A Brief Guide to Developing Direct Mail Creative for Testing Direct mail is our passion…
Often, our first step with a new client is to build a model. We take their current customer base and match it against a prospect database to identify shared characteristics. Then, we select prospects who are most similar to their existing customers for mailings from that prospect database.
Intuitively, prospects selected based on a customer model should outperform straight demographic or firmographic selected prospects.
Sounds simple, right? Having a data scientist running the process is a handy way to improve the likelihood of success. We’re lucky enough to have Sam Kublawi — one of the best — leading our efforts in this area.
In this post, we’ll pull back the curtain on some key steps he takes to enhance the likelihood of success while making a very complex process appear seamless to our clients.
Why Model for Mailings at All?
Here are the positives of a well-designed customer look-alike model for your mailings:
- Improved ROI: Targeting prospects similar to your customers means a greater likelihood of finding people who find your product and value proposition relevant.
- Prospect ranking: A look-alike model ranks prospects based on their similarity to your current customers. Selecting the top prospects of a look-alike model narrows your marketing effort down to the “crème de la crème” of potential prospects.
- Little or no cost to implement: We often negotiate waived modeling fees as long as the client agrees to purchase mailing data from the modeler.
- Convenience: It’s an opportunity to use your available customer data as a basis to find new prospects without having to conduct further internal analysis or outside research.
Add it up and it sounds like an easy and foolproof way to mail more effectively and increase ROI.
Do look-alike models ever fall short? Yes — obviously, there are no guarantees. But there are some important steps you can take to improve the quality of a model and the likelihood of in-market success.
How We Build Successful Models
This post is a guide, not a course, so these pointers are necessarily high-level. We’ll consider three broad categories:
- The data you provide as input for building models
- The data files we model against to find new prospects
- Assumptions and uses of model input and output data
Here’s a look at each of the above categories:
1: The customer data used to build the model.
Size matters. Building a model requires us to match your customer file to a large existing prospect database that incorporates thousands of consumer data points. Because match rates between customer and prospect database files can vary, we prefer a minimum of 5,000 records to build a model, and recommend 10,000 or more to add confidence to the build.
For large established companies, that’s usually not a problem. It can be for startups with a small number of customers, however. It’s especially tricky for companies that only collect email contact information. Since our analysis is derived from address-based data, we need to append physical addresses to that list. We’ll typically match 40-60% of emails, so in those cases, your customer file needs to be that much larger to perform the modeling exercise.
The attributes of your customer file will have a direct impact on the modeling results and, therefore, the prospect data it leads you to pull for a mailing. Here are a few things we’ll ask you to think about when pulling a customer file for modeling purposes:
- Does the customer “vintage” make a difference? If your customer base or your product mix has changed over time, you may want to omit the OGs.
- Can you calculate customer value? Sometimes it makes sense to model using purchasers of your most profitable product, or those who have exhibited the greatest customer LTV (Long Term Value).
- Is your sample pull representative of your customer file as it relates to your marketing objectives? A marketing discussion before a “representative sample” is pulled is always a best practice.
The goal is always to get the most representative customer file for modeling based on your current business and the objectives of your mailing.
2. The files modeled against.
The “quality” of the database(s) we match your data to is also important. Why is “quality” in quotes? Because while the actual data needs to be of excellent quality (accuracy, recency, etc.), none of that matters if your customer database has a weak match rate to that prospect database.
The best models happen when there is a strong correlation between the input data and the prospect database. We determine that based on the hit rate of customer data when compared to the records in the prospect file. The more records they have in common, the better the likelihood of a successful model.
Choosing a database for use in building a particular model requires an intimate knowledge of available databases and understanding the nuances that make each of them more or less appropriate for a given product, software or service.
Broadly, we look at four types of data sets to find quality prospects:
- Transactional databases
- Compiled consumer files
- Publishing/specialty lists
- Non-profit donors
There is a lot to understand about each list type, and that’s a topic for another blog. For now, each of these list categories has unique strengths based on from where, how, and how often they collect data.
Within data sets we’ll also want to make sure that we are selecting a prospect file that best aligns with the client’s business and mailing objectives. For example:
- For a fintech or finserv product, we’ll want to model against a database with exceptionally strong consumer financial data.
- If age is a critical indicator, we look at databases that corroborate age data from multiple sources to ensure the highest likelihood of accuracy.
We’ll build models concurrently among different databases and even within each database. Those are then back tested against the existing customer file. The models with the highest match rates to a customer file are selected for testing. The size of your test cells determines the number of models you can test in any one mailing.
3. Your assumptions and uses of the prospect data.
Perhaps the number one expectation to address upfront is that a customer look-alike model does not predict a response rate. We model customers with the goal of improving program performance overall. Only a model built with direct mail response data can be predictive of a direct mail response rate.
Another assumption that can prove costly is blind adherence to the mailing prospects the model selects for you. The model build is based on the interaction of thousands of variables. But it doesn’t know that the reason your customer base is 90% women is that you’re selling makeup. The model could lead to selecting a prospect file with a greater percentage of men. Age, income, and geography constraints are often important selection factors as well.
Somewhat related to the above point, it could be risky to pull key demos out of the model as sole guidance for targeting in other channels. Neural network modeling (AI) is based on the relationship between variables. Some target characteristics are going to be more dominant than others, but we recommend a single variable demo/firmo analysis of your customer data rather than a model for straight demo targeting.
Look-alike models are based on a specific point in time. As your business evolves, so will your customer base. Models need to be rebuilt regularly with updated customer data to capture these shifts.
Finally, keep in mind that the whole idea of modeling is to limit who you target. Modeling could omit smaller pockets of potentially profitable prospects. That’s why testing specialty lists in addition to modeling is a good practice. And why we sometimes model separate customer segments to get more granular with our targeting.
Overall, look-alike modeling can be a powerful tool for improving direct mail performance to meet customer acquisition goals. There is no one-size-fits-all modeling solution. Work with an experienced partner who can lend analytic insights plus deep knowledge of available prospect databases to optimize your program.
Gunderson Direct has long-lasting relationships with some of the country’s largest corporations, helping them to lower their customer acquisition costs and increase profits using address-based integrated direct marketing programs.
Drop us a line for more information on how our direct marketing expertise can help your business.
If you found this information valuable, sign up for our newsletter.