MarketSoft1

Summary

The campaign managers of a leading SUV website knew they had one of the best websites – well designed, attractive and informative. While they were generating quite a number of leads (those that filled the Test Drive Form) they were exploring ways and means to improve performance. The question was, would a machine learning program and resultant code, deployed in real-time, increase the leads? Yes, it did – with a potential of lead increase up-to seven times. Marketsof1 deployed a Support Vector Based machine learnt algorithm in real time to spot customers who were most likely to convert and that helped unleash a carefully orchestrated program to convert 3 times more within a month of deployment.

Problem

The campaign managers at the digital agency of a leading SUV website knew they had one of the best websites – well designed, attractive and informative.

The site was well laid out. It had about 30 pages of rich information such as:

Features
Specifications
Price
Gallery

Design
Inspiration
Reviews
Dealer
Locator
Adventure sections

Most optimized marketing programs in SEM and SEO led users to land on any page. Being a highly competitive category the campaign managers were constantly looking to improve performance of leads (those that filled the Test Drive Form). There were several questions that Google Analytics could not answer comprehensively. For e.g., did path or user journey matter? Is there a geographical predominance? What are the typical user behavior that is associated with leads? Google Analytics did provide key insights even though post-facto. Only about a third of the page had any visits. A specific page “Buyer’s Guide” was the most visited (over 75%) and almost all of the leads (98%) had visited that page! Should they target and retarget them visitors to that page? The issue was that the converse was bleak: 98% of the visitors to that page did NOT become leads either.

So whom should they target? And can they target them in Real-Time?

Method

Two months of transaction data or visits raw data was provided. The data had user identifier (cookie) and the page visited and the previous page visited. It also had several pointers to the customer such as city, state, country, ISP, speed of the connection, whether the access was from home of business and time of visit.

The transaction data was rolled up user-wise and nearly 20 new variables where created to enrich the analysis, such as, Total Pages Visited, Total Unique Pages visited, Time of Visit, Total time spent on site and Path.

Support Vector Machine, a supervised learning model was used to analyze the data for regressing / deciding who the most likely visitors where that would become leads.

The data was split into training set and validation set – a technique to check the efficacy of the model. Due to very low leads to visitor ratio, SMOTE technique was used to reduce the negative impact of the imbalance.

Results

The model provided credible answers to the question " who will become a lead " to the extent of 93% accuracy. Owing to keeping confidentiality and competitive advantage, let is suffice here to say that there were clear answers to whether path mattered, which pages, number of unique pages visited, total time spent, etc.

The key aspect what that the technique identified from as little as 10% of all the visitors, 7 times more leads (than the current levels) who were most likely to have become leads. The subsequent targeting and retargeting of these leads lead to threefold increase in actual leads. The capability to do so now is in real time.

Deployment

The managers deployed the SVM algorithm in real time in the website. It continuously remembers all the actions of the user on the site (even for several months at times) and continuously applies an algorithm to compute in real time the propensity of the user to become lead. Once the algorithm determines so a user to be having very high propensity, the campaign managers have devised a different set of campaigns to ensure that the user does become a lead.