In the presentation we focused on brand keyword bidding. In this case, paid clicks potentially substitute organic traffic. The advertisement effect we observe is the total AdWords traffic which includes not only the truly gained traffic but also the clicks of individuals that would have visited the website even in absence of SEM. The latter is what economists call the missing counterfactual. Missing as we do not observe it. Without this data it is impossible to calculate the true ROI.
The way to find the missing counterfactual is through controlled random experiments. The presentation contains a complete AdWords optimisation toolbox in GNU R necessary to run controlled experiments.
Turn Digital Reputation Threats into Offense Tactics - Daniel Lemin
Google Analytics and AdWords optimisation with GNU R
1. Google Analytics and AdWords optimisation with
GNU R
Hinnerk Gnutzmann & Piotr Śpiewanowski
flexponsive UG
Booster Conference, 9th March 2016
2. About flexponsive UG
• e-commerce consulting
• Big Data focus
• Qualitative user testing
• Academic (PhD in economics) and programming background
Contact
• mailto: spiewanowski@flexponsive.net
• web: https://www.flexponsive.net/
• t: @flexponsive
3. Topic of the day
• Marketing outcomes
• difficult to define
• even more difficult to measure
• Before Big Data: “Half the money I spend on advertising is wasted;
the trouble is I don’t know which half.” (John Wanamaker, 1838 -
1922)
• With Big Data: “AdWords brand keyword ads have no measurable
short-term benefis” (Blake et al., 2015) - 100% wasted?
• Open Questions:
• Incrementality Debate: Do AdWords campaings cannibalise organic
traffic?
• Quality: Are bought visitors good or bad customers?
• Heterogenity: Campaign effects differ between customers?
4. Agenda
1. Case study Brand Keyword: The Secret of vanishing AdWords ROI
2. What can we do?
• attribution models
• controlled experiments
• GNU R & Analytics: A Dream Team
3. How to do that?
• Google Core Reporting API & GNU R
• GA Query Explorer
• Configuring an experiment in AdWords
4. Analysis with GNU R
• Data wrangling, sampling, etc.
• GA replicate metrics
• Regression Analysis
5. Case Study II: adClicks and rain in Bergen
7. What happened?
• the AdWord is highly relevant to the search
• Navigational Query: The visitor wants to visit Skandiabanken.
• Customer knows the bank and maybe even has a service in mind
• Result: Probably the best keyword in the account
• Excellent CTR
• Very good conversion on-site
• CPC perhaps not so high
• Any questions?
• Organic result is the same!
• What would you click if there was no ad?
8. What happened?
• the AdWord is highly relevant to the search
• Navigational Query: The visitor wants to visit Skandiabanken.
• Customer knows the bank and maybe even has a service in mind
• Result: Probably the best keyword in the account
• Excellent CTR
• Very good conversion on-site
• CPC perhaps not so high
• Any questions?
• Organic result is the same!
• What would you click if there was no ad?
9. What happened?
• the AdWord is highly relevant to the search
• Navigational Query: The visitor wants to visit Skandiabanken.
• Customer knows the bank and maybe even has a service in mind
• Result: Probably the best keyword in the account
• Excellent CTR
• Very good conversion on-site
• CPC perhaps not so high
• Any questions?
• Organic result is the same!
• What would you click if there was no ad?
11. ROI
Problem: SEM expenditure a function not only of the campaign, but also
of the behavior and intent of consumer
12. The eBay study
• Blake et al. (2015), “Consumer Heterogeneity and Paid Search
Effectiveness: A Large Scale Field Experiment”
• Field Experiment: Does AdWords work for eBay?
• Very controversial results:
1. Conventional methods used to measure the causal (incremental)
impact of SEM vastly overstate its effect.
2. True effectiveness of SEM is small for a well-known company like eBay
3. Click substition: When the brand keyword AdWord disappeared,
almost all the users click on the organic result
4. Informative Advertising: AdWords work if a visitor gains additional
information through advertisement - AdWords had almost no effect on
revenues from existing customers - They found their own way to eBay!
13. What can be done? Attribution modelling
But how to know the true channel’s impact?
14. Attribution modelling
• a way to divide the “credit” for a sale between different marketing
channels
• if you don’t know what attribution model you are using, it’s “last
click” => you believe the sale only depends on the last ad the
customer saw before purchasing
• probably that’s not true: perhaps the customer had been following
the company blog for a long time, heard friends talk off-line about the
product, or saw many banner ads on different sides before making a
purchase
• problem: no good way to decide how to “attribute” between different
marketing channels
• results depend a lot on assumptions, which you cannot test
• similar problem: if you advertise your brick-and-mortar store on TV
and on radio, what drives the customer to your store?
15. What can be done? Controlled experiments
• Select by random treatment and control group, for example:
• Per user: A / B Testing
• By Geographical Region
• Assumption: Without experiment, both groups behave similarly
• Evaluation: difference in differences
• difference in the control group: Noise
• difference in treatment group: Effect + Noise
• Metrics: ∆TREATED − ∆UNTREATED
• Advantages of a geographical experiment:
• no multi-device tracking necessary
• easy integration with external data
• Caveat: Geographical groups really need to be comparable
(e.g. commuters)
17. GNU R and Google Analytics: Dream Team
1. Selection of the treated and control group
• Install R, generate a sample with GNU R
• Export: Copy & paste to AdWords
2. Data collection
• Google Analytics already configured
3. Aggregation and query
• In the cloud: Google Analytics Query Explorer
• Integration with RGoogleAnalytics
4. Evaluation: Estimation and Visualization
• All necessary functions available as packages in R
18. About R
• Programming language and software environment for statistical
computing and graphics, a dialect of S
• Quite lean; functionality is divided into modular packages
• Graphics better than in most stat packages.
• Useful for interactive work, but contains a powerful programming
language for developing new tools (user -> programmer)
• Very active and vibrant user community; R-help and R-devel mailing
lists and Stack Overflow
• Markdown packages for reproducable research and automated
reporting
• It’s free!
19. Install R
• Open Source for Windows / Mac / Linux etc.
• GNU R: https://www.r-project.org/
• RStudio IDE: http://www.rstudio.com
• Cheat Sheets to help!
• R Reference Card
• RStudio cheatsheets
• Package management via CRAN
install.packages('RGoogleAnalytics',
repos = "http://cran.no.r-project.org");
install.packages('plm',
repos = "http://cran.no.r-project.org");
install.packages('ggplot2',
repos = "http://cran.no.r-project.org");
20. Selecting Treatment Group
download.file('https://goo.gl/qVgiYp',
destfile='geoid.csv');
#Kommune level selection, but Fylke level also possible
regions <- read.csv('geoid.csv');
norway<-regions[which(regions$Country.Code == 'NO'
& regions$Target.Type == 'County'
& regions$Status == 'Active'),];
set.seed(1);
norway$isTreatment <- sample(c(0,1),
nrow(norway), replace =T)
write.csv(norway, file='norway.csv');
# paste into AdWords
writeLines(as.vector(
norway[which(norway$isTreatment == '1'),]$Canonical.Name),
file('treatment.csv'));
28. Google Analytics Core Reporting API & R
1. Create an “app”
• Google Developers page
• Enable Google Analytics API
• Create Credentials: OAuth client ID, Application type: Other
• Result: Client ID and Client Secret
2. Find your GA Profile ID
29. Setting up GNU R
client.id <- 'xxxxxxxxxxxxxxx.apps.googleusercontent.com';
client.secret <- 'xxxxxxxxxxxxxxx';
analyticsProfileId <- '111111111';
# redirect to google, paste, code
require(RGoogleAnalytics);
token <- Auth(client.id, client.secret)
# save
save(token, file = 'gatoken.txt');
# next time
token <- load("./gatoken.txt")
ValidateToken(token);
31. Real Data Example - www.flexponsive.net
kable(head(ga.data))
region date medium country sessions transactio
Brussels 20160229 referral Belgium 1
State of Parana 20160229 referral Brazil 1
Baden-Wurttemberg 20160229 organic Germany 1
Baden-Wurttemberg 20160229 referral Germany 1
Rhineland-Palatinate 20160229 referral Germany 1
(not set) 20160229 (none) Hong Kong 5
34. Tip3: Avoiding sampling
> ga.data <- GetReportData(ga.query, token)
Status of Query:
The API returned 1393 results
The query response contains sampled data. It is based on
XX.XX % of your visits. You can split the query day-wise
in order to reduce the effect of sampling.
Set split_daywise = T in the GetReportData function
Note that split_daywise = T will automatically ....
• “Sampling occurs automatically when more than 500,000 sessions
(25M for Premium) are collected for a report, allowing Google
Analytics to generate reports more quickly for those large data sets.”
35. Data Integration
• Wide Format: for each region and time a row
• Long Format: Region / time / dimension one line (EAV)
require (reshape2);
## Loading required package: reshape2
w <- reshape (ga.data, timevar = 'medium',
idvar = c( 'region', 'date'), direction = 'wide');
36. Data Integration: Almost finished
• Merge: Who is in which group?
ds <- merge (w, norway[, c ( 'Name', 'isTreatment')],
by.x = 'region', by.y = 'Name', all.x = T)
• Data set is ready!
• Comfortable DSL for data manipulation
• Use packages to minimize code
38. Case Study: Wanderlust
• an app “developed” for this presentation
• mysterious weekend getaway and short holidays booking engine
• supports inventory management of hotels and airlines
• seasonal demand fluctuations
39. Evaluation
• Simulated data for illustration: 3 summer months
• 1st August: experiment starts in 10 random provinces (fylke) -
AdWords stopped
• 1st August: start of school, search volume falls everywhere by 50%
• Scenario: 100% of visitors click organically when the AdWord invisible
• Randomization has decided:
• Sor-Trondelag (Trondheim): In the treatment group - from 1st August
no AdWords
• Hordaland (Bergen): In the control group - AdWords continue
40. Revenues in Sor-Trondelag (treatment)
60
80
100
120
Jun 01 Jun 15 Jul 01 Jul 15 Aug 01 Aug 15 Sep 01
date
transactionRevenue.total
41. Revenues in Hordaland (control)
60
80
100
120
Jun 01 Jun 15 Jul 01 Jul 15 Aug 01 Aug 15 Sep 01
date
transactionRevenue.total
42. Revenues in both Fylke
60
80
100
120
Jun 01 Jun 15 Jul 01 Jul 15 Aug 01 Aug 15 Sep 01
date
transactionRevenue.total
region Hordaland Sor−Trondelag
43. ROI Calculation - standard regression
require(stargazer);
out <- lm(transactionRevenue.total ~ isTreatment.cpc,
data = sd.w)
stargazer(out, header=FALSE, type='latex')
Table 2
Dependent variable:
transactionRevenue.total
isTreatment.cpc −48.358∗∗∗
(1.350)
Constant 111.350∗∗∗
(0.569)
Observations 1,748
R2
0.424
44. ROI Calculation - standard regression
• Standard OLS regression with binary variable == comparing means
• But not the right ones. In this case:
Revenues = β0 + β1 ∗ treatment
• The treatment takes value 1 for the treatment group after the
AdWords were stopped in Sor-Trondelag, otherwise 0
• As a result β1 represents the difference between the average revenues
in Sor-Trondelag in August and average revenues in Hordaland and
Sor-Trondelag in June and July
• That’s clearly now what we are looking for!!
47. ROI Calculation - Difference in Differences
• Difference in Differences estimator using fixed effects model with
binary varaibles allows to calculate the true effect of the treatment
• Econometrically we estimate this equation:
Revenues = β0 + β1 ∗ treatment + β2 ∗ before + γ ∗ fylke
• fylke is a matrix of binary variables for each district
• before is a binary variable takes value 0 in a period in which AdWords
were running in all districts and value 1 in period in which experiment
was started in some regions
• treatment takes value 1 for the treatment group in the preiod in
which the experimetn was started, i.e. after the AdWords were
stopped in Sor-Trondelag, otherwise 0
• The estimation result reveals the true impact of AdWords on
revenues in this data set
48. Discussion
• The Missing counterfactual - we do not know what else could be
happening - help: Experiment
• Challenge: Big Data without Big Code - Google Analytics & GNU R -
Very rich toolbox
• Result: Differences in Differences can work - note assumptions
49. Table of Contents
Intro
Brand Keywords
The eBay Study
Calculating the true ROI
Brand keywords with R
Configuring Experiment
Using Google Analytics API
AdWords experiment: an example
Regression Results