Jim answers questions from fellow Drillers
Hi again folks, Jim Novo here.
This month in the newsletter we answer questions on the nitty gritty of the actual discovery work by taking a very deep look into the whys and hows of segmenting customers. Straight-up and to the point, put on those data shoes and Let’s do some Drillin’!
Q: Hi Jim, I’m a great fan of your work!
A: Well, thanks for your kind words.
Q: I have a basic question for you. We are an online retailer and thus use email as the primary marketing communication channel (we do use Direct Mail to our best customers around holidays).
A: Those are smart choices. I’ve seen some stats on using direct mail to drive lapsed online customers either back online or into a store that are very encouraging, real money-makers for retail. Definitely worth testing, though in both cases, the product mix averaged higher ticket than your category typically does.
Q: However, we don’t have a set customer segmentation technique and thus no specific customer segments. One outside consultant, a statistician, had suggested looking at a new customer’s activity in the first 30 days and then classifying them into High Spender, Frequent Transactor, etc. segments. Not sure how well it works.
A: That’s quite unusual, I think. It would work in the first 30 days, but I think you would have to re-classify every 30 days using a scheme like that. Considering web-only behavior, the typical retail lifecycle beyond 2nd purchase (many buy only one time) is a ramping to a peak and then a more gradual, but still steep, falloff in purchases. The model above would not take this into account, and while the initial label might be accurate, it soon would not be. That’s not to say these kinds of models don’t work, but it usually takes years of testing and study to perfect them. “Data miners” often believe the numbers will simply tell them things like this, but they don’t take into account the human behavioral and other mitigating factors which may not be in the data.
For example, Recency and Latency are really “meta-data” about customer behavior; they are data created from other data. You can’t just look at the first 30 days of transactions and give a customer a label; customers have LifeCycles and you drive the highest ROI when you take advantage of knowing these cycles and acting on them to increase profits.
Q:Â Â I feel that we target our customers primarily by their category purchases, and not by any kind of behavioral model.
A: Category is often a secondary indicator, and probably more useful along the lines of writing copy than the timing of a promotion or offer. Your industry is full of stories about mis-targeting by category, e.g. I bought a book about something I have no interest in as a gift and you keep making offers to me like I want to buy every book in this category. But it really comes down to “when” first, and then “what”. The highest ROI promotions are always about “when” – the timing of delivery. “What” is pretty much secondary, since a dollar is a dollar no matter what category it comes from. Put another way, if in the end, you want me to buy a book – any book – and don’t really care which category I buy from, then I’m not sure “category” is anything other than a copy hint.
The exception to this would be if you find **known patterns** of category trending and are using those to generate incremental sales. For example, category of 1st purchase can often predict value of customer over time. Let’s say you know on average, people who buy gardening books eventually either stop buying altogether or continue on and buy interior design books. Given this choice, I would screen for people who are decelerating in their purchases of gardening books and start making interior design book offers to them. Some will stop buying altogether, but some will convert to interior design buyers. If you use a control group with this kind of test you will find out how many people you transitioned to interior design that *would not have transitioned without your promotions*. These people represent incremental sales and profits due to the promotions – their defection was prevented, and that has very high value.
Q: My marketing management thinks that segmenting customers is not worth the effort, since the cost of email is so low! We have over 15MM customers, with about 5MM active (have bought in the past 12 months).
A: Well, that’s a typical attitude, and there is some truth to it if you only look at the cost of delivery. There are two other costs, one tangible and one intangible. The most common tangible cost in online retail is subsidy cost, that is, the cost of a discount you didn’t have to give to induce purchase, which impacts margin. Do you remember the “ramping” after 2nd purchase I mentioned above? It is common for online retailers to blow a ton of margin discounting to people in this ramp who would have bought anyway. If you use control groups you can literally see it happening before your eyes.
For example, let’s say you take a group of customers who made their first purchase in the same month due to some promotion or ad, and they have all made more than one purchase. You split this group 50/50 into two groups, the control, which gets no e-mails, and the test, which receives e-mails with discount promotions. Over the next 60 days, the control group spends an average of $200 per person and the test (promotional) group also spends $200 per person – except you gave them $20 in discounts, so their sales are really $180 for the period. Multiply that $20 times a million customers and all of a sudden you are talking about real money, know what I mean? That’s subsidy cost, and it is as real as e-mail delivery is cheap.
The second cost is more intangible, but it manifests itself through declining response rates and unsubscribes. It’s the cost of a shorter LifeCycle caused by delivering too many promotions too often. In other words, the cost of irrelevance. By the time the customer is preparing for defection, they are ignoring your e-mails because there have been so many of them that were not relevant to the customer. So just when you need to make that big splash to retain the customer, they are no longer paying attention and defect.
Q: What’s your suggestion on a good way of segmenting them and how many segments do you normally recommend? Also, how often do we re-run the segmentation technique you recommend? Once every 6 months, e.g.
A: Hmm… Well, I rarely try to guess these things and simply let the data speak to me. Whatever the right segmentation is will be revealed by the behavior of the customers themselves in the data. Since you’re somewhat familiar with my stuff you probably know that the heart of it is either Recency or Latency, and everything else from there is just a further sub-segmentation.
But even if the population on average is mainly driven by Latency, you will certainly find sub-segments where Recency is the primary driver. In the end, it’s about increasing profits, and as I said above, the profits in High ROI Marketing are usually a function of timing. One of the components in the equation is reducing subsidy costs to active customers; the other is squeezing more profits out of defecting customers on their way out the door.
How can you figure this out? Start looking for patterns. Here’s a very simple example. What is the average number of days between purchases? Let’s say it is 40 days. When you do your promotions, make smaller offers to those with a purchase less than 40 days ago and larger offers to those with a purchase more than 40 days ago. Two segments, the offer is correlated to days since last purchase.
After this promotion, inside each of those two segments, you have two sub-segments: responders and non-responders. Aggregate the members of each of the four groups and compare: what is similar or different about them? Categories, time of day, day of week, price point? Ad they responded to?
This is a Latency-based approach, which often works better when the sales process is not completely controlled by the customer.
If the customer is in control, as she is in most retail situations, a Recency-based approach is probably better. Recency looks at time *since last purchase* rather than time *between purchases*. Do the same thing as with Latency above. When you drop your promotion, look at response by 30-day segments – last purchase <30 days ago, last purchase 31 – 60 days ago, last purchase 61 – 90 days ago, etc. You will see the customer LifeCycle right before your eyes. Then look at responders versus non- responders for each 30 day block. Are they similar? Different? Similar / different within a 30-day block? Similar / different when comparing between 30-day blocks?
You’re looking for patterns. When you start to see them, they form the basis of the next test, where you specifically target a known segment of people with specific characteristics who exhibit a known behavior. Then sub-segment, and so on. Two segments become four, four become eight, etc. You stop creating new segments when you can’t find a reason to create another one, there are no significant differences left to group people by. Again, the data itself tells you when you have reached the end of the segmentation possibilities.
Of course, with 5 million actives, that could take a LifeTime! The bigger question of course is this: how do you increase the number of 12 month buyers? The answer is slow down the defection rate by looking for it early, recognizing when it is beginning, and attacking it specifically. Most of the people “defecting at 12 months” really defected a long, long time before that, you just are not measuring it!
Hope that helps!
Download the first 9 chapters of the Drilling Down book: PDF