The following is a copy of the case study I presented with Tom Duncan of Positec USA at the Great Ideas Summit in NOLA earlier this month.  I hope you find it useful and I look forward to your comments.  Thanks again to Tom for funding and sharing his experience.

Predict the Impact of TV on Multichannel Sales from Tyson Roberts on Vimeo.

Tyson Roberts, our CEO, will be presenting with Tom Duncan from Positec USA, Inc. next Wednesday, March 2nd at the Great Ideas Summit in New Orleans. They will be presenting a case study that shows how we accurately predicted the impact of TV media on cross-channel sales during Q4 last year. Tyson and Tim will share a lot of great insights and some amazing results.

You can learn more about the event at http://www.eragreatideas.org/ and you can read a brief synopsis of the session with Tom and Tyson at http://blog.lucidcommerce.com/events/.

If you cannot attend the event but would like to receive a document form of the case study contact our sales team at sales@proceedmediagroup.com.

“This may seem simple, but you need to give customers what they want, not what you think they want. And, if you do this, people will keep coming back.” -John Ilhan

It always surprises me when companies have data that isn’t being used. If a company has been in business for any length of time, there are usually some customers who have already bought the product or service. Often ignored, this list of customer records can be the most valuable asset the company owns.

One of the first things that marketers are taught at marketing school is that they should know their customers. By examining your customers’ demographics, suddenly a whole world is opened up with information about who they are, where they come from.

Let’s have a look at some real-world results:

Below are six example products and what their age profiles look like. The Educational product appeals to very young age groups, where-as the Gripping bath mat appeals to older ages. In-between products find themselves favored by certain age brackets. Clearly different demographics buy different products, and this allows us to use demographics for targeting.

Figure 1: Products from top to bottom: (a) Educational program, (b) Hair care product, (c) Exercise product, (d) Handyman product, (e) Leaf blower, (f) Gripping mat for bath. The y-axis values are the percent higher or lower than the norm. For example, if 3% of product buyers were aged 70, and the norm for 70 year olds was 2%, then we would show this as a 100 * (3/2-1) = 50% higher than the norm.

Handyman product

Let’s now have a look at one product in more detail. Below are a set of 8 variables and their demographic readings compared to the population norm. As you’re looking through the graphs below, see if you can look at each graph and come up with some theories on who these customers are? (Y-axis on each is the same, so the bigger the bar, the more important or unsuual it is)

Figure 2: 8 Demographic variables and category readings for the Handyman Tool product. In all graphs the distributions are the percent higher or lower than the norm.

How did you go? Do you now have any thoughts about who these customers? By my reading the above scores indicate that the handyman product buyers tend to own trucks, are male, Republican, age 46-64, have higher incomes, are higher educated, lower home values (unexpected), and so on.

These graphs are fun to look at. However what about the other hundreds of demographic variables?

Yikes that would be a lot of graphs to read through!

In order to maximize understandability for a human analyst and capture only the most important information, we can select the highest positive disparity variable-values for each variable (eg. For Age, it might be 65), and avoid showing the other variable-values (eg. age 64, 60, etc) so that only the highest lift value from each major variable class is depicted.  The “most distinctive variable-values” report for the handyman product is below:

VariableValue Rank Income-EstimatedHousehold=$100,000-$124,999 1 Vehicle-DominantLifestyleIndicator=Truck Classification 2 Gender-InputIndividual=Male 3 Vehicle-KnownOwnedNumber=3 or More Cars 4 PoliticalParty-InputIndividual=Voter – Republican 5 Education-InputIndividual=Completed Graduate School 6 NumberofChildren-100%=No Children 7 Off-roadRecreationalVehicles=TRUE 8 Vehicle-MotorcycleOwner=TRUE 9 Vehicle-TruckOwner=TRUE 10 Vehicle-RVOwner=TRUE 11 AfricanAmericanProfessionals=TRUE 12 Hunting=TRUE 13 DIYLiving=TRUE 14 MilitaryMemorabilia/Weaponry=TRUE 15 HighEndAppliances-C=TRUE 16 Apparel-Men’s-BigandTall-C=TRUE 17 PhotographyandVideoEquipment-C=TRUE 18 Automotive,AutoPartsandAccessories-SC=TRUE 19 Woodworking=TRUE 20

Figure 3: Handyman tool product highest positive discrepancy variables ranked in order of highest lift to lowest, only one value from each variable class being shown.

The above profiles contain many useful insights and finally we may feel ready to start targeting them. However there is more yet that we can glean from our innocuous customer records.

Clusters

In the education variable there’s a curious “double peak” with higher than normal people with only High school education, and also higher than normal with Graduate degrees… It is possible, but it is suspicious. What could be the explanation?  Figure 4 shows another good example of a customer population with a double peak.

Here is where we can start digging deeper. Until now we have treated the customer population as a single monolithic group and targeted essentially the average. However this may not actually be the case. What if your customer base was organized into different clusters of customers? In fact, by targeting the average, it is possible that you may be targeting no-one at all (see Figure 5 left).

 

Figure 4: (Top) Age distribution for acne product showing a “double peak”. (Bottom) After running cluster analysis, we discover that there are well-defined clusters of young and middle aged adults which are making up the population, have different interests and attributes, and are causing the double-peak.

In order to target more precisely, we can use Statistical Clustering to automatically search out and identify clusters within the group of customer records. A variety of algorithms for clustering are available, but K-means is the simplest and so here’s a quick description – the algorithm breaks the population into K clusters or centroids (which are literally vectors just like the customer vectors), calculates membership of each cluster by finding the closest cluster to each customer, and then re-centers the cluster based on the mean of the members. The process repeats iteratively. Mathematically the algorithm is hill climbing in the negative direction of the derivative of an energy function defined by the sum of squares around each centroid, and so is trying to minimize energy or distances by getting its K cluster centers situated into dense regions so that spread from each cluster to customer is kept to a minimum.

Clustering the handyman product results in three interesting centroids – older, male, near-retirement or retired, republican-leaning, people who are highly educated, as well as younger, blue-collar, male, contractor types who are working and have not completed a college degree. There is also a tiny cluster of women. We can now calculate the value of each segment and micro-target these segemnts using TV media.

Figure 5: Examples of clustering on an artificial data set (normally distributed random noise). In the first example, the mean is being used to target four underlying clusters. One could argue that the mean is not actually targeting any of the profiles – it is actually targeting a sparse area with few customers! This shows how targeting the average customer can go awry. In the second example, one of the clusters is well positioned, but the other is a compromise target of three clusters. In the final example, 4 centroids are cleanly covering each of the 4 clusters. This corresponds to a campaign in which 4 micro-targets are being used to target each of the sub-populations.

Figure 6: Cartoon is called “know your customer” from Mike Bannon and Mordantorange.com.

Sometimes the insights from clustering can be just as important as the value of the technique in enabling micro-targeting. One company that we analyzed were selling an acne treatment and had traditionally marketed to teenagers. However the company was puzzled by some recent aggregated statistics that they were receiving from retailers that indicated that their product was being bought by mature-age people (which was essentially the average profile). By analyzing the population, we found that both teenagers, and middle-aged adults, formed distinct, compact, buyer clusters, with a different suite of characteristics; age (young vs middle aged), income (low vs high), occupation (student vs professional). The adults comprised both wealthier women who were buying for their children, as well as perhaps women who were buying as a treatment for menopause, and were actually a bigger cluster. The young people seemed to be buying the product as a treatment for themselves. Different messaging and TV programming needed to be employed to target both populations.

If you have some data laying around in spreadsheets, try analyzing it and finding out more about your customers. You may be surprised with what you find.

Admittedly this post is a bit of a sales pitch. But I think the experience described below serves as a good reminder for marketers that differentiation is often about results, not necessarily about how those results are achieved.

In my career I’ve always felt I need to differentiate myself. In marketing collateral, companies need to differentiate themselves and/or their offers. In sales, pitches need to clearly articulate what is different about what is being offered. If you are trying to convince someone that you have a better mouse trap you have to explain what is different about your mouse trap.

Today we worked through an exercise internally to understand how we best communicate what is different about our solution. Similar to some other TV media agencies and some data-oriented consulting firms (we compete against both) we offer advanced targeting, measurement, and optimization of TV media. And while the high-level pieces of what we do (targeting, measurement, optimization) sounds like what other companies offer, the results we are generating are very different.

We reviewed the results we were producing for our clients vs. the results other agencies were generating for those same clients prior to us being hired. Like most agencies we are hired to improve the results generated by TV media investments. We found a long list of staggering results that show how much our differences produce in terms of value for clients.  Some of the results from our list this morning include:

• Achieving 163% of our client’s retail sales objectives in 2010 (expectations were set based on other agency’s results)
• Increasing retail and online sales YOY 25% with only an 8% increase in budget
• Reducing cost per customer 72% in just 6 weeks
• Lowering cost per lead by more than 50% over the course of a 5 week campaign
• Etc.

While I’m happy to have context to share how great our results are, the results are not the point here… Or are they? Initially, these results were just proof points that we are clearly doing things differently. But what are we doing differently?

Well, there are a number of things we do differently. For example, we use new proprietary data and patent-pending algorithms to target media. We use patent-pending attribution algorithms and intelligent tests designed to enable statistical rigor in measuring how TV media affects retail and online sales. We optimize media based on unique exploration tests we run, the creativity of our team, and a maniacal focus on results that is pervasive in our culture.

The problem with the above list of differences is that people don’t really understand them. How do these propriety algorithms work and are they really better? What data do you have? Is it better than the data other agencies have? How do I know this mouse trap is really better? Marketers are constantly inundated with how agencies are different, but don’t know how to tell if the difference is better for them.

Well, after thinking about how we are different we decided the best way to communicate our difference to just communicate how different our results are. After all, that’s what the marketers care about. Who cares about the proprietary algorithm if the results aren’t measurably better?

So, our difference is our results. How we deliver the results is interesting- to us. And sure some sophisticated marketers care about our unique methodology. But in the end, it’s about results. And ours are different.

What do you mean by targeting?

My recent conversations with marketers have made it clear that the word targeting has been overloaded.  Depending on who you are speaking with, unintended meanings are likely being ascribed.  For example, if you mention targeting to a marketer who spends most of their time in addressable mediums such as online or direct mail, they’ll assume that you mean:

• Your advertisements are being seen only by consumers who meet your targeting criteria
• All other consumers are actively suppressed
• You’re paying a premium to your media partner(s) for this targeting service or technology

 In contrast, when broadcast marketers from TV or Radio say targeting, they don’t typically mean any of the above.  For this reason, I propose the following definition of targeting in hopes that it may be broad enough to generalize to all marketers:

Targeting is the multi-step process used by marketers to maximize the percentage of their advertising impressions that fall on qualified prospective customers.    The 4 fundamental steps of targeting.

 While there are many ways to target, the typical process contains the following steps:

1. Establish Your Target(s) – There are many different methods of deriving or constructing a target, but the fundamental aim is to produce a high-resolution profile of your most attractive customers, paying special attention to those attributes where they are meaningfully different from the population norm.  If your brand is already in market, the best method is often to derive your target from a profile of your most desirable existing customers.  For a new brand, targets cannot be derived and must be constructed.  This is often done using third-party, syndicated panel data, or from the marketer’s intuition, presumably informed by experience selling similar products.  The finished product should be a target customer profile that contains the demographic, behavioral, attitudinal and other attributes that are a) most strongly correlated with customer value, and b) most divergent from population norm.  In many cases, this process produces not a single profile, but multiple distinct populations of customers (clusters).  Clusters are outside the scope of this post, but not to be ignored.
2. Compute Your Allowable – The next step is understanding what it is worth to your brand to place an impression on a consumer who matches your target customer profile.  This is typically backed into.  First, calculate the average worth of a new customer, then compute an acceptable cost-per-acquisition.  Using your historical response rate, calculate the number of impressions delivered within target would produce a customer.  You now can compute an acceptable target cost-per-thousand impressions.  This is what you should be willing to pay to reach a thousand consumers who match your target customer profile.
3. Media Selection – Here, you’re tasked with evaluating the myriad media placement opportunities available, your aim being to maximize the similarity between the target customer profile you constructed above and the audience accessible through each placement.  In order to do this your media partner must be able to describe the composition of their audience along the same attributes that you found to be important in our target customer profile.  In an addressable medium, like direct mail or online, it is possible to evaluate this similarity on an individual basis.  Not so on TV.  At least, not yet (with an exception for some very small special cases).  In a broadcast medium, you’re forced to evaluate the similarity of the whole audience, which is a composite group of consumers who will be tuned in during your commercial airing.  In this way you are limited by the level of understanding  that a station or network has and can communicate about their audience.
4. Setting Your Bid – Having established a target, ranked the media placements by similarity and computed your allowable, you now know who you’re after, where they are and what you’re willing to pay for them.  One last step.  You now must discount each of the audiences by the waste factor.  This is the portion of the audience that does not match your target customer profile.  So, if the allowable you computed above was $5 CPM and the audience you are about to bid on is comprised of 50% in target and 50% waste, your bid would be $2.50 CPM.

Not as brief as I had planned, but clarifying I hope. 

So, what’s the problem with targeting on TV?

Television viewing audiences are poorly understood today.  This understanding greatly limits your ability to assess the number and concentration of target consumers you’ll be reaching.  In turn, making it impossible to accurately value these placements.  The solution is to enhance the understanding of these audiences.  There are many ways to achieve this.  This paper details how we address the targeting challenges TV presents.  Please share any approaches you’ve employed to address the challenge of targeting on TV.

We went dark for a couple of weeks but now we’re back in the swing of things after a brief holiday hangover. We’ve had an incredibly hectic 2 weeks and we are picking up right where we left off.

As I was hustling to work this past Wednesday I noticed my phone was being updated with new software from Motorola/VZW/Google. I received a bunch of updates to the OS, updates to apps, and the Madden 2011 game. I’ve been waiting for updates to the OS for months and then one day presto!

It took me all of 10 seconds to figure out why I received this update on Wednesday. On Tuesday Verizon Wireless announced it would offer the iPhone (obviously an Apple product). On Wednesday Motorola/Google made sure I received the updated software.

This was a smart move by the companies on one level because I have the latest software to keep me excited about my phone. And a game as a bonus because they’re such great companies. On the flip side, it pisses me off because they did it out of fear of competition rather than love for me as a customer. I feel this way because when I bought my Droid X in July they claimed I would receive an update to the OS with Flash back in September. It wasn’t until the iPhone announcement that I finally got my update.

Everyone knows few hard things happen in business without a forcing function. This experience was another reminder of that fact.