Using Predictive Analytics on Web Data
Analytics around Web data have been a standard part of a marketer's tool kit for years now. Free tools like Google Analytics have made it easy to track basic website metrics. At the same time, the market for paid Web analytics tools is growing rapidly, and is expected to generate close to a billion dollars in revenue by 2014.
Analytics has been a powerful tool even for companies not on the Web. In a Harvard Business Review publication, "Competing on Analytics," Thomas Davenport has examined how companies gain an edge using analytics. His research, which looks at companies like Marriott, Harrah's, and Capital One among others, highlights the role of analytics in functions ranging from supply chain and R&D to customer selection, loyalty, and service.
The key to the success for many of these companies -- and what companies of all sizes can learn from -- has been to not only look at metrics retroactively to analyze what happened, but also to develop models to predict optimal offerings for the future. In Davenport's words Marriott "has developed systems to optimize offerings to frequent customers and assess the likelihood of those customers' defecting to competitors," and the UPS Customer Intelligence Group "is able to accurately predict customer defections by examining usage patterns and complaints. When the data point to a potential defector, a salesperson contacts that customer to review and resolve the problem, dramatically reducing the loss of accounts."
This kind of analysis, known as "predictive analytics," is still not commonly employed at Web companies, or commonly available in Web analytics tools. Most of the common analyses and tests done by Web companies treat are centered on the notion of "visitors" to their website (transactional, one time relationship with consumers, typically driven by traffic coming from search engines) rather than "users" of their service (longer term relationship, typically involves creating a user account with the Web service). For lack of better terminology, I'll refer to these two distinct notions as "visit-centric" and "user-centric" models of the world.
In a visit-centric model of the world, common metrics include "time spent," "click through rates" and "conversions". Free tools like Google Analytics are widely used for these analytics. In a user-centric view, the metrics tend to be somewhat different. In his post "start-up metrics for pirates," blogger Dave McClure talks about the five steps in the customer lifecycle of a user-centric model: Acquisition, Activation, Retention, Referral, and Revenue (AARRR). Others have talked about using methods like cohort analysis to measure whether these metrics are improving for progressive cohorts of users.
Not only are the metrics different, the methods for optimization that are applied in the visit-centric world don't suffice in the user-centric world. In the former, you could run A/B tests to optimize "landing pages" and improve metrics on click-through rates and conversions per visit. Google Website Optimizer is a free tool for performing this optimization. In a user-centric world, a/b tests continue to be important. However, an important range of questions cannot be answered well, or optimized using a/b tests. For example, it's hard to associate metrics for users' repeat visits with the immediate impact of changes on any particular page. Some other examples of problems that are difficult to answer or solve using visit-centric analysis include:
- How often should you send marketing communication to your users, and who are the customers who are likely to respond positively to this communication?
- What actions, over the long run, help reduce user churn and increase repeat usage or increased revenue per user?
If a/b tests aren't going to be sufficient, what other analytics should companies build? And if revenues are going to materialize over a period of time, how can companies make sure they don't wait for months to understand the effectiveness of marketing dollars spent today?
That's where predictive analytics comes in. Historical customer data, which includes behavioral, transactional and demographic data, is mined to develop a model that predicts future behavior. Analytics companies like SPSS and Prediction Impact have talked about how to use predictive analytics for developing actionable predictions for each customer and decision optimization. These methods clearly need rich data about customers, which can form the basis for modeling. The good news is that it's possible for user-centric Web companies to have rich and high-integrity data about signed-in users who have created rich profiles. This is not true for visit-centric companies where predictive analytics runs into a lot of challenges, as noted in this post (note that all the challenges mentioned by the author implicitly assume a visit-centric model -- e.g. the author talks anonymity, or the inability to tie customer data to customer attributes).
It's interesting to note that other industries that have a similar view of the customer lifecycle as the AARRR model have used predictive analytics to good effect. For instance, in the travel and hospitality industry, predictive analytics techniques have been used for acquiring customers in a cost-effective manner, for fine-tuning marketing offers that increase repeat-usage, and for increasing revenue through effective cross selling and better yield management. In the mobile industry, similar techniques have been used to reduce churn (i.e., increase retention) and improve profitability.
Savvy Web companies have also started using these techniques. For instance, Facebook has used R in predictive analytics to answer questions like "Which data points predict whether a user will stay? And if they stay, which data points predict how active they'll be after three months?" The gaming company Zynga split its analytics team into two to become more proactive about analytics; while one team does the conventional reporting, the other "tests hypotheses and creates models using statistical and analytical methods." Looking at their success, it's clear that many others will follow suit.
If your Web service and your business model are built for repeat usage, you are probably already measuring metrics across the entire customer lifecycle from acquisition to repeat usage and revenue. By using predictive analytics, you can lower your cost of acquiring users, ensure sticky customers, and increase your revenue, just like Facebook, Marriott, and UPS.
Vijay Chittoor is a co-founder of Mertado Social Deals. He was previously director of product management at Kosmix. A former McKinsey consultant, Chittoor is a graduate of Harvard Business School and the Indian Institute of Technology, Bombay. He shares his thoughts on technology at his blog.