2002 May 01
Paramark PILOT
David M. Raab
DM News
May, 2002

“Optimization” is one of the few genuinely successful buzzwords in marketing technology today. It captures the desire to make the most of existing resources in an era when new resources are hard to come by. More subtly, it also implies a degree of control that many people would love to believe they have, even though they know they don’t. It’s nice to imagine yourself fine-tuning the dials on a marketing control panel–especially when your real job is more like riding a bronco in a rodeo.

But success has its price, and for a buzzword the price is clarity. “Optimization” does have a specific meaning: it is a mathematical technique that identifies the set of rules which best achieve an objective within a given set of constraints. But the term is now used much more loosely to describe any method for assessing multiple options.

PILOT (Paramark, 408-830-5993, www.paramark.com) is a good case in point. PILOT stands for Paramark’s Interactive Learning and Optimization Technology and is described by the company as a “real-time optimization platform”. So optimization is at the heart of the product’s market position. Yet what the system really does is automate the champion/challenger testing process. Atlthough this is optimization in the loose sense of finding the best choice, it lacks the constraint-balancing that characterizes true optimization.

Of course, what PILOT does is more important than what it calls itself. Its core function is to select offers: an external system such as a Web site presents PILOT with a customer and PILOT returns an offer ID. Later, PILOT receives information on the customer’s response. What’s special is that PILOT automatically identifies and recommends the offers that are most successful.

This isn’t as simple as it sounds. The system starts by conducting random tests of all offers, and evaluates results until it can declare a winner. PILOT calculates both the probability that a customer will accept each offer and the statistical confidence interval of the probability estimate. This ensures that the winner is declared only when it has truly been proven superior. PILOT will also restart testing if the winner’s results fall to a level where it appears another offer might be competitive. This lets the system automatically adjust to changes in customer behavior. Calculations and decisions are updated immediately each time a new interaction is processed.

PILOT can speed the learning process by pooling results from separate customer sources, such as different Web sites. The system reduces the weight assigned to this data as it builds its base of information within each source.

The definition of success is also fairly sophisticated. PILOT can target a discrete measure, such as orders placed, or a variable measure such as order value. Users can also assign weights to different measures and have the system maximize the combined result.

Another advanced function lets the system automatically identify customer segments that are most responsive to different offers. The segments are based on customer attributes, which can be presented to PILOT with each selection request or be loaded in advance to a separate customer database. Either way, PILOT will automatically assess which attributes correlate with which offers and will create new customer segments based on these attributes when appropriate.

Other features of PILOT are less exotic than its automated selection engine but still important. Users can set up multiple advertising campaigns, each with separate subcampaigns that might represent different advertising buys, areas on a Web site, or types of emails. Campaigns and subcampaigns can be assigned start and stop dates, budgets, maximum numbers of impressions, and system-specific parameters such as how long to wait before changing the winning offer.

Users also specify the set of offers available in each subcampaign. This lets them limit certain offers to specific customer groups. The offers themselves are assigned attributes defined by the user–such as color, positioning, and product. System reports show which offer attributes are most strongly related to customer response–providing useful insight into customer behavior.

PILOT does not use the attributes directly within offer selection, however. Instead, it tests each offer independently. This limits the number of offers that can be tested simultaneously: if there are too many, the system will never run enough tests to identify a clear winner. The actual limit depends on the the number of interactions and differences in offer performance. PILOT has optimized up to 100 offers in a campaign, but could probably handle many more.

Other reports provide logs of user- and system-generated events, show where and how often each offer was used, and display performance by campaign, subcampaign and offer. An “optimization” report estimates what results would have been if all offers had been used equally. It compares this with actual results to show what the vendor interprets as the improvement due to optimization. All reports are delivered as Web pages and allow the user to select options such as date ranges and metrics. Many provide graphs as well as tabular formats.

The current version of PILOT is run as a service hosted by Paramark. Users set up campaigns and offers via a Web browser. Integration with external systems varies slightly by application, but generally comes down to having the external system send transactions to PILOT and receive the Web address (URL) of the selected offer in return. Results are similarly tracked by sending transactions to PILOT when responses are received. The system uses cookies or session IDs to associate selections and results with a specific customer. PILOT itself does not store any customer data or transaction history, although it would be possible to load a customer database with attributes to help in making selections.

PILOT was originally developed to optimize online advertising campaigns. It has since been extended to handle Web sites and email and the vendor is now working on a version for call centers. The product was launched in late 2000 and has three major clients plus a number of pilot projects. Prices are based on volume and application and generally run $250,000 to $500,000 per year.

* * *

David M. Raab is a Principal at Raab Associates Inc., a consultancy specializing in marketing technology and analytics. He can be reached at draab@raabassociates.com.

Leave a Reply

You must be logged in to post a comment.