Lade Inhalt...

Seeding - Can Marketers Take Advantage of Known Network Measures?

Studienarbeit 2012 63 Seiten

Informatik - Wirtschaftsinformatik


Table of Contents

List of Figures

List of Tables

List of Abbreviations

1 Introduction

2 Literature review
2.1 Viral Marketing
2.2 Social network analysis
2.2.1 Concept and basic characteristics
2.2.2 Centrality
2.3 Determinants of a Successful Viral Campaign
2.3.1 Content
2.3.2 Environment
2.3.3 Incentives
2.3.4 Seeding Strategy

3 Empirical analysis
3.1 Hypotheses
3.2 Experimental design
3.2.1 Test environment
3.2.2 Test content and tracking
3.2.3 Incentives used in experiment
3.2.4 Seeding procedure
3.2.5 Execution of experiment
3.3 Empirical results
3.3.1 Participation
3.3.2 Timing implications
3.3.3 Exploratory data analysis

4 Summary, Conclusion and Recommendations for Future Research



List of Figures

Figure 1: Traditional marketing vs. VM

Figure 2: Motivation, role and impact of the communicator

Figure 3: Kite network

Figure 4: Relationship of qualitative and quantitative measures

Figure 5: a) random network, b) scale-free network, c) small world network

Figure 6: a) SIR model, b) SIS model

Figure 7: SIRS model

Figure 8: Visualization of test network graph including degree centrality

Figure 9: Preparatory work

Figure 10: Frequency distribution of degree and betweenness centrality

Figure 11: Example of our test content for one experimental setting

Figure 12: Treatment design

Figure 13: Venn diagram

Figure 14: Gantt diagram of seeding procedure

Figure 15: Frequency distribution of received JPs (AED and TED)

Figure 16: Frequency distribution of forwarded JPs (AED and TED)

Figure 17: Responses for 1st, 2nd, 3rd and 4th wave of experimental settings (AED and TED)

Figure 18: Frequency distribution of response time after initial seeding (AED and TED)

List of Tables

Table 1: Calculation of centrality measures

Table 2: Gender and centrality measures

Table 3: Overall participation (AED and TED)

Table 4: Responses (R) and infected reponses (IR)

Table 5: Indiviual probabiliy to respond (mixed effects logit model, AED and TED)

Table 6: Expected no. of responses (linear regression model, AED)

Table 7: Expected no. of responses (linear regression model, TED)

List of Abbreviations

illustration not visible in this excerpt

1 Introduction

“Viral Marketing adds fuel to the fire“. (Jurvetson, Draper 1998, p. 8)

Marketing offers three core strategies – push, pull and viral strategies. In a push strategy, marketers invite addressees to communicate a message. A pull strategy is build on active content requests by consumers (Wiedemann 2007). Viral Marketing (VM) is marketing strategy that utilizes the word-of-mouth to spread information about a physical product, a service or an idea via a digital network (Arndt 1967; Jurvetson, Draper 1998; Helm 2000; Subramani, Rajagopalan 2003). Advantages are its higher cost efficiency (Dobele et al. 2005) compared to traditional advertizing through mass media and the fact of higher persuasiveness of receiving a message through the recipients very own social network peers. In addition there is a high speed of diffusion, a high degree of message integrity and potentially an easier way to track the results and quantify the success of a campaign afterwards (Bampo et al. 2008, p. 274). A VM strategy usually aims at strengthening awareness of a product, retrieving consumer information or increasing sales (Langner 2006, pp. 220f). A great example for spreading information via the web was the 30-minute video documentary ‘Kony 2012’ , that aimed at raising awareness of inhumanity and suppression by an African warlord in Uganda. This video reached more than 30 million users in 48 hours (Halliday 2012). Four critical factors are crucial for the success of such a viral campaign:

1. Attractiveness of the content (Porter, Golan 2006)
2. Environment (Bampo et al. 2008)
3. Recipients behavior and incentives for forwarding the message (Arndt 1967)
4. Seeding strategy (Kalish et al. 1995; Libai et al. 2005; Bampo et al. 2008)

The different seeding strategies are of particular interest in this thesis because the propagation of the message differs, depending on the initial set of targets, which the initiator of a campaign defines. Targets in the context of digital networks can either be so called hubs, fringes or bridges, depending on the respective centrality measures (Bavelas 1948; Freeman 1977). Hubs represent individuals within a network that are well-connected, whereas fringes represent individuals with few connections. Bridges are individuals that connect different parts of networks that would not be connected otherwise (Bampo et al. 2008, p. 277). The big issue marketers have, is little knowledge about which parts of networks to seed first to provide an ideal propagation of a message. In literature, opinions dissent amongst authors. Most researchers recommend targeting well connected individuals and opinion leaders i.e., hubs, as this strategy supposedly increases rapid diffusion of the network (Coleman et al. 1966; Becker 1970; Kemper 1980; Anderson, May 1991; Iyengar et al. 2011). In contrast, a few authors are of the opinion, that seeding hubs is not an optimal strategy, due to the large amount of information hubs have to work through and filter (Simmel 1950; Sundararajan 2006; Porter, Donthu 2008; Galeotti, Goyal 2009). Furthermore, Watts, Dodds (2007) state that “large cascades of influence are driven not by influentials but by a critical mass of easily influenced individuals” i.e., hubs are not as crucial for the diffusion of a message inside a network as easily influenced recipients, who are very susceptible to adoption (Watts, Dodds 2007, p. 441). In addition, Van den Bulte, Lilien (2001) find that social contagion appears not to be influential for the diffusion of a network. Social contagion describes the behaviour, that individuals make their decisions based on the opinion or knowledge of their peers within their own private or business network. The ongoing debate in literature generates a need for a holistic empirical analysis on the role of hubs, brigdes, fringes and also their respective peers as initial targets in different seeding strategies (Hinz et al. 2011, pp. 33f; Bampo et al. 2008, p. 281). The goal of this thesis is to provide profound knowledge on ideal seeding strategies, given the structure of the underlying network.

The following structure describes the remainder of this thesis: We want to provide a short overview on the existing literature and the theoretical framework associated with the field of viral marketing. Hereafter, we provide basic knowledge to conduct a social network analysis and define centrality measures that will be of essential interest in our experiment. Then, we will focus on determinants for a successful viral campaign and especially recommendations for seeding strategies. The main part gives an overview on the preparatory work prior to the experiment and introduces the methodology of the latter. We use a test environment of 120 individuals for our experiment. Our test set-up runs different seeding strategies in parallel testing hubs, fringes, bridges and random sets. In addition, we also examine differences between using the entire network or only the neighbors of the biggest hubs as potential initial targets. A detailed analysis of the outcome of this empirical study follows. Last, we provide a summary of the major findings of this analysis and conclude with shortcomings of this evaluation and recommendations for future research.

2 Literature review

The following part introduces the origin, concept and goals of VM as well as role, motivation and impact of participants that act within a campaign. Hereafter, we define network measurements that are key elements in our analysis of the underlying social network structure. Next, we examine crucial factors for a successful viral campaign.

2.1 Viral Marketing

Our well connected society offers a great opportunity to spread marketing information through viral campaigns with tremendous speed (Godin 2001, p. 15). VM is a modern variation of classic word-of mouth marketing and is often referred to in literature as ‘word-of-web’, ‘word-of-mouse’, ‘buzz marketing’, ‘customer-to-customer’ or ‘peer-to-peer’ (Bampo et al. 2008, p. 273). It was first introduced in 1996 by venture capitalist Steve Jurvetson, who described the concept of Hotmail, a free email service. Hotmail attached a recommendation to sign up for their email service to every email users sent out. Users implicitly recommended the service through this notice at the bottom of their emails (Kaikati, Kaikati 2004, p. 9).

Typically, a viral campaign creates an attractive message, which invites recipients to forward the message voluntarily to their peers. Ideally, recipients of the latter have again a stimulus to forward the message to their respective peers (Phelps et al. 2004, p. 335). Initiators of viral campaigns take advantage of already existing private or business networks to spread their

illustration not visible in this excerpt

Figure 1: Traditional marketing vs. VM (Source: own illustration referring to Godin 2001, pp. 18f)

message (Galeotti, Goyal 2009, p. 509). In contrast to traditional marketing campaigns that focus on sending large amounts of messages directly to consumers, VM campaigns only infect the seeds that act as intermediaries to forward the message to their peers (Godin 2001, pp.

18f). If the infection of the network is successful, the message will diffuse the network like a virus from one consumer to the other (Kaikati, Kaikati 2004, p. 7). Figure 1 illustrates the different procedures of traditional marketing and VM.

The motivation of participants, who act as communicators i.e., participants that forward the message, can be extrinsic or intrinsic. Extrinsic motivation is based on giving the communicator an edge, whereas intrinsically motivated communicators forward the message, because they want to feel competently and increase their self-esteem by steering their peers (Deci 1975). Participants can have two roles of communication in viral campaigns – passive or active. One role is that consumers implicitly forward the message to their peers by using a product i.e., they act passively. A second option for consumers is to send a message and thereby actively participate in the viral process (Subramani, Rajagopalan 2003; Thomas, JR 2004). Viral campaigns can have a positive or negative impact on consumers. Especially messages with negative impact appear to diffuse the network faster than messages with positive impact (Swan, Oliver 1989). Negative impact in that context is dangerous for advertisers, as only few unsatisfied consumers can influence the fail or success of a product immensely (Solomon et al. 1999). Figure 2 illustrates these three dimensions: motivation, role and impact of the communicators.

illustration not visible in this excerpt

Figure 2: Motivation, role and impact of the communicator (Source: own illustration)

Given this information a viral campaign can have three goals (Langner 2006, pp. 220f):

- Strengthen brand awareness
- Retrieve consumer information
- Increase sales

Strengthening brand awareness appears to be the most common goal in VM. Consumers see a funny video clip or play an ad game online and unconsciously promote the brand, while busying themselves with or being entertained by the content. The second most common goal within the scope of a viral campaign is to retrieve consumer information. Relevant information, such as name, e-mail address or phone no. are usually retrieved through a hurdle in the referral process. Consumers have to leave this information to be able to use the content. Furthermore, a goal can also be to simply increase sales. One way is to provide a free service and in addition to provide premium fee required services e.g., free e-mail service with a premium spam filter or additional space. Another way is to provide the first version of a product free of charge, while later, probably more sophisticated, variations of the product are not free of charge anymore (Langner 2006, pp. 220f).

Once injected into a network, a marketing virus is hard to control. Due to this, the initiator of a campaign has to carefully select his/her target group prior to releasing the virus into the system. Especially the initial set of targets is important because consumers will most likely only forward the message to their respective peers, who might be interested in the content. If the initiator chooses the wrong initial set of target, the virus might not spread adequately and therefore will not reach the targeted group of consumers. Nevertheless, this referral process is very efficient as communicators of a message will only select peers that have a sufficient interest in receiving the message (Langner 2006, pp. 221f).

2.2 Social network analysis

Since individuals interact with each other in online networks, these ‘social networks’ have become an engrossing mainstream field of research (Zhou et al. 2006). The term social network is inseparably tied to the concept of influence. Granovetter's publications "The Strength of Weak Ties" (Granovetter 1973) and "Threshold Models of Collective Behavior" (Granovetter 1978) were very likely the first to inflame public interest for social networks and the spread of information. The revolution happened in the 1990’s, when physicists began publishing on small world networks (Watts, Strogatz 1998) and a year later Barabási, Albert (1999) examined the distribution of degree centralities (Freeman 2011). In the context of the influence theory, Gladwell’s “The Tipping Point” (Gladwell 2002) was one of the recent publications being responsible for the public fascination with social networks and increasing interest for related phenomena. For the field of VM, social networks also play an essential role, as they provide an ideal infrastructure to implement word-of-mouth mechanisms (Urchs, Koerner 2008). As the social network analysis (SNA) provides a sound basis to analyze networks and develop further recommendations for marketing activities, we will provide the necessary insights in the following chapter. Furthermore this basic knowledge is key to understand the methodology of our empirical analysis.

2.2.1 Concept and basic characteristics

SNA is the formal study of systems of people and their relationships among each other. It evolved from a combination of mathematics and social science fields, including graph theory, psychology and anthropology (Freeman 2011). SNA enables researchers to break down how communication works between different individuals, groups and companies, both from a mathematical and from visual point of view (Garton et al. 1997).

A crucial element for the SNA is graph theory, which is a subarea of mathematics. A graph is a simple way to illustrate network structures. It usually consists of vertices, which are connected through edges. Vertices are also referred to as nodes or point, while edges are also referred to as lines or arcs. Vertices usually represent actors, who can be individuals, groups or companies. In that context, edges represent relationships or interactions among the vertices, which can be directed (asymmetric) or undirected (symmetric). A ‘weakly-connected’ graph has at least one connection to one other actor of the network (Chartrand 1985; Bondy, Murty 2008).

For our analysis, edges represent relationships, which we call ‘ties’. These edges are undirected i.e., actors know each other. The strength of these ties can vary between actors and is mainly driven by four factors, which are the “amount of time, the emotional intensity, the intimacy (mutual confiding), and the reciprocal services which characterize the tie”. These ties can generally be distinguished as being strong, weak or absent (Granovetter 1973, p. 1361). Strong ties exist between relatives, friends and neighbors, while weak apply to acquaintances and indirect ties e.g., friend's relatives or relative's friends (Lin et al. 1981, p. 397). Absent ties define the absence of any relationship or ties, which are insignificant (Granovetter 1973, p. 1361).

Various network measures have been developed for the SNA. Commonly applied network measures in marketing are density, clique, and centrality. Density is the most widespread measure of network coherence. It can be between 0 and 1, while measuring the extent to which all potential edges between vertices exist. Actors in dense networks supposedly encourage each other to cooperate and to collaborate (Webster, Morrison 2004), have high social support and solidarity among each other, but on the downside also are expected to act conform to group consensus (Burt 1998).

Detecting cliques (Luce, Perry 1949) or subgroups in networks has been of unbroken interest in marketing (Webster, Morrison 2004). Webster, Morrison (2004) define a clique as “a subset of actors who all have direct connections to one another and no additional network member can be added who also has direct connections to everyone in the subset” (Webster, Morrison 2004, p. 13). Usually, networks contain several of these sub networks, which are small in size, but nevertheless overlap one another (Brown, Reingen 1987).

The following chapter discusses centrality as third, and for our analysis most important, measure.

2.2.2 Centrality

In order to provide an understanding with regards to the position or status of an actor in a network, its ‘degree’ can be calculated. The degree is formally defined as the number of ties, edges, arcs, lines, connections, peers or whatever you want to call relationships (Freeman 2011). This should not be confused with the idea of ‘degrees of separation’ in social networks, which is formally called path distance or geodesic distance (Bakhshandeh et al. 2011). Appendix A illustrates an example of how to calculate the degree.

Moreover, various centrality measures e.g., ‘degree centrality’ (DC), ‘betweenness centrality’ (BC), ‘closeness centrality’ (CC), ‘eigenvector centrality’, ‘information centrality’, ‘flow betweenness’, the ‘rush index’, the ‘influence measures’ etc. have been developed to enable the analyst to specify the individual importance of each actor and make implicit assumptions about the manner in which information diffuses a network (Borgatti 2005). DC, BC and CC are of great importance for the understanding of our empirical work[1] and shall therefore be explained in more detail. All three discussed centrality measures can take on values between 0 and 1.

DC is measure, which is easily calculated using the degree and the total number of actors in the network. The centrality of an actor within the network increases with DC and reflects a high number of edges to other actors (Freeman 1977; Freeman 2011).

BC considers three actors, measuring the shortest path or geodesic distance between a pair of actors. Afterwards we evaluate, whether the respective actor is an intermediary between the other two. BC increases with number cases for which the role of an intermediary is true (Jansen 2003:134ff). Hence actors, who do not have a high degree, can still play very important roles as bridges linking otherwise not connected sub networks (Freeman 1977).

CC examines, which actors are close to other actors. It is defined as the inverse of farness, which in turn, is the sum of distances to all other nodes. CC can only be calculated for weakly connected graphs, as the distance between vertices of partially disconnected sub networks is infinite (Wasserman, Faust 1994).

Table 1 illustrates the calculation methodology for the centrality measures in focus. Software like Pajek[2] and NodeXL [3] enable researchers to easily perform analyses on network centrality measures. We also used the latter for our empirical analysis.

Table 1: Calculation of centrality measures (Source: referring to Jansen 2003, pp. 137f)

illustration not visible in this excerpt

Figure 3 illustrates a kite network, which was developed by David Krackhardt. We use this kite network to substantiate the latter measures. Obviously, actor A has the most direct connections and therefore has the highest DC. Actor A is a 'hub' in this network, while actor I is a ‘fringe’, due to having the lowest degree centrality. However, actor A has only connections in his immediate cluster or clique. Actor B has a lower DC, than average, but has the highest BC and is therefore a ‘bridge’ of this network. Actor B plays an important role between two otherwise not connected sub networks. Actors H and I will not receive any information, if actor B blocks it – thus H and I will be cut off. Actors C and D have a lower DC, than actor A (the hub), yet the pattern of their direct and indirect connections enables them to connect with any part of the network quicker than anyone else. They have the highest CC and a very high visibility to what is happening in the entire network.

illustration not visible in this excerpt

Figure 3: Kite network (Source: own illustration referring to Krackhart 1990)

Figure 4 illustrates the relationship between qualitative and quantitative measures, which will be relevant in forthcoming chapters.

illustration not visible in this excerpt

Figure 4: Relationship of qualitative and quantitative measures (Source: own illustration)

2.3 Determinants of a Successful Viral Campaign

Any viral campaign is a complex marketing procedure, which success is dependent on several factors (Subramani, Rajagopalan 2003, p. 306). We analyze four crucial determinants (content, environment, incentives and seeding strategy) in the following part (Langner 2006, p. 222).

2.3.1 Content

The core element of every viral campaign is its content (Langner 2006, p. 222). It has to be attractive for the recipient to distinguish itself from the flood of messages that consumers are confronted with nowadays (Dobele et al. 2005, p. 148). The content is usually used as a kind of ‘bait’ for the product itself. This requires a high degree of creativity from advertisers or the respective creative agencies. The attractiveness of the content boosts the consumer’s need to share the message with his/her peers. The following attributes are essential for an attractive message:

- Joy or entertainment – the content is extraordinary, diversified and has a high entertainment value
- New and unique – only innovative campaigns revive interest of consumers, whereas copycats will not be as successful
- Helpfulness – useful content e.g., small tools to help consumers solve their everyday problems
- Free (or partly free) availability – the use of the content must be free of charge
- Easy communicableness – the content is easy to forward and share.

Apart from these core attributes the format is also relevant in order to be successful (Langner 2006, pp. 223f). In the past, print media and other traditional, non-targeted methods, such as radio, television or direct mail were the only options for marketers to go for. The internet has opened a tremendous opportunity for cost effective advertising (Scott 2010, p. 6). The use of frequently used documents and platforms is a particular prerequisite to thrive (Esch et al. 2010, p. 115):

- Actions (events, news and rumors)
- Animations (MS Powerpoint, Flash and eCards)
- Videos (stream or file)
- Games (online or offline)
- Documents (MS Word, MS Excel and PDF)

A viral message has to be sticky, smooth and persistent. Stickiness means that the message sticks to the recipient, who is therefore likely to become a communicator of the message. This is why emotions play an important role for viral messages, because strong emotions increase the probability of sharing content (Dobele et al. 2007, p. 292). Six primary types of emotions are important, according to a study of qualitative interviews (Dobele et al. 2007, p. 300):

- Surprise (generated when something unexpected happens)
- Joy (experienced when moving forward towards a goal or achieving a goal)
- Sadness (expressed when not in a state of being well, generated after a fearful event)
- Anger (expressed when personally offended or someone suffers an injustice)
- Fear (generated when anticipating a certain pain, threat or danger)
- Disgust (expressed when an aversion against something can be felt)

Emotions implemented in viral messages should be chosen carefully depending on the purpose of the campaign or image of the brand e.g., fun brands should rather use an element of joy, whereas elements of anger should be utilized for single issue campaigns that require an immediate response to an injustice. However, any campaign should contain an element of surprise, because it appears to be the dominant emotion identified by prospects (Dobele et al. 2007, pp. 301f).

In our analysis, we do not focus on the content of a viral campaign. Therefore, we try to create a message of constant attractiveness in our experiment.

2.3.2 Environment

The environment i.e., the structure of the social network, is of high importance for the performance of VM campaigns (Bampo et al. 2008, p. 288). The analysis of network characteristics is therefore essential (Barabási 2003, p. 6). In 1967, sociologist Stanley Milgram conducted an experiment, where participants from a city in Nebraska had to send a letter to an unfamiliar stockbroker in Boston using their very own peers. Milgram concluded that there is a maximum of six nodes, which connect individuals with each other. He called it ‘six degrees of separation’ (Milgram 1967). Similar, more contemporary, experiments conducted by Dodds et al. (2003) and Leskovec, Horvitz (2008) also found approximately six (Dodds et al. 2003) or more accurate 6.6 (Leskovec, Horvitz 2008) intermediates that connect two individuals in our modern society. This phenomenon of a high degree of crosslinks in our society is known as the small world theory. Therefore, the usage of mass media is not crucial to propagate information (Röthlingshöfer 2008, pp. 13f).

Driven by several technological innovations, the network data on interpersonal relationships between consumers has become accessible for researchers. Therefore, the impact of network characteristics on the propagation of a message can be assessed (Katona et al. 2010, p. 6). Bampo et al. (2008) modeled the size and connectivity of a population as a network. Their goal was to assess the performance of different real viral campaigns and the influence of the underlying network structure on its success. In their simulation models, Bampo et al. (2008) distinguish between three different types of networks (Bampo et al. 2008, pp. 277f):

- Random networks (Erdös, Renyi 1959)
- Scale-free networks (Barabási, Jeong 1999; Barabási, Albert 1999; Albert, Barabási 2002)
- Small world networks (Newman, Watts 1999; Watts, Strogatz 1998)

The degree of the nodes of a random network follows a binomial distribution. On average the degree of a node is defined as [illustration not visible in this excerpt], where parameter [illustration not visible in this excerpt] stands for the total amount of nodes and [illustration not visible in this excerpt] for the probability of a connection between two nodes. As viral campaigns typically have a large [illustration not visible in this excerpt] and a very small [illustration not visible in this excerpt], the average degree remains moderate. Thus, it is very unlikely that any of the nodes is connected to a significant portion of other nodes. Scale-free networks provide a useful representation of many real networks e.g., the World Wide Web or the electric power grid of western part of the United States. The degree of its nodes supposedly follows a Power-law distribution. It allows for few nodes to have a very high degree, whereas most of the other nodes have rather few connections. Furthermore, the mean degree remains moderate. Two characteristics, dynamic growth and preferential attachment, are both important features of social networks which also apply to scale-free networks. Studies by Dorogovtsev, Mendes (2003) and Drineas et al. (2004) analyzing small e-mail networks based on the traffic of servers show, that the latter networks also appear to follow a Power-law distribution. In addition to the above mentioned features, this fact makes especially scale-free networks interesting from a VM perspective. Small world networks are characterized through high clustering and therefore a short average distance between nodes. This class of small world networks has first been introduced by Watts, Strogatz (1998). These characteristics form a network simulation model, which is especially interesting for VM. It reflects relationships that are established through physical proximity, due to friendship or professional relationships (Bampo et al. 2008, pp. 277f). Figure 5 illustrates these three bespoken graph types.

illustration not visible in this excerpt

Figure 5: a) random network, b) scale-free network, c) small world network
(Source: Huang et al. 2005)

The following work does not evaluate the success of VM campaigns in different network environments. However, our analysis is based on a real social network structure that fits closest to the above mentioned scale-free network characterized through a small mean degree and few nodes with a high degree i.e., hubs. Therefore, we do not modify the underlying environment as a determinant for success


[1] For an analysis of correlations between network measures, please see Valente et al. 2008.

[2] Batagelj, Mrvar 1998: Pajek (Slovene word for spider) is a program for Windows to analyze and visualize large networks. It is freely available and for non-commercial use.

[3] Smith et al. 2010: NodeXL is a free and open add-on for Excel 2007 or 2010 that provides social network extraction, analysis, and visualization features in the context of a spreadsheet.


ISBN (eBook)
ISBN (Buch)
1.5 MB
Institution / Hochschule
Technische Universität Darmstadt – Electronic Markets
viral marketing seeding social network analysis



Titel: Seeding - Can Marketers Take Advantage of Known Network Measures?