ADAPTIVE EXPERIMENTATION

IN INTERACTIVE MARKETING:
THE CASE OF VIRAL MARKETING
AT PLAXO
KIRTHI KALYANAM, SHELBY MCINTYRE, AND J. TODD MASONIS

KIRTHI KALYANAM
is J.C. Penney Research Professor and

Director, Internet Retailing, Retail

T his case study shows how interactive marketing campaigns can be sys-
tematically fine-tuned and made more productive through adaptive experi-

Management Institute, Santa Clara mentation. It details the use of adaptive experimentation in a viral marketing

University, Santa Clara, California; campaign at Plaxo, Inc., a company that provides Internet-based updating

e-mail: [email protected] of personal contact information. The experiences of Plaxo highlight that even
for a product that is amenable to viral marketing, growth is not guaranteed.
SHELBY McINTYRE To achieve a desired level of growth, Plaxo identified the product features that
is in the Marketing Department contributed to greater adoption conversion of recipients of its marketing mes-
at Santa Clara University; sage and improved them through continuous experimentation. To overcome
e-mail: [email protected] potential negative side effects of aggressive viral growth, the company used a
carefully crafted feedback loop via internet alert services that tapped into the
J. TODD MASONIS
blogging community. This practice allowed management to better under-
is Founder and Vice President, Plaxo
stand negative perceptions of the product and take timely corrective actions.
Inc, Mountain View, California;

e-mail: [email protected] © 2007 Wiley Periodicals, Inc. and Direct Marketing Educational Foundation, Inc.

The authors thank Ram Shriram of

JOURNAL OF INTERACTIVE MARKETING VOLUME 21 / NUMBER 3 / SUMMER 2007

Sherpalo Ventures for his help Published online in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/dir.20086

and support.

72
 Journal of Interactive Marketing DOI: 10.1002/dir

INTRODUCTION records, including its patent applications. Due to this

unique access, we are able to piece together the
One of the virtues of direct marketing is the ability to sequence of events behind Plaxo’s successful viral
undertake some degree of experimentation with test marketing campaign.
mailings. This case study of Plaxo, Inc., an Internet ser-
vice that updates changes in personal contact infor- Plaxo launched its product with a viral marketing
mation across users’ electronic address books, shows campaign that ultimately attracted 5 million users in
how continuous adaptive experimentation1 focused on 3 years. However, although Plaxo’s product is amenable
a viral marketing concept generated a tight feedback to viral marketing, the company quickly learned it did
loop with its customers. This practice enabled the not guarantee high growth. Months after a launch and
company to go beyond a tipping point of customer a brief spurt in growth, daily adoptions dropped to
adoption into spiraling growth. just a handful of new users per day.

Viral marketing is a seductive concept. In a viral mar- Given lackluster growth and a declining cash posi-
keting campaign, a company uses the influence of its tion, Plaxo used a viral growth equation concept and
own customers to promote a product or service to an experimentation process for optimizing its “viral
prospective customers (Hanson & Kalyanam, 2006; engine.” The equation helped explicitly identify
Leskovec & Huberman, 2006). Viral marketing is an features of the product that impact viral growth.
Internet adaptation of marketing using the word- Experimentation and natural selection were used to
of-mouth effects, a phenomenon originally identified improve product features that impacted viral growth.
by Rogers (1995) in the context of the diffusion of This highly reasoned approach raised the viral index
innovations. Viral marketing models promise hockey beyond a “magic” threshold and helped the company
stick growth with little or no marketing expenditures move beyond a tipping point in its market. Within
(Jurvetson, 2000). The availability of a large network 3 months of implementing the concept, its user growth
of users, coupled with a low-cost communication mech- increased to a desired level of 5,000 new users per
anism, drives the realization of this promise. Hotmail day.
is one of the earliest cited examples of a successful viral
However, as the viral engine took off, it set off red
marketing campaign (Montgomery, 2001). Viral mar-
flags among the Internet community. Criticism of
keting used to be restricted to online companies, but it
Plaxo in blogs, chat forums, and press articles tar-
is now increasingly used by marketing managers at
nished it as an invasion of privacy and a generator of
Fortune 500 companies.2 However, what seems less
spam. It seemed that Plaxo had become like the myth
understood is that most viral attempts are likely to fail
of the Ebola virus, growing so aggressively that it was
(Wilson 2000). Unless a viral engine reaches a point of
in danger of killing its hosts. At that point, Plaxo
“internal combustion,” the effort is not likely to gener-
was faced with a painful dilemma, abandon the viral
ate sustained growth (Gladwell, 2000).
approach or ignore the negative press, risking further
damage to its brand.
This article takes the reader behind the scenes at
Plaxo, Inc. We collected information by interviewing
In the end, Plaxo found a way to strike the right
managers and examining the company’s archival
balance between spiraling viral growth and being a
good net citizen. The company hired a privacy officer
1
The term “adaptive experimentation” is defined in the AMA to get a fresh perspective and to investigate the
Marketing Dictionary as an approach (and philosophy) for man- privacy concerns that naturally occur with an address
agement decisions, calling for continuous experimentation to estab- book updating service. Plaxo changed product
lish empirically the market response functions.
features to moderate viral growth and established an
2
Burger King’s The Subservient Chicken and Coq Roq, Ford
active outreach program to turn its critics around.
Motor Company’s Evil Twin campaign, Heinz’s Ketchup Against
Tomato Cruelty campaign, McDonald’s “McRib Farewell Tour,”
Office Max’s Elfyourself campaign—see http://en.wikipedia.org/ The next section describes the problem of managing
wiki/Viral_marketing for many examples of major corporations contact information and Plaxo’s solution. Following
moving into viral marketing campaigns. that, the product launch is discussed together with

ADAPTIVE EXPERIMENTATION IN INTERACTIVE MARKETING 73

Journal of Interactive Marketing DOI: 10.1002/dir

the development of the viral equation and its imple- This step involved considerable switching costs for
mentation. Data on viral growth is presented, followed new users.
by an analysis of the negative effects of rapid viral
growth and how Plaxo learned to strike the right bal- The founders believed that the symmetric nature
ance. In closing, the article highlights some key of these existing networks limited their initial value. In
lessons learned. turn, that limited value inhibited rapid growth.
Therefore, the lack of rapid growth meant that the net-
CONTACTS JUST KEEP ON MOVING works could not support their value proposition. Instead
they tended to fall into disuse and ultimately fail.
In 2001, Todd Masonis and Cameron Ring were room-
mates at Stanford University. They saw the need for The founders thus saw the need for a system that
a service to update e-mail addresses and other contact (1) does not require membership in a private network
information that many individuals were struggling to as a prerequisite for providing substantial benefits,
maintain on their own computers. They found that (2) does not impose significant switching costs on
many of their own personal contacts were constantly adopters, (3) increases the value of its service by
out of date as college friends graduated, took jobs, got rapidly acquiring new members, and (4) provides a
married, and frequently moved. global address book of members and nonmembers.

Maintenance of up-to-date contact information Plaxo used e-mail to create an asymmetric updating
between friends, family, business associates, clients, service. In the Plaxo system, a user initiated update
and customers has always been a challenging task. In requests using e-mail. Since e-mail traverses propri-
a typical year, most people change at least some item etary networks, it enables updating across asymmet-
of contact information such as a phone number, a fax ric networks. The melding together of a proprietary
number, an e-mail address, or a physical address. One network and other networks using e-mail differenti-
study indicated that in a typical year, approximately ated Plaxo from previous approaches.
35% of Internet users change an e-mail address and
33% of mobile phone users change one of their num- The inherently viral property of the Plaxo service sug-
bers (Weidlich, 2000). Also, approximately 40 million gested that viral marketing would be a good fit
people change physical addresses in a given year. (Kempe, Kleinberg, & Tardos 2003). In 2002, the Plaxo
Furthermore, out-of-date contact information leads to management team decided that instead of paid adver-
losses of friendships or business opportunities for pro- tising, a viral campaign would establish the product
ductivity and revenue. Hence, a self-updating address and grow the user base. The company reasoned that as
book appeared to be a desirable service. Plaxo users initiate update requests to contacts in
their address books, they would expose and potentially
Services Based on Symmetric versus convert these contacts to use Plaxo.
Asymmetric Networks
In 2002, several conventional services did provide
WE HAVE BUILT IT. WILL THEY COME?
online storage and maintenance of contact informa-
Plaxo launched its service with a PR campaign. Next,
tion. However, these services were only able to provide
attention was focused on seeding the network from
updates to the contact information of other users of
the “top” (Weidlich, 2000; Kempe, Kleinberg, &
the same service—a situation where the value of the
Tardos 2003). Venture capitalists, who had big
service depended on the size of the installed customer
Rolodexes and were considered influential, were
base of the service (Shankar & Bayus, 2003). Plaxo
signed up as the first users of the product.3 When
management termed such services “symmetric,” in
that they required both parties of the information
exchange to be members of the service network. 3
This relates to the issue of effective product sampling to introduce
Another crucial issue with these conventional services a new product and generate word of mouth. Jain, Mahajan, and
was that they required the user to switch from their Muller (1995) conclude that target sampling to opinion leaders
current address book to the one provided by the service. should be more effective than neutral sampling.

74 JOURNAL OF INTERACTIVE MARKETING

 Journal of Interactive Marketing DOI: 10.1002/dir

FIGURE 1
New Users per Day

these venture capitalists became Plaxo users, they requests into new users. This intuition was formalized
sent out e-mails requesting updates to the contacts in into the following equation:
their address books.
Viral Index
V
N * Cr  1 (1)
Figure 1 graphs new users gained per day over the
first 6 months after the launch date of the product. This simple equation proved to be a powerful diagnos-
As can be seen, there was a spike in new users that tic for Plaxo managers. The Viral Index V indicates
accompanied the product launch. However, this spike the relative size of each new generation of customers
in new users soon dropped to a trickle. As people compared to the previous period. For example, when
returned to work in January, there was another brief V
1.1, each new generation of customers is 1.1 times
spurt in new users. But this soon settled down to an the size of the previous generation. When V  1.0,
average of 50 new users per day. At this rate, the com- each new generation is smaller than the previous one
pany would acquire only 18,000 new users per year. and the number of new customers eventually goes to
Viral adoption of the product would not meet Plaxo’s zero. V is indeed a magic number, and V
1 a magic
expectations, and Plaxo might never reach the tipping threshold. In an Internet marketing context, the viral
point (Gladwell, 2000). Management was left wonder- equation can be implemented by regularly collecting
ing whether the initial strategy to rely on viral mar- information on N and Cr from the company’s informa-
keting was flawed. tion systems. Managers at Plaxo calculated V at regu-
lar intervals and thereby monitored the traction of the
THE VIRAL EQUATION viral engine.

We can confirm that viral growth requires by V  1

The management team decided to take a closer look
generating cumulative adopter curves for different
at their execution of viral marketing. Intuitively, it
values of V. The cumulative adopter curve can be gen-
seemed that as long as every new user was able to
erated using N and Cr in the viral equation.4 Let us
bring at least one other new user, the user base would
suppose that on average each Plaxo user sends out
continue to grow. The number of new users that a cur-
rent user could bring to the user base, in turn,
depended on the number of update requests sent out 4
Godin (2001) provides a similar example in the context of a greeting
by the user (N) and the conversion rate (Cr) of those card company.

ADAPTIVE EXPERIMENTATION IN INTERACTIVE MARKETING 75

Journal of Interactive Marketing DOI: 10.1002/dir

800
V=0.99, V=1.00 V=1.01 V=1.01
700 N=5 N=5 N=5 N=2.89
Cr=0.198 Cr=0.20 Cr=0.202 Cr=0.35

Cumulative Number of Adopters 600

500

400

300

200

100

0
1 10 20 30 40 50 60 70 80 90 100 110
Period

FIGURE 2
Graph of Cumulative Number of Adopters

update requests to 5 people (N
5). Suppose also that In this sense V
1 is indeed a magic threshold.
each recipient of an update request has a 50% chance Finally, Figure 2 shows the difference between two
(Cr
0.5) of signing up to use the service. In this curves, both with V
1.1, but one with Cr
0.202
example, V
5*0.5
2.5. If Plaxo starts with 10 peo- (20.2%) and the other with Cr
0.35 (35%). Even
ple, then the size of the successive generations of new though the viral index is the same for both curves, the
adopters and their activities are as follows: curve with a higher Cr grows much faster, indicating
that the conversion rate is very important.
Generation 1: 10 new users sign up and send 50
update requests This viral equation is very simple. It does not explicitly
model more detailed effects such as clustering or the
Generation 2: 25 new users sign up and send
possibility that an e-mail today might result in a lagged
125 update requests
conversion over several time periods into the future.
Generation 3: 63 people sign up and send 315 However, even with this simple equation, it is possible
update requests. to forecast the number of new users that sign up in each
successive generation due to referrals from current
Notice that the second generation is 2.5 times the first users. This forecast would be analogous to the number
generation and that the third generation is 2.5 times of new adopters due to imitation effects in the Bass
the second. This is consistent with a V of 2.5. new product diffusion model (Bass, 2004 [1969]).5

Figure 2 uses this approach to generate and graph 5

Following Mahajan, Muller, & Bass (1995), if a(t), the number of
data on the cumulative number of new customers for new adopters at time t, is given by
each time period based on four different viral equa- q
a1t 2
p3 m  A1t 2 4  A1t 2 3 m  A1t 2 4
tions. For V
0.99 the cumulative growth curve m
peaks out and growth stalls. This would be the case where p is the coefficient of innovation, m is the total number of
potential adopters, A(t) is the cumulative number of adopters in
for all V  1. For V
1 the user base grows, but in a
time t and q is the coefficient of imitation. The second term in the
linear fashion. For values of V  1 we see nonlinear equation is the number of new adopters in time t due to internal
growth. In fact, even for a value of V
1.01, which influences or imitation effects. This is analogous to the forecast pro-
is barely greater than 1, we get exponential growth. vided by the viral equation.

76 JOURNAL OF INTERACTIVE MARKETING

 Journal of Interactive Marketing DOI: 10.1002/dir

However, Plaxo management focused on using the viral metric in the viral equation. The effect is two-fold:
equation as a tool for assessing alternative page treat- First, it comes from a friend or associate of the indi-
ments, the focus of this case study. vidual (rather than from Plaxo), and, second, it con-
tains a customized message that is composed directly
LINKING THE VIRAL EQUATION TO for that individual by someone who knows him well
PRODUCT FEATURES (certainly better than Plaxo, anyway). Rogers (1995)
has argued that mass media channels are relatively
This decomposition of viral growth was a powerful more important in generating awareness of an inno-
breakthrough for Plaxo. It allowed management to vation, whereas interpersonal communications are
focus on the product features that influenced N and important for persuasion. Hence, the persuasive value
Cr. Figure 3 links the two key metrics in the viral of person to person e-mails and the potential value of
equation to different product features. The viral mar- personalization.
keting problem was transformed into a product fea-
ture and design problem—aspects under the control The “Update Contact Wizard” (not shown) is Step 1 in
of management. Plaxo’s update sequence. A Plaxo member uses this
screen to select who should be sent an update request.
Figure 4 is “The Address Book Status” screen that Users can select contacts on a “one by one” basis, but
motivates the user to send out e-mails (Phelps, Mobilio, there is also a “check all” feature, the latter of which
David, & Raman, 2005) for contact update information. maximizes the number of update requests sent out
It provides a crucial statistic, namely the percent of (maximizing N ).
contacts that are currently out of date. This screen also
informs the user with notification of any pending Plaxo also has a feature that scans e-mails to find
update requests and the status of these requests. addresses in the headers and footers of e-mails.
These features serve to motivate the user and are a call Syntax-checking algorithms are applied to eliminate
to action. They influence N in the viral equation. invalid e-mails, and to avoid duplication with e-mail
addresses already present in the address book.
The “Personalizing Message” screen (not shown) The edited set of e-mail addresses is presented as an
affects conversion rates (Cr), which is the other crucial additional set of contacts under the address book.

FIGURE 3
Plaxo’s Viral Equation

ADAPTIVE EXPERIMENTATION IN INTERACTIVE MARKETING 77

Journal of Interactive Marketing DOI: 10.1002/dir

FIGURE 4
Address Book Status

This feature provides the Plaxo user the convenience To offer the user a way of checking on the number of
of automatically extracting e-mail addresses that are e-mails being sent out, Plaxo provided a bar chart
not in the address book. This set of contacts extracted that ranks the frequency of past communications with
from e-mail would typically be incomplete, not includ- each contact. The frequency chart feature also allows
ing other relevant information such as addresses and Plaxo to capitalize on the “strength of strong ties”
telephone numbers. Hence, this set of contacts would (Barabasi, 2003). The logic is that contacts with a
benefit significantly from an update request. From a high frequency of communications could be strong ties
viral marketing perspective, this feature has the and that the update requests would potentially have
powerful effect of expanding N. a high impact on them, thereby providing a higher
conversion rate (Cr).
While maximizing N is desirable from a viral spread-
ing perspective, it does lead to the possibility of many Plaxo also provided a “Review Selected Contacts”
inappropriate e-mails being sent out. As one of the screen (not shown) that reconfirmed to users the con-
founders points out: tacts (names and number) being updated. This allowed
users to catch and fix unintended errors before they
There is a lot of junk in peoples’ address books. sent the update requests.
By “junk” I mean people that the individual barely
knows. Over time old entries may be people the TINKERING AT THE MARGIN WITH
individual doesn’t even remember any more. When NATURAL SELECTION
Plaxo users decide to send out update requests
to their whole address book that includes e-mails to Figure 5, Screen A, shows one alternative of the
the junk part of the address list. “Presentation of Update Request” screen. This is the

78 JOURNAL OF INTERACTIVE MARKETING

 Journal of Interactive Marketing DOI: 10.1002/dir

Screen A

Screen B

FIGURE 5
Subtle Differences in Update Request Screens

ADAPTIVE EXPERIMENTATION IN INTERACTIVE MARKETING 79

Journal of Interactive Marketing DOI: 10.1002/dir

screen that the contact sees when he or she receives the inviting the contact to join the Plaxo service. This was
update request. The user’s information was provided in the one opportunity Plaxo had to display an explicit
a business card format that includes reply buttons to marketing message and hence was referred to as a
“change” or “confirm” the information. “sell page.”

Figure 5 shows the subtle differences between alter- The Plaxo sell page (not shown) is crucial for viral
native screens (Screen A vs. Screen B) for “Update adoption. Plaxo experimented with a number of
Request” e-mail messages. The Screen B alternative versions of this page to discover pages with high
showed a card with business and personal informa- conversion rates. As these descriptions show, many of
tion and a stand-alone “update” button. These mes- these changes were subtle and are “tinkering at the
sages are examples of many more subtle changes that margin.” A priori, it would be difficult to determine
altered different features of the pages. Many versions which versions would work better. Plaxo’s approach was
of these pages were deployed simultaneously and the very data driven. The company experimented with
company used field experimentation to select the win- the alternative versions to see which performed better
ning combination of features. over the short term and the medium term.

Once the contact finished updating or confirming the Figure 6 presents a conversion rate table for dif-
information, he or she is shown a confirmation page. ferent Web page versions. The company measured
Plaxo used this page not only to confirm that the conversion rates over a short duration (1–2 days),
update was complete, but also to present a message medium duration (7–14 days), and longer durations

FIGURE 6
Conversion Rates for Versions of Web Pages over Different Time-Frames

80 JOURNAL OF INTERACTIVE MARKETING

 Journal of Interactive Marketing DOI: 10.1002/dir

(7–35 days). For these Web pages, conversion was

measured based on the degree of accomplishment of
the intended purpose of each Web page. For example,
the “Presentation of Update Request” pages were
intended for users to request updated information via
e-mail to nonusers and possibly download the client
software and launch the service. On the other hand,
alternative pages for updating information were for
the user to update their own contact information.

Figure 7 shows the results obtained with this

reasoned experimental approach.6 Eventually, around
May 21, 2003, the number of e-mail update requests
(N) tipped. The 7–14 day conversion rate, which
initially had high variability, moved above 1.0 during
the week of July 2, 2003, and developed a rising
trend. This demonstrates the effect of iterative exper-
imentation on Cr.

To summarize, Plaxo found that its viral engine did

not grow automatically. Management developed a
viral equation to identify the various product features7
that would affect the components of viral growth.
Management then took a very data-driven, adaptive-
experimentation approach to select among various
FIGURE 7
The Growth in the Number of Update Requests
versions of product features.

AGGRESSIVE GROWTH CREATES Where the viral model can go wrong is when you get
A BACKLASH numerous requests coming close together with
the same Plaxo branding on the messages. Then the
The infamous Ebola virus does not spread more widely nonusers start feeling that they are updating Plaxo
because it kills off its host too quickly. In an analogous when they really only want to update their friend.
way, a reduction in trust of the Plaxo service might
threaten to kill off new adoptions, which might stop the Thus the viral engine can be too aggressive and stim-
viral spread much as in the Ebola syndrome. Aspects of ulate too many outgoing e-mails. At first thought, this
the Plaxo product, the viral model, and the environ- might be stimulating the nonuser to join Plaxo, but
ment combined to create trust problems for Plaxo. actually, in some instances, it serves instead as moti-
vation to not join the service.
In reality, Plaxo was a proxy, collecting information on
behalf of its users. It was important to make sure that One blogger suggested that since Plaxo did not seem to
nonusers had the perception that they were updating have a revenue source, the company must be selling
their friend’s address book, and not Plaxo. However, personal contact information to third parties. Other
frequent exposure to Plaxo messages and branding companies such as Gator, at the time, were being
started to have the unintended effect of creating the accused of doing this. Figure 8 provides an example of
wrong impression. As the founders explained: a negative blog.

6
See Montgomery (2001) for an extended analysis about applying Unfortunately for Plaxo, an e-mail asking for personal
quantitative analysis on the Internet. contact information is a key tactic for an identity thief.
7
Rosen (2000) identifies product features that people “buzz” about. In fact, a special term, “phishing” (as in “fishing for

ADAPTIVE EXPERIMENTATION IN INTERACTIVE MARKETING 81

Journal of Interactive Marketing DOI: 10.1002/dir

Plaxo is a mean system and should not be supported. It requires you to "opt out" separately for every person who might
Plaxo you. That is ludicrous! There should be a law that requires a "global opt out" option. I've received many Plaxo requests
and they are almost all from people who've never personally e-mailed me, let alone taken time to talk to me. Why on earth do
they need my contact information I don't know. Especially because it's not hard to find. Plaxo is a very selfish, egocentric
way to run one's personal world. It's basically a spam system fueled by vanity and self-importance that bludgeons people into
participation. That's almost as bad as Google holding a cache of the entire Internet but not allowing anybody to scrape its
pages. There's a simple rule: Do to others as you would have them do to you. Plaxo and Google have launched a new
morality: Be so powerful as to get away with stuff that nobody can afford to let you not get away with. Oh, may you crumble
and I tread on your ruins and celebrate with the free people we shall all be. I think these corporations are all nests for demons,
empty shells, vehicles for eliminating responsibility, genies who have all the rights of humans with none of the moral
responsibilities. Corporations - you will serve mankind in hell as the self-destructive machines that consume all wicked
fictions. Thank you, and With best wishes.
Posted by: Andrius Kulikauskas | November 3, 2005 07:56 PM

Same thing happened to me this week. Someone asked me to update my info, and when I did, holy crap, it sent it to everyone
I'm linked to and it's been too long for me to know who that is. And since I had changed some stuff I wouldn't have changed
if I had known it was being transmitted to the Pope and his wife, I was a bit chaffed as well.

Posted by: jeneane | November 3, 2005 11:35 PM

FIGURE 8
Examples of Negative Blogs

information”), has developed to describe the practice.8 Plaxo’s original approach was to maximize the number
Therefore, when there are a number of Plaxo e-mails of outgoing e-mails to stimulate the most growth.
being sent, it is easy to get apprehensive that some- However, this approach created the perception of
one is phishing. Even if the e-mail is from a trusted Plaxo as a spam generator. The founders explained:
contact, the information might still be hijacked by
someone surreptitiously reading that e-mail. With When someone leaves a company they often send
concerns of identity theft on the rise at the time of the out their updated contact information. With Plaxo,
Plaxo launch, this concern was being magnified. the updating is so easy that far more contacts
are included. Some sales person might have 5,000
Plaxo also sent tracking information to users to individuals on their contact list, so imagine what
convince them to join. This inadvertently created a happens if they change a cell phone number and
perception of a violation of privacy. As the founders decide to update everyone in the list with that
explained: information. Some of those 5,000 contacts are not
going to even remember the person.
In the update that we used to send, the update
request said something like “You have received 11 Nonusers can get annoyed because they get little
update requests. You should join Plaxo, and then perceived value from their updating efforts. This is
you wouldn’t have to do all this updating.” But we exacerbated by the frequency of requests and heightens
learned that this just turned off the nonuser all the the perception of these requests being spam-like. Plaxo
more because we were telling them that we kept messages can also cause a workflow diversion. Many
track of how many messages they were getting. people attack their e-mail stream with a “get through it”
It came across as very big brother. attitude. Going through the whole updating process
during an e-mail session heightens the perception of
8
Wikipedia defines phishing as “a form of social engineering, spam, it being an unwelcome interruption at that time.
characterized by attempts to fraudulently acquire sensitive
information, such as passwords and credit card details, by mas-
querading as a trustworthy person or business in an apparently STRIKING THE RIGHT BALANCE
official electronic communication, such as an e-mail or an instant
message. The term phishing arises from the use of increasingly In November, 2003, Plaxo hired a privacy officer.
sophisticated lures to “fish” for users’ financial information and His assignment was to build trust among the com-
passwords.” See www.wikipedia.org. Accessed February 26, 2006. munity. The key actions taken by the privacy officer

82 JOURNAL OF INTERACTIVE MARKETING

 Journal of Interactive Marketing DOI: 10.1002/dir

included toning down the product features, providing The privacy officer became very active in the blogo-
a netiquette policy, addressing security concerns, and sphere, in user forums, and on Plaxo’s own Web page.
better clarifying the privacy actions and then commu- Web-alert services were used to notify the company
nicating the new policies quickly and broadly. immediately of negative posts. The privacy officer
made it his practice to respond to these quickly. He
Among the product changes was a modification to was able to contact many key bloggers and journalists
make the contact wizard less aggressive. In the initial and succeeded in turning several of them from critics
design, when the software was installed, the “update into supporters of the company.
contact wizard” was automatically launched. The ini-
tial screen came up with all the contacts highlighted Plaxo also discovered that a small percentage of users
and a button encouraging the user to update contacts. was generating the most complaints and problems for
As users quickly went through the steps, the system the service. These “bad apples” would send update
would blast out an e-mail to everyone in the address requests to thousands of users at a time and more
book; often hundreds of messages. While this is very frequently than a casual user. Since most users
viral in terms of maximizing N in the viral equation, would use the service responsibly, a small number of
it can cause trust issues (and the Ebola syndrome) in bad users would generate a huge amount of e-mail
the long run. In the newer versions of the software, and detract from the overall value of the system. Plaxo
this “send all” option was removed. Instead, a chart added automatic throttling features to limit the number
was added to inform the users of how frequently they and frequency of update requests and proactively dis-
had been contacting each person in their address abled these problem users.
book. The user was then expected to select those at
the top of the frequency-of-contact list, individuals to These sophisticated and nuanced actions allowed the
whom they would have strong ties. company to sidestep the Ebola syndrome. In the
months following these actions, negative posts about
The company also developed a policy on “proper neti- Plaxo declined dramatically compared to the previous
quette for Plaxo members,” trying to encourage them months. The hockey stick growth continued, with the
to maintain proper empathy for their contacts as they company growing to 15 million users by August 2006.
sent out update requests. Members were advised to
(a) not publicize a businFess or event on Plaxo, (b) send
update requests only to people who clearly know them KEY LESSONS
and would want to receive their information, (c) per- With the decline in effectiveness of traditional advertis-
sonalize the update request, (d) make sure to provide ing, and fundamental changes in the media landscape,
plenty of information about who the requester is, and marketing managers are turning to Internet-based
(e) respect others’ privacy and not bombard them with techniques such as viral marketing. One of the essential
multiple update requests. benefits of the Internet environment is the ability for
continuous experimentation of “word of mouth,” but
The company developed a “privacy and security” page, now with feedback about each contact and its effective-
along with a privacy comparison matrix. The sell page ness. Since viral marketing does not require any media
also was changed to contain a statement to the effect investment, it is very tempting. However, this study
that “Plaxo is very concerned with your privacy.” suggests that viral marketing is hard work. Rather
than the use of just marketing dollars, it requires inge-
To counter phishing fears and to provide a secure envi- nuity, tenacity, and attention to nuances through
ronment, the company implemented several security research.
services (e.g., VeriSign and eTrust) and began to state
all this clearly and often on the Plaxo Web page, to the This study provides a number of important lessons
bloggers, and in the update request e-mails themselves. about viral marketing. First and foremost, it shows
The messages also informed users that privacy policies even when a product is inherently amenable to viral
are enforced by the FTC and that not following them marketing, fast growth is not guaranteed. It may
would lead to federal prosecution of Plaxo. need active management. Our findings support the

ADAPTIVE EXPERIMENTATION IN INTERACTIVE MARKETING 83

Journal of Interactive Marketing DOI: 10.1002/dir

arguments provided by prior work (Godin, 2001; was achieved by tasking a single individual with both
Wilson, 2000) that managerial action can impact the product development and customer growth.
success of a viral campaign. For example, Godin (2001,
p. 28) notes that “it is possible to dramatically improve Articles on viral marketing have tended to focus pri-
the chances that [an] idea virus will catch on and marily on how to grow the customer base. They have
spread.” ignored the possibility that viral campaigns can cre-
ate negative effects such as perceived violations of
Second, the fact that the readiness of the viral engine privacy. This case study is unique in its analysis of the
is perhaps the most important and primary aspect of occurrence and nature of negative effects. The case
any viral marketing campaign. The network should study shows that monitoring blog postings for nega-
be seeded after the viral engine is known to be ready tive perceptions can be fruitful. Managers should
through adaptive experimentation. One road to fail- appreciate the distinct and complementary nature of
ure, as Plaxo almost learned, is to reverse any part of the information obtained from iterative experimenta-
this priority sequence, such as seeding the network tion versus those from blogs. The former improves
before the product features are adequate. This priori- product features for growth, the latter provides the
ty sequence does not seem to be well appreciated in feedback loop as to how growth objectives should be
the literature (see, for example, Rosen, 2000). balanced against negative perceptions that viral cam-
paigns can create. As the case of Plaxo demonstrates,
The concept of a viral equation is not new. It has been this type of information is useful in overcoming the
articulated as a mathematical description of how a Ebola syndrome in viral marketing.
virus-like spread occurs in a social system (Granovetter,
1978, 1983; Schelling, 1972, 1978; Miller, McIntyre, & Aggregate diffusion models like the Bass model are
Mantrala, 1993). However, this case study shows that based on observing only one variable—the number of
the viral equation can be implemented to diagnose and new adopters at time t. The model then partitions
improve the health of the viral engine in two broad the new adopters into external and internal influ-
steps. First, the two key metrics in the viral equation ences. Viral marketing campaigns provide direct mea-
can identify the product features that impact viral surements on multiple variables including, but not
growth. Next, these features can be improved quickly limited to, (1) the number of new adopters that can be
and incrementally using a process of iterative experi- tracked to external influences, (2) the number of new
mentation. adopters that can be tracked to internal influences,
(3) the number of messages sent by the previous
In a viral marketing campaign, a manager should cohort group (N), and (4) the Conversion rate (Cr) of
focus on customer conversion rate. Anemic conversion these messages. Incorporating these variables into
cannot ultimately be overcome with massive messag- the diffusion model could be an interesting topic for
ing. Plaxo learned that turning up the volume (high N) future research.
meets resistance and further depresses conversion cre-
ating a downward spiral. While the viral index V must
be greater than the magic threshold of 1.0, this can be REFERENCES
achieved with many combinations of Cr and N, in the Armstrong, A., & Hagel, J. (1996). The Real Value of On-line
short run. But for longer run success, it is raising Communities. Harvard Business Review, May–June.
Cr that provides for healthy growth. Barabasi, A.-L. (2003). Linked: How Everything is Connected
to Everything Else and What It Means for Business,
Science and Everyday Life. Cambridge, MA: Plume Books.
In the case of Plaxo, small and subtle changes in fea-
Bass, F. (2004). A New Product Growth for Model Consumer
tures had a big impact on growth. An important lesson
Durables. Management Science, 50(12) supplement
here is that managers should not underestimate the [Originally published 1969], 1825–1832.
effects of “tinkering at the margin.” Quick and itera- Godin, S. (2001). Unleashing the Ideavirus, Dobbs-Ferry,
tive experiments seem to be well suited to identifying NY: Do You Zoom, Inc.
improvements. This tinkering at the margin requires Granovetter, M. (1978). Threshold Models of Collective
an extreme level of synergy between marketing and Behavior. American Journal of Sociology, Vol. 83,
product management. In Plaxo’s case this coordination 1420–1443.

84 JOURNAL OF INTERACTIVE MARKETING

 Journal of Interactive Marketing DOI: 10.1002/dir

Hanson, W., & Kalyanam, K. (2006). Internet Marketing Moore, G. (1991). Crossing the Chasm, HarperCollins, N.Y.,
and eCommerce. Thomson: Mason, OH. NY.
Jain, D., Mahajan, V., & Muller, E. (1995). An Approach for Phelps, J. E., Mobilio, L., David, P., & Raman, N. (2005). Viral
Determining Optimal Product Sampling for the Marketing or Electronic Word-of-Mouth Advertising:
Diffusion of a New Product. Journal of Product Innova- Examining Consumer Responses and Motivations to
tion Management, 12, 2, 124–135. Pass Along E-mail. Journal of Advertising Research, 44,
Jurvetson, S. (2000). What Exactly is Viral Marketing? Red 333–348, Cambridge University Press.
Herring, 78, 10–112. Rogers, E. (1995). The Diffusion of Innovations (5th ed.). The
Kempe, D., Kleinberg, J., & Tardos, E. (2003). Maximizing Free Press: NY.
the Spread of Influence through a Social Network. Rosen, E. (2000). The Anatomy of Buzz, Currency-
Conference on Knowledge Discovery in Data, Proceedings Doubleday: NY.
of the Ninth ACM SIGKDD International Conference on Schelling, T. (1972). Dynamic Models of Segregation.
Knowledge Discovery and Data Mining, Washington, Journal of Mathematical Sociology, Vol. 1, 143–186.
D.C. 137–146. Schelling, T. (1978). Micromotives and Macrobehavior.
Leskovec, J., Adamic, L., & Huberman, B. (2006). The Norton and Company: NY.
Dynamics of Viral Marketing. Proceedings of the 7th Shankar, V., & Bayus, B. L. (2003). Network Effects and
ACM Conference on Electronic Commerce, 228–237. Competition: An Empirical Analysis of the Home Video
Mahajan, V., Muller, E., & Bass, F. (1995). Diffusion of New Game Industry. Strategic Management Journal, 24,
Products: Empirical Generalizations and Managerial 375–384.
Uses. Management Science, 14(3): 979–989. Weidlich, W. (2000). Socio-dynamics: A Systematic
Miller, C. M., McIntyre, S. H., & Mantrala, M. K. (1993). Approach to Mathematical Modeling in the Social
Toward Formalizing Fashion Theory. Journal of Sciences. Harwood Academic Publishers: Amsterdam,
Marketing Research, Vol. 30(May), 142–157. 2000.
Montgomery, A. (2001). Applying Quantitative Marketing. Wilson, R. F. (2000). The Six Simple Principles of Viral
Techniques to the Internet. Interfaces, 31, 2, 90–108. Marketing. Web Marketing Today, 70.

ADAPTIVE EXPERIMENTATION IN INTERACTIVE MARKETING 85