How to Use Big Data in Your Digital Marketing

Stephen Hammill
Marketing Technology Director

Part 1 : Descriptive, Prescriptive, and Predictive Analysis

So let’s say you came here because you work in marketing and are tasked with helping your client or business make better sense of all their data. No problem! You are an Excel master and can run pivot tables around the competition.

There’s just one problem. There’s too much data — millions of rows. You can barely open the CSV files without crashing your computer, nevermind create meaningful insights from it. Your IT department can help, but writing SQL queries can be laborious, and you are a marketer, not a data scientist. Moreover, the data files themselves are not playing nice with each other.

You have Big Data problem, and you need a plan to deal with it.

Let’s define Big Data here as multiple sets of data, not easily combined, and too big to analyze through traditional means. How do we build it up and break it down? We start by looking at some steps through the lense of a real world example: a big brand, consumer facing vendor with millions of loyal customers and thousands of locations around the world.

By uploading clients’ data files into the Cloud, we ease the strain on our systems, and greatly decrease the time necessary to run queries.

Step 1 : Descriptive Analytics

It sounds great, an endless stream of details — orders, revenue, traffic, conversions — expect much like a library of books in an unfamiliar language, much of the data remains locked away from the decision makers who could make the most use of it. This is where the descriptive phase comes in.

This client’s trove of user and sales data went back about a year, but with no internal means to organize it or allow for the creation of any audience segments, and audience segments are what they needed.

Since these files were too large to be analyzed through a traditional software like Excel, and since we didn’t have a robust set of servers to allow for internal database builds, we instead looked to cloud-based options.

This is where Google BigQuery comes in for us. We use BigQuery to house all of our larger data sets for faster recall. Since those massive fields aren’t taxing our agency’s computers and servers, we can instead focus on moving into analysis more quickly. (Note: There are dozens of great cloud-based and local database systems out there. Since eROI is a Google house, this tool works well with many of our analytics platforms.)

Once uploaded, we can start defining the data tables, making sure duplicates and missing fields are handled.

For this client, the order data had a unique ID, but it didn’t match the IDs from customer data files. Let’s cross reference the email addresses to match IDs from one file to the other. Success!

Next we want to establish clear baseline metrics, ones that can be entirely attributed to either a specific channel or campaign. This will be crucial when it comes time to forecast on our findings. For our big brand client, luckily, their conversion data was already broken down this way, so we could easily attain baseline conversion rates.


With that complete, the data is now clean and accessible, but not visualized or ordered in an obvious way. While we could certainly export the now filtered audience segments back into a spreadsheet, what we want to be able to do is look for patterns in the raw data to create our own findings.

Step 2 : Prescriptive Analytics

This step can be interpreted a few ways. Maybe the client has asked for specific findings and insights from their data to match up with already established criteria. Or maybe it’s up to you to determine what rules to apply and set up that framework for the client (we do this a lot).

Then there are times when the client will ask for something based on an internally-held set of beliefs or assumptions, and we discover that they were coming to a faulty conclusion. With our big brand client, we had data telling us that a younger segment of customers for a recent campaign were spending far less money per order than the whole. This might lead us to assume that younger customers tend to spend less by virtue of their age bracket.

The problem with that assumption is it ignored another element. During the same time window, our client was offering a free item to all customers. Looking closer at the data we saw that younger customers were twice as likely to redeem this free item, which, in turn, was reducing their check size. It was a small finding, but big in implication. Further internal testing proved it out.

From here, our smart customer segments can evolve from the client’s assumptions to illuminate new buying behaviors. In this particular case, we created a new segment of users who love free offers and customization options.

We use data visualization tools like Tableau to tie directly into our datasets, allowing for fast visual cues of trends and outliers, and to help us eliminate unnecessary variables.

Avoiding confusion between correlation with causation (two things happening concurrently does not mean they cause each other to happen), is easier said than done. The real trick is to eliminate variables whenever possible, while not eliminating the ones that have direct impact on what you’re analyzing. I like to use the analogy of a staged murder mystery, with the data standing for various characters found at the scene of the crime. Some are innocent bystanders, who can be disregarded rather quickly; others are potential suspects. Maybe the crime was committed by two people in conjunction.

As data analysts we are detectives before anything. And like good detectives, we marry our experience with observations, always asking questions, even when the answers might seem obvious.

Step 3 : Predictive Analysis + Forecasting

Chances are your client is less interested in seeing their data correlate to past or present performances as to future growth and sales. The question is always “what can these insights do for us going forward?” This is where marketing data analysis can truly shine, in our ability to make highly accurate predictive statements about future behaviors and performance.

With our example client, with nearly a year of steady conversion rates and traffic patterns for their app, we could confidently make assumptions for revenue growth broken down into the newly created segments.

Our agency insights come at the intersection of data sets, in this case even including data captured apart from our client.


You may recall during the 2008 presidential elections how a blog called FiveThirtyEight managed to predict the outcome of the Electoral College with better accuracy and much sooner than nearly all the traditional political pundits.* The reason for their success lay in predictive analysis. Their poll aggregating models were devoid of personal biases, options, or “gut” feelings that were the norm for decades.

If you’re the one doing the analysis for a company or client, what you find in the data may not always jive with current theories or assumptions. How you communicate those new insights is just as important as discovering them.

*They were so good at this, and other predictive analyses, that they were purchased by Disney-owned ESPN in 2013.

Coming soon: more tips on data visualization and how to make dashboards that actually work.

Stephen Hammill
Stephen Hammill, Marketing Technology Director at eROI.