Sonamine Website

  • Increase font size
  • Default font size
  • Decrease font size
E-mail Print PDF
AddThis Social Bookmark Button

Network analysis increases conversion rates by 340%

Word of mouth (WOM) marketing is all the rage, but few commercial firms providing WOM products or advice can point to any robust proof that it is better than regular marketing.  This study analyzes how different segments used for a direct marketing campaign performed.  It has a few unique aspects that help come close to supporting the effectiveness of WOM.

1 – In selecting target segments for a direct marketing campaign, the researchers created a group of prospects based on whether they had a recent telephone conversation with an existing customer of the product.  This “recent conversation” property is the best approximation to “word-of-mouth” sharing that I have come across in the literature.  It is called “network neighbors” in the study.  Using call record data in this way allowed the researchers to study whether direct communication between prospects and customers inflenced the purchase decision.  No mining of the conversation contents took place since these were mobile calls, but imagine what the email providers can do with their data!

2 – Intuitively, marketers believe that birds of the same feathers flock together.  Therefore it is hard to know if the network connections are just masking or proxies for demographic and soci0metric similarities.  In this study, Hill and team found a way to separate out the similarites from the network neighbor effect.

One finding is that the take rate of the network neighbors was 3.4 times higher.  3.4 times higher is absolutely unbelievable as a commercial outcome.  

Is network based targeting and marketing any better than current forms of marketing?

First a little bit more detail from the study.  The marketing team had created 21 different segments for the direct marketing campaign.  These were made up of good prospects with the right demographics and customer relationship.   Hill and team created another segment, 22.  Segment 22 were made up of network neighbors that were excluded by the first 21 segments.  In other words, prospects in 22 would not have made the initial marketing list cut, they would have been excluded.  Segment 22 made up about 1.2% of the final marketing list, and had many network neighbors.  Note that there were network neighbors in the segments 1-21 as well.

Looking at the results of segment 22 then provides a good comparison between network marketing and traditional marketing. 

1 – the network neighbors in segment 22 had take rates of 0.83%; the network neighbors in segments 1-21 had take rates of 1.35%  So it would seem that the additional factors (loyalty, demographics etc) used in segments 1-21 improved conversions.

2 – however, the non network neighbors of segments 1-21 had a take rate of only 0.28%  This means that the network neighbors of segment 22, who had “bad” prediction factors such as loyalty and demographics, actually responded 3 times more than the non-network neighbors with higher loyalty and better demographics.  

When creating targeted prospect lists,  direct marketers are often leaving out a significant source of new customers.  Using network connections allows marketers to find new high-potential leads from non-traditional pools.

Did WOM reach customers who were not in the campaign list?

If word of mouth occurred, it would have increased the product adoption without the need for any direct marketing campaign.  In other words, the existing customers would have spread the products to their friends.  in this study, Hill and team were able to observe the take rate of network neighbors that did not receive the direct marketing.  This is another unusual aspect of this study.

The take rate of network neighbors not marketed to was 0.11%.  Recall that the take rate of the “good” prospects who were not network neighbors was 0.28%, while the take rate of network neighbors in segment 22 (aka not so good marketing prospects) was 0.83%

“Although they were not even marketed to, their take rate is almost half that of the non-NN targets — chosen as some of the best prospects by the marketing team” (italics original)

So the next logical question – what is the baseline take rate of the non-NN prospects who did not receive any direct marketing?  Despite difficulties getting actual numbers for the denominator (the population of non-NN that was not marketed to), Hill and team estimated that the take rate of this group to be well under 0.01%

In this telecommunications example, word of mouth is happening, is increasing the baseline product adoption by almost 10 times.  This allows telecom companies to put a value on customer satisfaction.  Unsatisfied customers who would not recommend their products to their friends cost the company some of this organic word-of-mouth customer acquisition.

Will network based marketing replace traditional segmentation?

So can marketing based on communications between your target prospects actually replace current slice-and-dice methods using age, gender, income, geographical location etc?  Hill and team decided to study this further by studying the network graph made up of existing customers and the people they spoke with.  Other graph attributes such as ” number of linked prospects”, “connected to influencer”, “overlap of graph with an existing customer’s graph”  were added to the predictive models, in addition to the standard user demographics and customer relationship variables such as loyalty. 

There were 3 predictive models studied.  All network attributes only, traditional attributes only (loyalty, demographic and geographic) and a combination.  The metric studied was something called the Wilcoxon-Mann-Whitney statistic, essentially an area-under-the-curve (AUC) measurement.  An AUC = 1.0 means that every time a randomly chosen taker is picked, there is 100% probability that she will be ranked higher than a randomly chosen non-taker.  So AUC=1.0 means your model is perfect.  AUC=0.5  means that your model is practically useless. 

The all network model had an AUC of 0.71; the all traditional model had AUC of 0.66  The combination model had AUC of 0.71  Two interesting points to note.  First, the network attributes did a better job than traditional marketing variables.  This may be a little counter-intuitive.  Every marketer has been taught, in school and by their peers, that an individual’s characteristics determine their segmentation and marketing effectiveness.  But this finding shows that there is an alternative way to determining the best prospects without using any sensitive personal information about the consumer, such as age, gender and income.  

Secondly,  in the combined model, adding the traditional variables into the network attributes did not improve the model one iota.  ”The similarities represented implicitly or explicitly in the network attributes seem to account for all useful information captured by traditional demographics and other marketing attributes.”

Much of the consumer marketing industry today is founded on learning about the individual consumer.  Many firms spend a lot of money to buy and augment their consumer data.  The hope is that better demographics and geographic information can lead to better marketing success.  Imagine if all we need is communication and social network data instead of this demographic data.

One lingering question, you need existing customers before you can utilize network-graph based marketing methods.  How do you get this “seed list” of customers?

Return-On-Investment (ROI) from using network based marketing

Increasing conversion rates by 340% Part 5 – ROI
May 5, 2009 by sonamine
This post is really about why a marketer would use the network based methodology over traditional methods, in addition to having 3.5x conversion rates.  Really the title should read, saving direct marketing costs while achieving target conversions. After all, it boils down to $.

First, one more graphical finding from the study.  The graph below shows the lift curve when using traditional attributes of loyalty, demographics and geography, and traditional attributes WITH network attributes.   The x-axis plots the % of the selected mailing generated by the 2 models.  The graph plots the amount of cumulative sales that would have resulted if you mailed out a certain percentage of the prospect list.  As you can see, the trad+network model performed better.   Unfortunately, the authors did not provide a lift curve for the pure network attributes model.  But, we have an approximation we can use, bearing in mind that trad+network model performed no better than a pure network attribute model, we can essentially consider the lift curve of the trad+network as the same as the pure network curve.



We can use these lift curves to do a simple ROI calculation.  To facilitate this, I have added 2 lines to the curve above.

Assuming we want to achieve an 80% cumulative value of sales, we can see how much savings we can make on the mailings using the network model.  The savings are in 2 forms (a) the cost of the direct mailings.  So in this case, we would save on about 15% of the mailings.  You can see this by following the horizontal line that starts at 80% on the y-axis, and finding where the line intersects the lift curves.  Taking the difference gives you the savings.

The number of mailings was not provided due to confidentiality reasons, but let’s assume it was 1M.  The savings here would amount to 150,000 individual pieces.  At $1 a piece, this would amount to savings of $150,000.  Not bad.  (b) There might also be some cost savings from not having to purchase augmented data from suppliers such as Acxiom.  This cost is difficult to quantify so we’ll leave it out of the calculations. 

Assuming take rates of about 0.5% (the actual number was not reported), that means 80% cumulative sales represented 4000 purchases out of 1M mailings.  Assuming a simple product profit of $150 over a year, the total profit due to this direct mailing  =  $600,000. 

The ROI using traditional attributes = 600,000 – 700,000 = -100,000 = negative 14%

The ROI using traditional + network attributes = 600,000 – 550,000 = 50,000 = 9%

Products that are suitable for network marketing

In the realm of WOM, do the products themselves influence effectiveness of viral spreading?  Hill and team went one step further and tried to compare the network effects of 2 products.  The product that they marketed was a “new technology”; the second product released by the same mobile communications firm was a new pricing plan.  The goal is to compare how these 2 products did along WOM measures.

The measurement of comparison was something called “network neigborness.”  Over a period of 8 months within one year of the respective product launches, Hill and team measured the % of new customers that had communicated previously with a user of the product in question.  These percentages were taken for both the new technology and the new pricing plan.  The results are shown in the chart below.



For the pricing plan, new network-based customers averaged a constant 3%.  For the new technology, that % was rising.  The dip shown in month 5 corresponded to the mass marketing campaign of the main study, leading to many new non-network-neighbor customers.   One possible explanation for the difference in network effect is that a new pricing plan is but one piece of information in a vast number of pricing options.  As a result, the “new” pricing plan might not stand out in people’s memory.  After all, humans are limited by the cognitive biases of availability heuristic and contrast effects.  Additionally, talking about a new product provides the socially desirable attribute of “coolness”. 

Product and brand marketers looking to leverage word of mouth and buzz marketing, with companies like P&G Tremor and BzzAgent, should consider whether their product has the proper characteristics for newsworthiness.


Reference
Network-based marketing: Identifying likely adopters via consumer networks.  Hill, S., Provost, F., Volinsky, C.  Statistical Science, 2006, Vol 21, No.2 256-276.

 

User Login

               No account yet?