Viral marketinsome viral campaigns have performed phenomenally well - for example, a viral video by Unilever in 2006 generated over 2.3 million views in its first 10 days and three times more traffic to the company's website than a Super Bowl commercial had - many campaigns falter. Ralf van der Lans of HKUST and his co-authors Gerrit van Bruggen, Jehoshua Eliashberg and Berend Wierenga set out to increase a campaign's prospects by devising a model to predict how many customers a viral campaign can reach, how the reach evolves, and how this depends on marketing activities.

"Because messages from friends are likely to have more impact than advertising, and because information spreads rapidly over the Internet, viral marketing is a powerful marketing communication tool. But not all viral marketing campaigns are successful and it's important that marketers be able to predict the returns on their expenditures and thus how many customers they will reach," they say.

Their model takes an approach from epidemiology that describes a branched, or stepped, rate of infection. In marketing campaigns, "infection" comes from seeding a group of customers with a marketing message in the hope that the customers will pass the message on to their contacts, and those contacts to their contacts, and so on. Seeding is either through email or online or offline advertising.

The authors' model gathers information at every point - from the moment when the customer receives the invitation, to whether they open it, click to the campaign website, land at the website, participate in the campaign and forward the message to their friends. That information is then used to make predictions.

To test the model's effectiveness, the authors applied it to a campaign held from April 1-May 5, 2005 by a large financial services company. Seeding was carried out in the first 14 days through 6,400 banner clicks and 28,758 seeding emails. By the end of the 36 days, there had been 228,351 participants, all of whom had visited the campaign website and provided their contact information.

The database gathered by the model included not only the identity of the participant, but also their date of participation, source and date of invitation, number of e-mails sent to friends, and number of friends who were already participating or invited.

That information was processed to predict the campaign's reach with remarkable accuracy. By Day 7, before the seeding was even completed, the branching model predicted a reach of 221,429 participants. None of the other models tested by the authors even came close to the true figure.

In addition, the model showed which methods of seeding were most effective. E-mails from a friend carried a 0.26 probability that a person would participate in the campaign against 0.12 for an e-mail sent by the company. But banner clicks carried a 0.34 probability, probably because customers who click on banners are already interested in the campaign.

The authors were also able to apply the model to "what-if" scenarios to see how many more customers would be generated by additional investment in banners or seeding e-mails at different stages in the campaign.

"The term 'viral' may (incorrectly) suggest information spreads automatically. But marketers need to actively manage the viral process to facilitate the spread of information," the authors say.

"Our model incorporates the effects of marketing activities such as seeding e-mails, bannering and traditional advertisements on the viral process, which standard branching models don't allow for. This enables marketers to accurately forecast the reach of a campaign within a few days. We believe our model is a valuable tool to develop and optimize viral marketing campaigns."