Learn Methods, Metrics
(site map)

[ Home ]    [ FAQ ]    [ Download ]    [ Contact / About / Privacy ]

Get the Drilling Down Book!
http://www.booklocker.com/jimnovo

Prior Newsletters:
http://www.drilling-down.com/newsletters.htm
-------------------------------

In This Issue:

# Topics Overview

# Best Customer Retention Articles

# Event Driven Models in B2B

# Dealing with Extended Latencies

Topics Overview

Hi again folks, Jim Novo here.

Man, February went by very quickly.  So I'm late with the newsletter, my apologies.  Some of the blame goes to the newly launched Web Analytics Association.  I'm Co-Chair of the Education Committee and, well, there's a ton of work to do!  From the site:

"With your help, we can develop and offer training and certification to advance the professionalism of our trade.  We can encourage institutions of higher learning to add web analytics to their curricula".  I'm supposed to do that.  Want to help?  Become a member and then please volunteer!

Speaking of web analytics, I did a seminar with Brent Hieggelke of WebTrends for the AMA (American Marketing Association) on web segmentation and visitor / customer retention.  It's free to check out, though they want some personal info from you.  See:

http://tinyurl.com/3kt6b

OK, let's do some Drillin'...

Best Customer Retention Articles
====================

Big Ideas 2005
February 7, 2005  Target Marketing
Wisdom from the "gurus" on what will be hot this year in marketing, as direct and database marketing continue to bleed into and influence all marketing strategies and programs.  ROI based on customer value, the fall of mass marketing, the rise of analytics, and more.

Bring 'Em Back
February 21, 2004  Target Marketing
A nice primer on one of my favorite topics - defining a customer defection.  You can't increase profits by retaining customers if you don't define what a defection is first!  The article also supports a frequent suggestion of mine - quit paralyzing yourself by trying to nail it down 100%.  Put a stake in the ground, start somewhere.  Then it's a matter of using the right metrics, testing, and timing.

-------------------
If you are in SEO and the client isn't converting the additional visitors you generate, you can help them make it happen - click here.
-------------------

Questions from Fellow Drillers
=====================

Event-Driven Models in B2B

Q:  My company is really interested in pre-qualifying leads driven through the internet channel based on perceived interest  e.g. downloading a particular white paper, or returning, responding to an offer, what have you.  We haven't implemented this, but this scoring mechanism is sort of happening behind the scenes and is collected in a database for future use.  I know about your RFM and I think it's very appropriate here based on behavioral activity, but my sense is that I could also use significant events to help pre-qualify you.

How feasible is this sort of additional event scoring and what do you think you'd discover at the end of the day when you spoke to your marketer who ultimately tried to contact these leads based on this scoring?  i.e, guy comes into the car dealership, but the dealer already knows he can pre-qualify you for a $40,000 car.  Smart harvest marketing or just a waste of IT resources?

A:  This kind of thing is done all the time in B2B lead gen, and many of the "lead management" software packages have this capability built in, more or less.  They're usually not very smart.

RFM is really not going to be the ticket here, because you have too many discrete events, and the definition of "Frequency" and "Monetary" in this environment is elusive.  In other words, the RFM definition of Frequency is repeated occurrences of the same event, not a series of different events.  So in B2B lead gen, RFM is usually not optimal.  However, Recency and first cousin Latency are still very valid, and in the beginning, the early discovery,  Latency can help you organize the data.

The easiest way to start with something like this is similar to the way Pay-Per-Click is managed - you look at the end result and track back to the beginning, for example, I am buying all these search terms.  At the end of the month I sort by sales volume and can see which phrases generated the highest sales, and further, which ended making / losing money.

So what you do is focus on the "end game", whatever that is, and compare success to failure.  When comparing, what was the source of the lead / what actions were taken / in what order were they taken / what was the timing of these actions for successful outcomes?  And for failures, the same?  Compare and contrast success and failure, and you are on your way to building your own model.  This real world info provides you a "base" from which you can set up "tests" that can prove / disprove what effects source of the lead / what actions were taken / in what order were they taken / what was the timing of these actions have on outcome.  The trick to all this is tracking all the events to the lead so you have historical info.

Further, if you can collect the data, and there are a lot of significant behavioral events (a page view probably is not, a newsletter sign-up probably is), and you think the data is "clean", you can use machine intelligence to help run some of this data.  Search for the models CART and CHAID for more info.  You get into some tricky business here, and you can NEVER let the machine over-rule common sense, but these models can be helpful in the "sifting" process as you look for predictive variables.  Free / cheap software is available to run these so-called "Decision Tree" models, see here.  At the higher end, check to see if there is a copy of SAS or SPSS around the company, they are usually equipped with these models or "suite" add-ons are $3000 - $5000.

Jim

-------------------------------

Dealing with Extended Latencies
=====================

Q:  Is latency, as a metric, out of the question when the spread of the number of days in a latency period is so wide that to average them out and call the resultant figure "Acceptable days to date of predicted purchase" would seem meaningless?  I am thinking about the disparity in latency between customers who are Heavy, Moderate and Low users.

A:  I'm not sure I have enough context to understand the question (what are you trying to accomplish by using the metric?) but Latency is what it is.  In other words, you take your clue from the existing behavior itself.  If the average Latency for a certain segment is 2 years, well, it is, and that's not too long or too short, it just is.  Whether you can act on that information is another story; it depends on what you are trying to accomplish.

For example, average Latency on major home appliances, depending on brand, is anywhere from 5 to 10 years.  Is that too long of a "spread" to make the metric useful?  No.  It just is what it is, and you deal with it.

Now, it could be what you are really getting at has more to do with failing to identify the defection.  In other words, you are trying to average customer Latencies that are "open-ended" or infinite and you're coming up with a useless number.  If that's what you mean by "latency period is so wide that to average them out would be meaningless" it sounds to me like you need to call the defection so there is some endpoint you can use to measure with.

For example, let's say you have 10 years of data and it includes people who bought only one time 10 years ago.  Clearly, those people are no longer customers.  To include them in a Latency analysis would create a meaningless number, both arithmetically and behaviorally.  You have to put a stake in the ground and declare "no activity for over 2 years, they are a defected customer and will not be included in the analysis".  This takes the "infinite Latency" problem out of your hands.

Another approach (and I'm just guessing you are talking about retail) is to look at 2nd purchase Latency - average number of days between 1st and 2nd purchase.  Let's say it's 90 days.  So from now on, any new buyer who doesn't make a 2nd purchase within 90 days of the first is considered a defected customer and excluded from any analysis, because literally, they are likely to have infinite Latency and the "spread" is meaningless.

A third approach taken by people who have been in this kind of business a long time and trust these metrics is to set a threshold for being a customer at all.  At HSN (after many years of testing), we came to the point where we didn't even consider you a "customer" until you made a 2nd purchase.  We kept track of the number of 1x buyers, but that's about it.  We discovered there was no profitable way to create a "relationship" with these buyers, so they were not considered customers from a marketing perspective.

Q:  In this case (Latency period is so wide), is Recency the way to go?

A:  Well, again, I lack enough context to answer that question specifically because I don't know what you are trying to accomplish and with what kind of business.  But in general, Recency is a better bet when the behavior is not predictable, and Latency is a better bet when the behavior tends to be cyclical or repeat at regular intervals.  So for general retailing, Recency is a better predictor of likelihood to purchase.  If you are looking at the oil change business, Latency would be a better predictor of likelihood to get another oil change at a certain point in time.

Latency and Recency can be used at the same time on different segments, it's not an "either or" situation.  For example, you could use Latency for Heavy and Moderate, Recency for Low, if that makes sense in your situation.

Let's say that in your case "Low" means there really isn't enough transactional activity to determine Latency.  So you use Recency for that segment.   This approach is often used in remote retail, I call it a "one and done".  After first purchase, you allow Recency to expand, so that multi-buyers "tip their hand" and you don't end up making offers you did not have to make to get the 2nd and 3rd purchase.   At some point (typically 45 - 90 days out), the majority of multi-buyers have shown themselves, and you make one offer to the remaining 1-time buyers.  Those that respond become multi's, the majority of the rest will never buy a second time and it's a waste of money to promote to them - this is what I mean by "one and done".

In other words, you give them some time to prove whether they are customers or not, then you provide them one last shot by giving them a nudge.  Anybody who does not respond is now a defected customer and that's it.  If they come back and prove themselves to be a customer, fine, we'll include them in all the customer marketing.  But if they don't, it's  relationship game over.

Q:  My purchase of your book has generated a second purchase!  By a colleague of mine.

A:  Thanks for that!  If you'd like to elaborate on the specifics of what you are trying to accomplish, perhaps I can be more helpful!

Jim
-------------------------------

Booklocker.com     Amazon.com     Barnes & Noble.com

Find Out Specifically What is in the Book

Learn Customer Marketing Models and Metrics (site article list)