A/B Testing for Startups
by Aleksandr Osiyuk
Product Analyst at MacPaw
t.me/osiyuk
MacPaw products
CleanMyMac Setapp Gemini Gemini Photos
Clean External
Drives to Get
More Space
The first
subscription
store for Mac
apps
Wipe Duplicate
Files Off Your
Mac
Optimize
iPhone photo
storage
Startup must run A/B tests another ways
that large companies
Analyst & PM expectations
Analyst & PM expectations
Analyst needs to:
● apply a scientific approach
● dig into metrics
● try different methods of analysis
Analyst & PM expectations
Analyst needs to:
● apply a scientific approach
● dig into metrics
● try different methods of analysis
PM needs to:
● fix a winner as soon as possible
● know how much money it produces
Factors affected the testing time
Factors affected the testing time
● Base conversion rate
Factors affected the testing time
● Base conversion rate
● Relative increase in conversion rate
Factors affected the testing time
● Base conversion rate
● Relative increase in conversion rate
● Daily number of users
Base CR
Relative
increase
New CR Sample size
Case 1 1% 5% 1.05% 626,230
Case 2 2% 5% 2.1% 309,927
Case 3 3% 5% 3.2% 204,493
Case 4 4% 5% 4.2% 151,776
Case 5 5% 5% 5.3% 120,145
Case 6 10% 5% 10.5% 56,885
Case 7 15% 5% 15.8% 35,798
Case 8 20% 5% 21.0% 25,255
Case 9 25% 5% 26.3% 18,929
Base CR
Relative
increase
New CR Sample size
Case 1 10% 1% 10.1% 1,414,680
Case 2 10% 2% 10.2% 354,138
Case 3 10% 3% 10.3% 157,602
Case 4 10% 4% 10.4% 88,767
Case 5 10% 5% 10.5% 56,885
Case 6 10% 10% 11.0% 14,312
Case 7 10% 15% 11.5% 6,401
Case 8 10% 20% 12.0% 3,622
Case 9 10% 25% 12.5% 2,332
Tips to speed up testing
Tips to speed up testing
● Use metric that is closely correlated to final conversion
Tips to speed up testing
● Use metric that is closely correlated to final conversion
● Testing big changes
Tips to speed up testing
● Use metric that is closely correlated to final conversion
● Testing big changes
● Divide users in the middle of a funnel
CR from 1 step to 3 step
CR from 2 step to 3 step
Base CR: 50/1000=5%
Relative increase: 10%
Sample size: 30 244 (30 days)
1 step
1000
2 step
200
Target action
55
5
50Base CR: 50/200=25%
Relative increase: 10%
Sample size: 4 753 (24 days)
CR 25%
CR 5%
CR from 1 step to 3 step
CR from 2 step to 3 step
Base CR: 50/1000=5%
Relative increase: 10%
Sample size: 30 244 (30 days)
1 step
1000
2 step
200
55
5
50Base CR: 50/200=25%
Relative increase: 10%
Sample size: 4 753 (24 days)
CR 25%
CR 5%
A/A test
Target action
Tips to speed up testing
● Use metric that is closely correlated to final conversion
● Testing big changes
● Divide users in the middle of a funnel
● Increase visual contrast to be noticed by users
Tips to speed up testing
Users
every day
Actions CR
Relative
increase
Sample
size
n, days
1000 120 12% 10% 11657 11.7
Tips to speed up testing
Users
every day
Actions CR
Relative
increase
Sample
size
n, days
1000 120 12% 10% 11657 11.7
500 50 10% 8%
Tips to speed up testing
Users
every day
Actions CR
Relative
increase
Sample
size
n, days
1000 120 12% 10% 11657 11.7
500 50 10% 8%
1500 170 11.33% 9.41% 14041 9.4
Tips to speed up testing
CR
Relativeincrease
Tips to speed up testing
● Use metric that is closely correlated to final conversion
● Testing big changes
● Divide users in the middle of a funnel
● Increase visual contrast to be noticed by users
● Calculate minimum detectable effect (MDE)
Another ways to speed up testing
Another ways to speed up testing
● Metrics: binary / non-binary
Another ways to speed up testing
● Metrics: binary / non-binary
● Analysis: bayesian / frequentist
Another ways to speed up testing
● Metrics: binary / non-binary
● Analysis: bayesian / frequentist
● Hypotheses: innovate / optimize
Experiments with monetization
Experiments with monetization
Experiments with monetization
Print magazine - 20$
Digital magazine - 10$
Print magazine - 20$
Digital magazine - 10$
1
29%
2
71%
AOV: 13$
Experiments with monetization
Print magazine - 20$
Digital magazine - 10$
Print & Digital - 20$
Print magazine - 20$
Digital magazine - 10$
Print & Digital - 20$
1
11%
2
30%
3
59%
AOV: 17$
Experiments with monetization
Digital magazine - 10$
Print & Digital - 20$
Digital magazine - 10$
Print & Digital - 20$
1
61%
2
39%
AOV: 14$
A/B testing for Insights
A/B testing for Insights
● Check more metrics and user flow
A/B testing for Insights
● Check more metrics and user flow
● Segmentation (old/new, traffic, country, behavior etc.)
A/B testing for Insights
● Check more metrics and user flow
● Segmentation (old/new, traffic, country, behavior etc.)
● Mix overlapping / chained tests
A/B testing for Insights
● Check more metrics and user flow
● Segmentation (old/new, traffic, country, behavior etc.)
● Mix overlapping / chained tests
● Multiple comparisons problem
A/B testing for Insights
● Check more metrics and user flow
● Segmentation (old/new, traffic, country, behavior etc.)
● Mix overlapping / chained tests
● Multiple comparisons problem
● Simpson’s paradox
Male Female
568 of 2000 (28%) 372 of 2000 (19%)
Simpson’s paradox
Male Female
568 of 2000 (28%) 372 of 2000 (19%)
Highly competitive
subjects
168 of 1200 (14%) 270 of 1800 (15%)
Lowly competitive
subjects
400 of 800 (50%) 102 of 200 (51%)
Simpson’s paradox
А/В testing features in apps
А/В testing features in apps
● Why do you need to split the traffic on server-side?
● If there is no internet? (А/А-test in apps 50/50? Case about
custom settings)
Calculator: bit.ly/2JswQWc
t.me/BigQueryThanks for attention!
Telegram channel about А/В-testing:
t.me/ABtesting

A/B testing for startups

  • 1.
    A/B Testing forStartups by Aleksandr Osiyuk Product Analyst at MacPaw t.me/osiyuk
  • 2.
    MacPaw products CleanMyMac SetappGemini Gemini Photos Clean External Drives to Get More Space The first subscription store for Mac apps Wipe Duplicate Files Off Your Mac Optimize iPhone photo storage
  • 3.
    Startup must runA/B tests another ways that large companies
  • 4.
    Analyst & PMexpectations
  • 5.
    Analyst & PMexpectations Analyst needs to: ● apply a scientific approach ● dig into metrics ● try different methods of analysis
  • 6.
    Analyst & PMexpectations Analyst needs to: ● apply a scientific approach ● dig into metrics ● try different methods of analysis PM needs to: ● fix a winner as soon as possible ● know how much money it produces
  • 7.
  • 8.
    Factors affected thetesting time ● Base conversion rate
  • 9.
    Factors affected thetesting time ● Base conversion rate ● Relative increase in conversion rate
  • 10.
    Factors affected thetesting time ● Base conversion rate ● Relative increase in conversion rate ● Daily number of users
  • 11.
    Base CR Relative increase New CRSample size Case 1 1% 5% 1.05% 626,230 Case 2 2% 5% 2.1% 309,927 Case 3 3% 5% 3.2% 204,493 Case 4 4% 5% 4.2% 151,776 Case 5 5% 5% 5.3% 120,145 Case 6 10% 5% 10.5% 56,885 Case 7 15% 5% 15.8% 35,798 Case 8 20% 5% 21.0% 25,255 Case 9 25% 5% 26.3% 18,929
  • 12.
    Base CR Relative increase New CRSample size Case 1 10% 1% 10.1% 1,414,680 Case 2 10% 2% 10.2% 354,138 Case 3 10% 3% 10.3% 157,602 Case 4 10% 4% 10.4% 88,767 Case 5 10% 5% 10.5% 56,885 Case 6 10% 10% 11.0% 14,312 Case 7 10% 15% 11.5% 6,401 Case 8 10% 20% 12.0% 3,622 Case 9 10% 25% 12.5% 2,332
  • 13.
    Tips to speedup testing
  • 14.
    Tips to speedup testing ● Use metric that is closely correlated to final conversion
  • 15.
    Tips to speedup testing ● Use metric that is closely correlated to final conversion ● Testing big changes
  • 16.
    Tips to speedup testing ● Use metric that is closely correlated to final conversion ● Testing big changes ● Divide users in the middle of a funnel
  • 17.
    CR from 1step to 3 step CR from 2 step to 3 step Base CR: 50/1000=5% Relative increase: 10% Sample size: 30 244 (30 days) 1 step 1000 2 step 200 Target action 55 5 50Base CR: 50/200=25% Relative increase: 10% Sample size: 4 753 (24 days) CR 25% CR 5%
  • 18.
    CR from 1step to 3 step CR from 2 step to 3 step Base CR: 50/1000=5% Relative increase: 10% Sample size: 30 244 (30 days) 1 step 1000 2 step 200 55 5 50Base CR: 50/200=25% Relative increase: 10% Sample size: 4 753 (24 days) CR 25% CR 5% A/A test Target action
  • 19.
    Tips to speedup testing ● Use metric that is closely correlated to final conversion ● Testing big changes ● Divide users in the middle of a funnel ● Increase visual contrast to be noticed by users
  • 20.
    Tips to speedup testing Users every day Actions CR Relative increase Sample size n, days 1000 120 12% 10% 11657 11.7
  • 21.
    Tips to speedup testing Users every day Actions CR Relative increase Sample size n, days 1000 120 12% 10% 11657 11.7 500 50 10% 8%
  • 22.
    Tips to speedup testing Users every day Actions CR Relative increase Sample size n, days 1000 120 12% 10% 11657 11.7 500 50 10% 8% 1500 170 11.33% 9.41% 14041 9.4
  • 23.
    Tips to speedup testing CR Relativeincrease
  • 24.
    Tips to speedup testing ● Use metric that is closely correlated to final conversion ● Testing big changes ● Divide users in the middle of a funnel ● Increase visual contrast to be noticed by users ● Calculate minimum detectable effect (MDE)
  • 25.
    Another ways tospeed up testing
  • 26.
    Another ways tospeed up testing ● Metrics: binary / non-binary
  • 27.
    Another ways tospeed up testing ● Metrics: binary / non-binary ● Analysis: bayesian / frequentist
  • 28.
    Another ways tospeed up testing ● Metrics: binary / non-binary ● Analysis: bayesian / frequentist ● Hypotheses: innovate / optimize
  • 29.
  • 30.
  • 31.
    Experiments with monetization Printmagazine - 20$ Digital magazine - 10$
  • 32.
    Print magazine -20$ Digital magazine - 10$ 1 29% 2 71% AOV: 13$
  • 33.
    Experiments with monetization Printmagazine - 20$ Digital magazine - 10$ Print & Digital - 20$
  • 34.
    Print magazine -20$ Digital magazine - 10$ Print & Digital - 20$ 1 11% 2 30% 3 59% AOV: 17$
  • 35.
    Experiments with monetization Digitalmagazine - 10$ Print & Digital - 20$
  • 36.
    Digital magazine -10$ Print & Digital - 20$ 1 61% 2 39% AOV: 14$
  • 37.
  • 38.
    A/B testing forInsights ● Check more metrics and user flow
  • 39.
    A/B testing forInsights ● Check more metrics and user flow ● Segmentation (old/new, traffic, country, behavior etc.)
  • 40.
    A/B testing forInsights ● Check more metrics and user flow ● Segmentation (old/new, traffic, country, behavior etc.) ● Mix overlapping / chained tests
  • 41.
    A/B testing forInsights ● Check more metrics and user flow ● Segmentation (old/new, traffic, country, behavior etc.) ● Mix overlapping / chained tests ● Multiple comparisons problem
  • 42.
    A/B testing forInsights ● Check more metrics and user flow ● Segmentation (old/new, traffic, country, behavior etc.) ● Mix overlapping / chained tests ● Multiple comparisons problem ● Simpson’s paradox
  • 43.
    Male Female 568 of2000 (28%) 372 of 2000 (19%) Simpson’s paradox
  • 44.
    Male Female 568 of2000 (28%) 372 of 2000 (19%) Highly competitive subjects 168 of 1200 (14%) 270 of 1800 (15%) Lowly competitive subjects 400 of 800 (50%) 102 of 200 (51%) Simpson’s paradox
  • 45.
  • 46.
    А/В testing featuresin apps ● Why do you need to split the traffic on server-side? ● If there is no internet? (А/А-test in apps 50/50? Case about custom settings)
  • 47.
    Calculator: bit.ly/2JswQWc t.me/BigQueryThanks forattention! Telegram channel about А/В-testing: t.me/ABtesting