts
Automated
Actions
Aggregated
Reporting
Results API
Custom StatsDashboards Debugging
Analysis Methods
Data Export
Custom ETL Pipeline
Develop your own custom integration to bring
experimentation results into your analytics tool.
Analysis Methods
+
Custom
Integration
Framework
1. Can’t measure what matters
3. No confidence in the results
2. Can't leverage the results Teams can analyze the results
• Understand how teams are doing analysis
• Connect experiment data with their workflows
• Scale by enabling everyone to analyze
“I want tools our teams can trust
and have faith that the results will
manifest”
1. Can’t measure what matters
3. No confidence in the results
2. Need help to analyze experiments
1. False Discoveries
2. Skewed Results (Outliers / Bots)
3. Data Discrepancies
Top reasons teams lose
confidence on experimentation?!
1.False Discoveries
2. Skewed Results (Outliers / Bots)
3. Data Discrepancies
Top reasons teams lose
confidence on experimentation?!
Results Confidence
Stats Engine
Validate decisions with statistical confidence
! Error control for multiple hypothesis testing
! Prevents bad decisions due to ‘peeking’
! Configurable risk tolerance
1. False Discoveries
2.Skewed Results (Outliers / Bots)
3. Data Discrepancies
Top reasons teams lose
confidence on experimentation
Income per Person
$50-70K Billions!
Average Order
$500
Revenue
per Order ($)
Results Confidence
Revenue
per Order ($)
Results Confidence
Extreme Orders!
Average Order
$500
Revenue
per Order ($)
Results Confidence
Outlier Threshold
Average Order
$500
Extreme Orders!
Outlier Filtering
Improve the integrity of your decisions by
filtering outliers from the results.*
! Enabled on demand
! Calculated in real-time
! Configurable thresholds
Results Confidence
*Availability: Coming in beta November 2017. Compatible with revenue and custom value metrics.
Bot Filtering
Enhanced Bot Filtering to improve the
fidelity of the results
! Enhanced bot filtering that complies to industry
standard (IAB/ABC List) for web analytics
! Bots automatically filtered from the results
OPTIMIZELY
Results Confidence
1. False Discoveries
2. Skewed Results (Bots / Outliers)
3.Data Discrepancies
Top reasons teams lose
confidence on experimentation
Numbers don't match!
1. Product Differences
2. Implementation Bugs
3.Event Timing Issues
Leading factors
of analytics
discrepancies
visitor counter: +1
First Optimizely event
First 3rd-party analytics event
Optimizely
visitor counter: +1
Other Analytics
Results Confidence
visitor counter: +1
First Optimizely event
First 3rd-party analytics event
Optimizely
visitor counter: +1
Other Analytics
+2
+1
Results Confidence
Hold/SendEvents API
Mitigate discrepancies between
Optimizely and your web analytics.
window.optimizely.push({type: "holdEvents"});
window.optimizely.push({type: "sendEvents"});
HoldEvents: Instruct Optimizely to hold
the events in a browser queue.
SendEvents: Instruct Optimizely to release
the events from the browser queue.
Results Confidence
visitor counter: +1
First 3rd-party analytics event Optimizely &
Other AnalyticsFirst Optimizely event
Results Confidence
visitor counter: +1
First 3rd-party analytics event Optimizely &
Other AnalyticsFirst Optimizely event
reduction in discrepancies
related to event timing
90%
up to
Results Confidence
1. Can’t measure what matters
3. No confidence in the results
2. Need help to analyze experiments
Teams have confidence in the results
• Identify sources of bias or discrepancies
• Filter them out from the results
• Scale by enabling everyone to analyze the results
with confidence
Decisions at Scale
!Grow adoption by enabling teams to measure what matters to them
!Grow adoption by enabling teams to analyze using their own workflows
!Maintain adoption by giving teams confidence on the results
Thank you!

Opticon 2017 Decisions at Scale

  • 1.
    opticon2017 Decisions at Scale: Empoweryour Team to do More with Experiment Data Giannis Psaroudakis Product Manager, Optimizely
  • 5.
    Bing had 8years of consecutive market share growth Source: Comscore
  • 6.
    “The growth ofexperimentation is the major reason Bing is profitable and its share of U.S desktop searches nearly tripled.” Ronny Kohavi, GM of Analysis and Experimentation, Microsoft Source: Comscore
  • 7.
    Old Reality Cultureof Top-Down Innovation Embrace Success Make Decisions Follow Orders Bottom-Up Innovation Embrace Failure Validate Decisions Follow Data
  • 8.
    Experimentation Hero 10’s experiments/ year Experimentation Program 100’s experiments / year Culture of Experimentation 1000’s experiments / year Scaling the Experimentation Program
  • 9.
  • 10.
    Mission Make Customer DataUseful Through Experimentation
  • 11.
    ResultsEvents Stats EngineAudiencesMetrics Process signals across customer experiences Target user groups with common attributes Measure the impact of hypotheses Understand the impact of hypotheses Validate hypotheses with statistical confidence 6.5+Billion Events Daily 1.2 + Billion Experiences Daily 50+Million Unique Users Daily
  • 12.
  • 13.
  • 14.
    Empower every otherteam in her organization to make validated decisions with high velocity.
  • 15.
    Center of Excellence Support SalesMarketing E-commerce Engineering
  • 16.
    Center of Excellence Support SalesMarketing E-commerce ? ? ? ? ? Engineering
  • 17.
    Can't leverage theresults Experimentation is not useful Can’t measure what matters to them Reduced adoption of the program Don't have confidence in the results Losing trust in experimentation
  • 18.
    “I want tosupport the measurement needs of all teams across the organization” 1. Can’t measure what matters 3. No confidence in the results 2. Can't leverage the results
  • 19.
    E-commerce Product Manager “I wantto measure the impact of changing our e-commerce experience on our revenue metrics”
  • 20.
    Sales Leads Manager “I wantto measure the impact of our online customer outreach campaigns on offline conversions, such as phone calls”
  • 21.
    App Performance Engineer “I wantto measure the impact of our code changes on the application’s page load time”
  • 22.
    Social Media Marketing Manager “Iwant to increase the impact of our social media campaigns by reducing the bounce rate on landing pages”
  • 23.
    Sources E-Commerce Sales Leads Engineering Social Media Metrics Web/Mobile InternalSystem Server Web/Mobile Revenue Offline Conversions Page Load Time Bounce Rate
  • 24.
    Sources X Event API Bringyour own signals into experimentation no matter where they live. Web, Mobile & Full-Stack Offline Conversions Business Intelligence Secure Environments Event APISnippet / SDKs Optimizely X
  • 25.
  • 26.
    Metrics Builder Measure themetrics that matter most to your business with greater precision and flexibility. Metrics ! Intuitive interface ! New metric types ! Metrics calculated in real-time
  • 28.
    Bounce & ExitRage Metrics Conversions Revenue Custom Value Measure distinct user actions Measure any numerical goal checkouts per visitor, page views per visitor Measure transaction amounts revenue per visitor (RPV), average order value (AOV) Session Duration Measure engagement duration avg. time spent on news article pages, avg. time spent on site Measure user abandonment page load time per visitor, purchased items per order bounce rate on homepage, exit rate on checkout page Metric Description Examples (2018) (in beta November 2017)
  • 29.
    Teams can measurewhat matters to them • Identify the signals and metrics teams need • Incorporate them into experimentation • Scale by enabling everyone to measure what matters to them 1. Can’t measure what matters 3. No confidence in the results 2. Can't leverage the results
  • 30.
    “I want teamsto have flexibility in the way they analyze the results” 1. Can’t measure what matters 3. No confidence in the results 2. Can't leverage the results
  • 31.
    Analysis Method E-Commerce Sales Leads Engineering SocialMedia Results Page Analytics Software Raw Data Spreadsheet
  • 32.
    Analysis Methods Results Export Expandyour analysis workflow with a variety of new and improved methods to export your results. Results API CSV Export Raw Data Export Analytics Integration
  • 33.
    Analysis Methods GET /v2/experiments/{experiment_id}/results ... "metrics":[ { "name":"items in cart" "results": { "123456": { "is_baseline":false, "lift": { "confidence_interval": [ 1.152505, 2.612566 ], "is_significant":true, "significance":0.9885, "value":0.44 "visitors_remaining":2601 } } ... } Your Application Custo m Alerts Automated Actions Aggregated Reporting Results API
  • 34.
    Custom StatsDashboards Debugging AnalysisMethods Data Export Custom ETL Pipeline
  • 35.
    Develop your owncustom integration to bring experimentation results into your analytics tool. Analysis Methods + Custom Integration Framework
  • 36.
    1. Can’t measurewhat matters 3. No confidence in the results 2. Can't leverage the results Teams can analyze the results • Understand how teams are doing analysis • Connect experiment data with their workflows • Scale by enabling everyone to analyze
  • 37.
    “I want toolsour teams can trust and have faith that the results will manifest” 1. Can’t measure what matters 3. No confidence in the results 2. Need help to analyze experiments
  • 38.
    1. False Discoveries 2.Skewed Results (Outliers / Bots) 3. Data Discrepancies Top reasons teams lose confidence on experimentation?!
  • 39.
    1.False Discoveries 2. SkewedResults (Outliers / Bots) 3. Data Discrepancies Top reasons teams lose confidence on experimentation?!
  • 40.
    Results Confidence Stats Engine Validatedecisions with statistical confidence ! Error control for multiple hypothesis testing ! Prevents bad decisions due to ‘peeking’ ! Configurable risk tolerance
  • 41.
    1. False Discoveries 2.SkewedResults (Outliers / Bots) 3. Data Discrepancies Top reasons teams lose confidence on experimentation
  • 42.
  • 43.
  • 44.
    Revenue per Order ($) ResultsConfidence Extreme Orders! Average Order $500
  • 45.
    Revenue per Order ($) ResultsConfidence Outlier Threshold Average Order $500 Extreme Orders!
  • 46.
    Outlier Filtering Improve theintegrity of your decisions by filtering outliers from the results.* ! Enabled on demand ! Calculated in real-time ! Configurable thresholds Results Confidence *Availability: Coming in beta November 2017. Compatible with revenue and custom value metrics.
  • 48.
    Bot Filtering Enhanced BotFiltering to improve the fidelity of the results ! Enhanced bot filtering that complies to industry standard (IAB/ABC List) for web analytics ! Bots automatically filtered from the results OPTIMIZELY Results Confidence
  • 49.
    1. False Discoveries 2.Skewed Results (Bots / Outliers) 3.Data Discrepancies Top reasons teams lose confidence on experimentation
  • 50.
    Numbers don't match! 1.Product Differences 2. Implementation Bugs 3.Event Timing Issues Leading factors of analytics discrepancies
  • 51.
    visitor counter: +1 FirstOptimizely event First 3rd-party analytics event Optimizely visitor counter: +1 Other Analytics Results Confidence
  • 52.
    visitor counter: +1 FirstOptimizely event First 3rd-party analytics event Optimizely visitor counter: +1 Other Analytics +2 +1 Results Confidence
  • 53.
    Hold/SendEvents API Mitigate discrepanciesbetween Optimizely and your web analytics. window.optimizely.push({type: "holdEvents"}); window.optimizely.push({type: "sendEvents"}); HoldEvents: Instruct Optimizely to hold the events in a browser queue. SendEvents: Instruct Optimizely to release the events from the browser queue. Results Confidence
  • 54.
    visitor counter: +1 First3rd-party analytics event Optimizely & Other AnalyticsFirst Optimizely event Results Confidence
  • 55.
    visitor counter: +1 First3rd-party analytics event Optimizely & Other AnalyticsFirst Optimizely event reduction in discrepancies related to event timing 90% up to Results Confidence
  • 56.
    1. Can’t measurewhat matters 3. No confidence in the results 2. Need help to analyze experiments Teams have confidence in the results • Identify sources of bias or discrepancies • Filter them out from the results • Scale by enabling everyone to analyze the results with confidence
  • 57.
    Decisions at Scale !Growadoption by enabling teams to measure what matters to them !Grow adoption by enabling teams to analyze using their own workflows !Maintain adoption by giving teams confidence on the results
  • 59.