The question I attempt to answer in this post:
Is there a scientific way of segmenting customers based on a number of dimensions?
We all know that we can plot the shape of a curve on a two dimensional graph or draw the shape of an object on a three dimensional graph. However once we have crossed the number of perspectives to be more than three dimensions the mind starts to wonder how one can visualize the shape of the object. Trying to model the outcome of a process that has multiple dimensions is more complex than can be represented in a euclidean space.
Even more difficult is to find the optimum of that shape. So let us say that we wanted to find the lowest cost of marketing to the community of a million customers. Further we know that they interact with us via multiple channels, such as web browsers, email, chat rooms, call centers, mobile phones, tablets. We also know that the process of communication is either initiated by the sender ( marketer)or the receiver (prospect or customer). The stage of the receiver’s buying cycle also has an influence on the outcome of this interaction. The awareness of the brand, the price sensitivity, the affluence of the receiver, the promotional offers on the table are just a few of the influencing factors. The array of factors that influence the outcome of this marketing game are too many to articulate. So how do we model this complex world of b2c or b2b marketing?
Early in our life we have been taught to use a cryptic language to represent ideas. The language of mathematics. So very early on we learned that we can represent a straight line by an equation. We also learned to define the line by its slope, the height of the Y axis where it intersects it and a pair of points on a two dimensional graph by a set of points like (X,Y). We can use the same approach to represent the marketing scenario mentioned above by a multidimensional shape.
This process is called modeling. We attempt to fit the abstract representation of the real world, to an equation that defines the shape of this world. In marketing there are two questions we try to answer.
What is the probability of a favorable outcome ( someone buying something )?
What is the amount of revenue that can be generated if the outcome is favorable? In other words if one were to buy, how much will they buy?
The first question has a YES/NO kind of binary answer and the second question has a discrete number ($58.25) kind of answer. Both are estimates but their nature is different. So the technique to answer them is also different. The first is called modeling response and the second is called modeling revenue.
The first question is of classification of the outcome as a yes or no. The technique used for this is called logistic regression.
The second question estimated the return amount. The technique used for this is called linear regression.
Now humans are creatures of habit, so we assume that they will behave the same like they have been behaving under normal circumstances. In effect, they will tend to fall back (regress) to their habit. This hypothesis gave rise to the method of observing the behavior of the receiver and coming to the conclusion about the two answers we are seeking. Marketers and the statisticians have been using this method to observe the behavior of the last year’s buyers, and create a equation to predict the likelihood of purchase and better still, the amount of money likely to be spent by the customer.
By the use of these techniques we can create a score (very much akin to the FICO score we all are measured by lenders). We can then use this score to rank the customers from the highest to the lowest by the probability of buying. We can also rank the customers by the amount of likely money to be spent. These two scores themselves reveal a lot.
We could decide whom to send a marketing collateral and whom not to send it. Thus by holding off sending it to the lost causes or the sleeping dogs we can improve the return on investment of our marketing efforts. After all it costs real dollars for marketing.
We could also multiply the probability of buying with the amount of money likely to be spent by the customer, to create a final score for each customer. By ranking the customers top to bottom by this final score we can get the best of the best customers to market to. Now you know why some of us are such magnets for junk mail!!!
We sponsored a survey about Marketing Analytics in the cloud with The Relevancy group. Check it out.
Once you have identified the customer segment that you want to target, the next monumental task in front of the direct marketing folks is to grab the attention of the audience and keep it on the message that you want to convey. The composition of the email, starting with the subject line, the images in the body, the language in the text all can be tweaked to create the right effect. In this role you actually act like a movie director. Just like a movie director uses the light, sound, background score, the action scenes, the photographic angles, the expressions of the actors to create the right mood and setting for the story that he wants to tell, you need to harness all the skills that you have to engage the receiver.
Today, the email marketing industry is exploring the impact of the design by using multiple different versions of the email content. Matching the right content with the right segment of receivers does magic in improving the response rates of the email campaign. With email, the ability to try different combinations of the creative and monitoring the responses is quite inexpensive and instant, so we can experiment to find the effective email combination that works for a segment of receivers.
Subject Line testing and analysis
Generally the subject line testing is the first thing that can be tried with significant impact. Some of the questions that one can try to answer by testing are:
Is a short subject line better than a long one?
What is the optimal size of the subject line?
Is a question in the subject line better than a statement?
The objective here is to improve the open-to-sent metric. The more catchy the subject line, the more likely the receiver is to open your email. This is usually used with great effect by crime thriller writers—”Sex, Lies and Money” they say will always improve ratings. Perhaps, there is something to a catchy heading. Our objective, is to measure the improvement in the open-to-sent rate. So, let us say that you have created a set of four segments of customers, whom you are going to send the same email. Look at some of the following headlines used by Amazon, WSJ.com, Bloomberg etc.:
“Camera, Photo & Video Lightning Deals”
“Dear China, it’s over…”
“Is Best Buy a Better Sell?”
These subject lines tend to stand out in an already overcrowded inbox of the receiver.
Image Size Analysis
Another thing that has become very effective is adding images to the email message that draw attention to themselves while conveying the product you are trying to sell. As a marketer you want to test the impact of the size of the image. Perhaps a single image of the appropriate size, tells the whole story. Creating multiple versions of the email with varying images of different sizes can give you a set of different versions of the email. These can be sent to different groups of customers in a single segment. Use these versions of the email to monitor the response rates of the various versions.
Play with the Call to Action (BUY) Button
The next thing to work with is the “call to action” button that the receiver can click on. The size the color and the position of this crucial element can make a difference between the receiver buying the product or moving on to the next email. Try to place this at the top of the page, or the lower left corner, center of the page, right below the image etc.
When you combine these various elements of the email –header, images, call to action buttons etc. you can come up with quite a few versions of the same email message.
Then you can send the traditional email you have been sending to one group of the chosen segment, and send the different versions to the rest of the groups of your segment. Most Email Service Providers (ESP) will allow you to manage this effectively. The key activity is to gather the data from these email blasts and store them in a database for analysis. In a short four weeks you can have enough data to guide you into making some empirical decisions about the versions that work and the versions that can be discarded. This forms your baseline for monitoring the improvement in response rates by using the right version to the right segment. See figure 1.
Figure 1: Comparison of email campaigns by subject line
The merger of content versions and the segments of your target list needs you to track the responses over a period of time to find the optimum creative combination for each segment. So you need to track the individual email version sent, the rate of response of the receiver across the various response segments, and a large number of other variables that influence the outcome of your email campaign. You can’t achieve this without having a very large database for marketing department. This database should focus on the individual receiver, the emails sent to them, their response to various creative combinations. The size of this database can be significant and performance can be a major consideration. Suddenly the meaning of ‘BIG DATA’ becomes abundantly clear to you. A typical email campaign database can store upwards of a few billion records over a 3-5 year period. To develop long-term insight, the marketing department needs to start with a long-term plan. If it is not possible to have a VLDB database in-house due to cost and IT support considerations, search for a vendor who can provide such a database for you. Without this kind of database, your ability to improve the response of the campaign is quite difficult, if not impossible.
The current trend of cloud based services offers a very effective solution. In a future blog I will review the characteristics of a database engine which is most suitable to the marketing activities. Do you need a columnar or a row oriented database engine, a massively parallel database engine? The database technology does uniquely differentiate the marketing analytics solution.
Everything in marketing is about being relevant to the context of the customer. Having said that, when we set out on achieving this lofty goal we run into a number of challenges. One of the major consideration is the definition of the size of the opportunity.
Let us assume that we email every one of our email list of 10 million members.
Let us define response rate in marketing parlance.
Response rate = number of responders/number of impressions in the performance window.
Performance window can be defined as the time after the email reaches the inbox of the target audience. This can be as short as 24 hours or up to two weeks depending on the frequency of your waves of emails.
So let us say that
- we sent 10 million emails
- 500,000 people opened their emails
- and of these 5000 clicked at least one link in your email
- within the two weeks after we sent the email,
then our response rate would be 5000/10,000000 = 0.05%.
We could also look at another important ratio here.
Click to Open Ratio = 5000/500,000 = 1%. These two response rates are typical of the email marketing numbers. Let us see why.
The general notion is to send more emails to increase the number of responders. However, by the definition as the number of impressions increase the denominator of our formula increases thus reducing your response rate.
Our executive management reviews the response rates and mandates that we do better, or perhaps we don’t share these rates with our management as they are too low and proactively decide to do something about the response rates.
Let us rethink our approach
If we take the same “spray and pray” technology of the newspaper, billboard era, then we have not made any progress. The best part about the digital world is that technology allows us to sense and measure more signals than was possible in the offline world. If we don’t take advantage of this measurement we haven’t made progress.
So let us think about narrowcasting rather than broadcasting. I want to make an offer only to those people who may be interested. That way I can keep my number of impressions to a minimum and improve the number of responders. However, wishful thinking alone doesn’t make this happen. We decide to think different.
One of the team members has a bright idea,- why don’t we segregate our buyers from non buyers? surely, they behave differently. Then the discussion drifts along the lines of demographics and geographic segmentation. These dimensions are easy to use, and so, we dive in headlong in this approach. But we soon discover that in the next couple of our campaigns the response rates don’t budge. If you have been there, don’t worry, we all have been there.
A Better Approach
Let us start by looking at how we can understand the context of the buyer. We can start by segmenting the customers into a number of segments by the following attributes based on their past buying behavior. Some of them are:
- Buying stage
- Style preference
- Price Sensitivity
- Social acceptance
- Quality consciousness
- Return behavior
- Channel preference
- Opt In status
With Today’s technology, we can observe the potential customer unobtrusively over a long period of time and collect data about them. Once we have done that for a while, we can start to group them into various clusters. Some clusters can be as follows:
Just browsing/prefers to browse online/brand conscious/likes contemporary styling/wants to buy first/local < 10 miles/discerning of quality/last bought a year ago/last purchase was $300/Return behavior unknown/Professional/ male/retail buyer/Opted-In
Actively buying for self/Brand Neutral/Modern style lover/Balances price to performance/seeks recommendations/local <5 miles from store/compromises on quality/last bought 90 days ago/last purchase was $100/rarely returns/upper class/female/online buyer/opt-In
Such clusters, once defined allow us to predict the future behavior of new members to the cluster. As we gather important clues about individual customers, we can start to plan separate campaigns to address every unique cluster.
Each of these listed variables influence the buyer behavior to a varying degree. So, perhaps we can study the last two years of the behavior of this customer segment and calculate the rank of each variable by its influence on the buying behavior of the members.
We can also find the strength of influence of each variable on the outcome. This is the correlation between individual variable and the outcome. So let us try to find an equation that can explain in mathematical terms, the influence of these variables on the final outcome. Will the receiver respond or not? So we classify receivers based on their likelihood of response.
Once we have done this classification, we can calculate the probability of response. Now even within responders there is going to be a probability of response by each receiver. So we need to rank the likely responders based on their probability of response (e.g. 0.85). Then we can create a cut off threshold (let us say 0.75) . Anyone with a probability above this threshold should be emailed, the rest of them can be safely ignored.
This is one part of the equation. We have identified the segments of customers we want to communicate with. THE WHO of our story is defined.
However if we still need to establish the context of the receiver of our communication when we are going to email them. So we still need to try to explore the recent activities of the receiver. Our ability to track the receiver’s search terms, the visit to the various websites, the blogs visited, the products reviewed, the products they have pinned, the stores they have visited, give us their stage in their journey to purchase. This also gives us the clue about their stage in the buying cycle and what may be of interest to the receiver. This allows us to do some controlled experiments with the type of emails we can send them. It also answers the question -when is the best time to send them the emails.
This approach of using the data to guide us through the unknown territory of marketing new receivers is called Data Science. Using the historical data to analyze the habits of customers needs a scientific approach. It is methodical, time-consuming but the surest path to marketing success.
In a future blog, I will explore the world of subject line testing, the composition of the email, the contrast colors, the sizes of the images, the positioning of the call to action links. These activities make our emails more effective as marketing messages.
If you are like most marketing managers the top most thing in your mind is to generate more revenue with your marketing spend. Perhaps your performance is measured on it. In effect you are expected to invent the perpetual machine which takes no input but generates infinite output, or so it seems.
Most companies today use emails as a main way to communicate their marketing messages with their customers. Yet the response rates are so low that the gut reaction to poor email responses to your email marketing program is to increase the email frequency to improve the number of responders.
However this leads the customers to perceive your messages as an irritant if not spam. Even if you are not flagged as a spammer or an irritant there is a strong possibility that the value of your message is diluted, thus creating a longterm loss of brand value. Is there anything that an organization with a modest budget, can do to improve response rates without barraging your customers with unwanted emails?
Fortunately, there is a way to engage your customers without bombarding them with emails and yet improve the response rates of your email marketing.
Let us look at how you can segment your customers with whom you want to communicate.
Broadly speaking there are four different types of customer categories:
- The persuadables
- The sure bets
- The lost causes
- The “Do not Disturbs”
The persuadables are those that are likely to be seeking a product or service and are familiar with your brand and aware of your offerings. These customers are likely to welcome your email because it solves a problem they are trying to solve. Perhaps they are interested in buying a product you are offering and so your email seems to be well-timed. Here the need is met just in time, or there is an untapped desire, unspent disposable income that you can access by sending the right message at the right time to the right person. The persuadables are also the customers who will spend higher if targeted.
The sure bets are those customers who are very familiar with your brand and offering. These customers may buy irrespective of receiving an email/catalog/sms/coupon. You may potentially waste your money by sending them emails, or better still reduce the profitable revenue generated by your marketing program by offering them coupons. This is preaching to the choir.
The lost causes are those who are never likely to respond to marketing messages as they are either not interested in buying, or they have been won over by your competition. Sending them emails may be fruitless and you are better off trying your message somewhere else.
The Do Not Disturbs: The fourth category of customers are those who are likely to be loyal customers but who don’t want to be disturbed by frequent emails. Sending them emails is likely to turn them off. You can lose a good customer due to poor marketing. Generally these customers feel slighted that you don’t know them and get put off by your marketing emails. This is a risk you can’t afford to take, as it would mean losing a good but infrequent customer who buys a lot when ever they get to your store or website.
The question by now you must be asking is all this is good but how do I segment my customers in these four categories. I will get to the process of effective segmentation later first let us look at the historical and current situation.
Historically experienced marketing managers have developed an intuition based on observing the behavior of their customers.
- When did they last buy?
- How frequently do they buy?
- How much do they buy?
The trade term for this formula or expertise is called RFM (Recency, frequency and monetary) value. For years this has been a mechanism used to segment customers by these three dimensions and target them with marketing messages. But this technique has been overused. Along with this the avenues for buying have increased significantly as well. Besides retail stores, there are now e-commerce sites and mobile apps where the buyer can exercise the right to buy. They can buy in their bedroom late at night, in their pajamas, or buy while they are riding a car during their daily commute. So the customer is getting empowered to buy anything, anywhere, anytime.
The advertising influences on a customer are increasing multifold. Google search, ratings and reviews, social media bragging by friends about what a great deal they got are routine. So what in your marketing really worked? What can ou attribute the sale to? This is the holy grail of marketing today. My point is that just a three-dimensional analysis of customers doesn’t give enough insight into their buying behavior. Obviously a better way to analyze customer behavior is needed.
Over the years direct marketing companies have used predictive modeling for creating multiple segments of the customers based on a large number of variables that are likely to influence the buying behavior. Obviously you couldn’t mail the catalog to all the people in the country as it costs real money to get the catalog in the hands of a customer. Even if it costs $0.50 per catalog to send a 50 page catalog to a customer the numbers quickly add up when you mail the whole population multiple times a year. Hence the need to improve targeting. Marketing managers have developed very deep expertise to increase the return on investment of the marketing dollars. In direct marketing the predictive modeling is used to calculate a purchase propensity score (the probability of purchase multiplied by the amount of money the customer is likely to spend) for each customer. This gives a sense of the success of the campaign before any mailing is done. Use of this technique has not been applied to email marketing mostly due to the cost of modeling and scoring the customers. There is also a notion that it costs very little (at least relatively!) to send an email blast, so I might as well send it to every one of my customers.
Both the cost of the modeling and the almost negligible cost of emailing have kept this approach from being used for email modeling.
However our experience over the last few years has been quite the opposite. Typically most companies are happy if they get 1%- 2.5% rate of response to their email marketing. But using the approach I am about to outline, we have experienced response rates in the 12-15% range. Initially, when we reviewed the numbers we didn’t buy them, but when the rates continued to keep coming up again and again it became conviction that we are on to something.
In the next few posts I will attempt to articulate this approach and look forward to your feedback. What we will attempt to learn together are the issues involved and how to overcome these to attain the marketing nirvana of “sending the right message to the right customer at the right time based on their moods, likings, buying stage and buying behavior”. Stay tuned.
[contact-form][contact-field label='Name' type='name' required='1'/][contact-field label='Email' type='email' required='1'/][contact-field label='Website' type='url'/][contact-field label='Comment' type='textarea' required='1'/][/contact-form]