What makes a Web site popular?

50 November 2003/Vol. 46, No. 11 COMMUNICATIONS OF THE ACM

BY PETROS KAVASSALIS, STELIOS LELIS, MAHMOUD RAFEA, AND SEIF HARIDI
WHAT MAKES A
WEB SITE POPULAR?
To ensure a constant increase in user
interest, clicks, recommendations, loyalty, and
market share, first understand the information
flows and connection networks surrounding
each Web site in cyberspace.
Human cities in the physical world are unevenly
distributed centers of economic and social activity
characterized by a variety of structural elements,
including paths, edges, districts, nodes, and land-
marks. They organize themselves in time and
space as the result of the principles of conver-
gence, conflict, randomness, and site planning.
Likewise, information cities are envisioned as
thriving over the Internet, inhabited by tens of
thousands, even millions, of users (humans and
software agents) exhibiting complex navigational
behaviors and participating in cooperative online
buying, selling, chatting, interaction, collabora-
tion, and socializing. Evidence of this activity is
apparent in such highly popular Web portals as
www.ebay.com,, www.yahoo.com, www.aol.com,
�
COMMUNICATIONS OF THE ACM February 2004/Vol. 47, No. 2 51
ILLUSTRATION BY KARINE DAISAY

52 February 2004/Vol. 47, No. 2 COMMUNICATIONS OF THE ACM
and www.amazon.com that represent information
and/or commercial interests, business-to-business
hubs, and aggregations of virtual communities, some
routinely hosting millions of Internet users per day.
To explain the historical large-scale agglomerations
of human activity in the global economy, including
the concentration of social and economic activity into
physical cities and the clustering of entire business sec-
tors in specialized industrial and service districts like
Silicon Valley, Hollywood, and the City of London,
economists have developed a set of theories and ana-
lytical tools, synthesizing them into what they call the
New Economic Geography [6]. At the heart of the
responses it provides to the question “Where and why
does economic activity occur?” is the critical role of
“increasing returns.” In line with [3], which reintro-
duced the concept into modern economics, increasing
returns operate within industries, markets, and busi-
ness as positive feedback reinforcing the elements that
achieve success or aggravating the effect of loss.
Accordingly, proponents of the New Economic
Geography argue that concentrations of human pop-
ulations and their economic activity arise and are sus-
tained by some form of increasing returns—backward
and forward links associated with large local mar-
kets—initiating both circular and cumulative
processes, with agglomeration being self-reinforcing
[6]. It now appears the same strategy of increasing
returns explains the intriguingly similar rules describ-
ing the Web economy’s evolution and growth.
Increasing returns contribute to directing dispropor-
tionately large numbers of Internet users to the
largest, most compelling, most popular Web sites. But
how do increasing returns operate and what are their
implications for the Web-based economy?
Explaining Growth
The Web ecosystem includes an enormous variety of
sites and online pages. Evidence suggests that the
process of growth in this expanding business envi-
ronment follows two notable patterns:
Few large sites. Many small Web sites, but few large
ones, operate within the Web economy, so a relatively
few sites host the majority of Internet users; or, sur-
prisingly, as demonstrated in [1], the distribution of
visitors per Internet site follows a universal power law,
similar to the one found in the human population
distribution of the world’s larger cities.1
Quick growth. Popular Web sites grow quickly; for
example, Yahoo!, AOL, Amazon, and eBay have built,
as reported by consulting firm Morgan Stanley [10],
some of the fastest-growing, most valuable brands in
history, achieving that status relatively inexpensively.
Any attempt to explain Web growth would seem to
require a theory accounting for both patterns simul-
taneously. To begin with, a power law might success-
fully emerge from a stochastic model with the
assumption that the expected growth in popularity of
a Web site fluctuates in an uncorrelated fashion from
time interval to time interval around a positive mean
value and is independent of the site’s size [1]. How-
ever, the problem with this approach is that growth
rates selected randomly are exogenous, so they neither
reflect the history of the system’s evolution nor
explain the quick growth rates of the most popular
Web sites.2 To abandon the idea of random growth,
Web economists need a model in which Web site
growth and visitor population agglomeration emerge
endogenously from the behavior of individual
agents—the sites themselves and their visitors—and
their interactions. An increasing returns-based model
is one in which the presence of some sort of agglom-
eration economies influence the decisions of Internet
users concerning their visits to particular Web loca-
tions while allowing the final outcome to reflect the
history of the decision process.
But how might increasing returns in cyberspace be
modeled? Following a suggestion in [6], one can explore
a modeling approach involving random networks of
interaction among Internet users and Web sites, rather
than one involving random growth of the size of Web
sites. As [6] explains, the “randomness that creates the
power law may not involve random growth but random
‘connections’ in space. For example, imagine port cities
that serve the interior along a transport network formed
with random connections among transport nodes, with
the direction of the preferred connections reflecting
accidents either of history or geography. Alternatively,
we could suppose that the connections lie in some
abstract space of industry linkages” as might exist
between suppliers and manufacturers.
A computational model involving two superim-
posed interaction networks with random connections
in cyberspace might reproduce such randomness, thus
creating a power law regularity on the Web. In this
model, the first network (Web site connections) links
the sites; included are nodes corresponding to Web
sites and edges representing the links among them
that transport users from one site to another. The sec-
ond network (word of mouth) organizes social inter-
actions among Internet users, allowing for
word-of-mouth information propagation within a
1Many studies suggest that the distribution of larger cities worldwide follows a power
law in which a city’s size is inversely proportional to its rank in a list of cities ordered
by population [6].
2It would be more reasonable to expect the magnitude of growth fluctuations for a par-
ticular Web site to decrease with its size. Such a decrease is empirically the case of the
fluctuation in the growth rates of business firms [11].

structure consisting of local ties and long-range con-
nections. Each network frames the choices made by
Internet users about visiting sites in the sense that the
way information propagates introduces an informa-
tion-feedback mechanism into the process of compe-
tition among Web sites for market share
(information-based increased returns). For example,
Internet users stochastically
decide to visit Web sites with
probabilities depending on
numbers of links pointing-in
to the site (in-links); con-
versely, sites attracting large numbers of visitors
become more pointed-in than others (circular causa-
tion). On the other hand, sites that users learn about
through word of mouth depend on which sites other
users have already visited; Internet users are thus more
likely to learn about popular sites than unpopular
ones (information contagion).
Computational Implementation
We implemented this model in a large-scale agent-
based environment called the Web-Simulated Econ-
omy, developed at the Swedish Institute of
Computer Science (with the collaboration of the
Atlantis Group at the University of Crete) using
Mozart, a distributed software architecture (www.
mozart-oz.org); it allowed us to experiment and pro-
gressively produce global dynamic behavior. After t
time steps, the model leads to a scale-free state with
the distribution of visitors across Web locations fol-
lowing a power law (see Figure 1a). The model
incorporates six assumptions:
Two populations. Two small populations of agents
that increase exponentially over time represent Internet
users and Web sites with diverse amounts of offerings.
Portfolio of Web sites. Internet users organize their
site preferences in portfolios of choices, including their
most frequently visited sites. As the process evolves, a
portfolio may be updated with new sites identified via
word-of-mouth information propagation; at each time
step, some percentage of Internet user-agents query
other agents (friends and acquaintances) to recom-
mend their own favorite sites. At the same time, user-
agents explore the Web on their own, visiting new sites
by following the out-links of the sites they’ve already
visited. They might include these sites in their portfo-
lios if they find them interesting or useful (compared
with previously selected sites).
However, users are relatively loyal
to their portfolios, adding new
sites as they perceive value in
accessing and navigating them.
Utility function. To form and
update their portfolios of sites,
user-agents employ a utility func-
tion with two arguments: the
“performance characteristic” of a
site (sites are conceived as prod-
ucts, so a different performance
characteristic is attributed to each one for determining
its performance in terms of natural attractiveness, or
intrinsic quality); and the “match” between user pref-
erences and site offerings.
Web-site investment. Sites can deploy investment
strategies to improve their performance (in terms of
attractiveness), thus influencing the process of portfo-
lio formation and update. Investment is either soft (to
sustain performance) or aggressive (motivated by “ani-
mal spirit,” or greedy self interest), hoping to capitalize
on increased growth rates. Most such investments in
the Web economy follow the mimic model in the
sense that ambitious sites look to replicate successful
competitors’ investment strategies.3
Network structures. A small-world network, or a
network structure in which the average shortest path
between any two users is small while the clustering
coefficient is large, can help describe the dynamics of
social contacts among Internet users and mediate
word-of-mouth information propagation [12].
Inversely, the Web-site connection network emerges
from the stochastic decisions of Web sites to point-in
to popular sites, with the quantity of outgoing links
varying across sites according to an intrinsic prefer-
ence for employing a directory strategy involving the
categorization of links based on themes. (Some sites,
especially large-scale Web directories, constantly add
links, categorizing them to provide pointers to as
many popular Web resources as possible [8].) The
number of outgoing links also depends on a site’s rate
of growth; for example, popular sites, when employing
COMMUNICATIONS OF THE ACM February 2004/Vol. 47, No. 2 53
Proportion
of
sites
Proportion
of
sites
binned distribution
power law fit, p=1.19
Number of in-links
10
0
10
-2
10
-4
10
0
10
2
10
4
Proportion
of
sites
binned distribution
power law fit, p=1
Number of users
10
(a) (b) (c)
0
10
-2
10
-4
10
0
10
2
10
4
binned distribution
power law fit, p=1.25
Number of out-links
10
0
10
-2
10
-4
10
0
10
2
10
4
Figure 1. Fitted power law
distributions of numbers
of site: (a) visitors; (b)
in-links; and (c) out-links.
3Many business analysts argue that investments in the Web-based economy are boosted
through imitative behavior, perhaps because of a general fear that online competition can
quickly take users, customers, and profits away from companies that fail to constantly
improve their sites and provide the investment to support the effort [7].

a directory strategy, nat-
urally increase their
outflow of links more
rapidly than their less popular counterparts.
Entry strategies. New sites enter the Web economy
at different time steps using entry strategies of rela-
tively high initial investments that grow larger and
larger as the overall Web economy grows larger.
Beyond its capacity to accurately reproduce the
power law regularity, the overall model achieves inter-
esting results in terms of Web market efficiency and
Web economy organization, explaining, in the lan-
guage of organizational economics, how the scale-free
nature of the Web emerges in practice (see Figure 2).
The model produces several notably interesting
results. First, sites are rewarded for bringing in more
and more users by way of relative performance, or
their ability to compete in the marketplace, rather
than absolute performance, or their natural attractive-
ness or product quality. In many cases, sites with rel-
atively equal performance differ significantly in the
numbers of users they are able to attract and keep. We
verified that the effect of word-of-mouth information
propagation, combined with a hierarchical explo-
ration pattern privileging the best-connected Web
sites, is a powerful mechanism for promoting sites
that establish themselves quickly and become well
known quickly, while possibly excluding sites with
relatively good performance in terms of attractiveness.
Second, newly established sites may achieve a top-
ranked position, indicating weak correlation between
site age and number of user visits. The reason some sites
manage to top the charts so quickly is that once the pos-
sibility of economic behavior or
strategic investment is accorded to
newcomers, newly established sites
can, with a positive probability,
quickly accumulate large numbers
of incoming links (in-links), thus
surpassing older sites. Accordingly,
the model obtains only a limited
correlation between a site’s age and
the number of incoming links the
site acquires.
Third, visitor distribution
across Web sites is not the only
factor following a power law; fac-
tors that decay as a power law
include the number of incoming
links a Web site receives during the
course of the model (in-links), as
in Figure 1b, and the number of
outgoing links sites point-out
(out-links), as in Figure 1c. Such
behavior is another validation-against-reality test for
model results, the first being visitor distribution.
Comparative Advantages
The scale-free nature of the Web is explained else-
where in the context of preferential attachment-
based models [5], which assume a network
underpinned by two structural mechanisms: contin-
uous network expansion through addition of new
vertices and the preferential attachment of new ver-
tices to sites already well connected. An increasing-
returns-based model may obtain the same results by
modeling behaviors that induce positive feedback
into the process of Web sites competing for market
share. The more a site is visited, the more users are
aware of it and the more additional links it receives
(any given Web site generally wants to and does
point-in to popular sites). The more users learn
about (via word-of-mouth) and discover the site (via
user navigation paths reflecting the direction of the
links), the more visits it receives. In addition, since
the model also involves investment by Web site
owners hoping to improve site performance (in
terms of attractiveness), economic variables enter it
directly. Such investment generates potentially
diverse Web site performance. The model suggests
that the interplay between a large variation in the
landscape of Web-site performance and the complex
structure of the networks in which agents are
embedded can produce a power law.4
54 February 2004/Vol. 47, No. 2 COMMUNICATIONS OF THE ACM
User
portfolio
update
portfolio
word-of-mouth
exploration
visitors
Site
Web users' social network
Web sites' network
– matching
– performance
evaluation
create links
growth
more links
investments
replicate strategies
innovation
Figure 2. Graphical representation of
the Web-simulated economy model.
4See [2, 9] for other explanations of power law, also based on the interplay between
variation in quality and the complexity of the structure hosting individual agent inter-
action.

Moreover, in analyzing these increasing returns,
one makes an interesting observation about the
sources of growth in the Web economy. The model
reveals a specific growth process on the Web relating
to particular institutional structures, that is, the net-
works within which individual navigation and site
behavior are associated. Information economies in
general, and the Web economy in particular, generate
specific institutional structures consisting of random
information flows involving networks of interaction
among individual agents; as these structures are spe-
cific to information economies, they propel the
growth process for particular sites, quickly claiming
market share.
The model proves that the exceptionally high
growth rates achieved by the most popular Web sites
can be explained by information-based increasing
returns, information-feedback mechanisms in the
competition for market share, and links among sites
generated through point-in conventions. In this
regard, one can see (by observing a number of simula-
tions) the effect of parameter d, which represents the
number of sites Internet users visit through their per-
sonal explorations. When one eliminates this assump-
tion (d = 0), thus deactivating the Web-site-connection
network (a powerful generator of increasing returns), a
very different picture emerges, with no particularly
popular sites and relatively slow growth rates. As the
parameter d increases (implying a fluid transport net-
work like the Internet itself ), the number of sites with
quick growth rates increases considerably. The density
of the information propagation network, whereby
many users adopt word-of-mouth information atti-
tudes, seems to have similar influence, though weaker,
on the emergence of the fastest growing sites.
Finally, this broad perspective, which combines
behavioral and economic assumptions, explains the
scale-free nature of the Web in two ways: in terms of
information flowing within the social networks of
Internet users and in terms of connections among
sites via links associated with pointing-in to popular
Web sites. Web sites employ directory strategies to sat-
isfy existing users and attract new ones, but how
intensively a site employs them depends on its own
growth and popularity. In this context, growth and
Web competition are emerging phenomena on top of
information network structures.
Conclusion
E-marketers should thus investigate and leverage the
long-term ramifications of these structures to help
predict the behavior of Internet users toward their
organizations’ Web sites and to identify the best
Web-based ways to promote information about their
products. An increasing-returns approach has the
notable advantage of being able to identify the
sources of population agglomeration and growth on
the Web and, modeling them, provide useful
insights on how to diffuse marketing information
across the multiple networks in which user prefer-
ences are embedded. Software like TouchGraph
(www.touchgraph.com), for visualizing networks of
interrelated information, may be available within a
few years to Webmasters and e-marketers who want
to observe and analyze the networks involved in Web
site positioning and evolution.
References
1. Adamic, L. and Huberman, B. The Web’s hidden order. Commun.
ACM 44, 9 (Sept. 2001), 1–4.
2. Amaral, L., Buldyrev, V., Halvin, S., Salinger, A., and Stanley, H.
Power law scaling for a system of interacting units with complex inter-
nal structure. Physical Rev. Let. 80, 7 (Feb. 1998), 1385–1388.
3. Arthur, B. Increasing Returns and Path Dependence in the Economy. Uni-
versity of Michigan Press, Ann Arbor, 1994.
4. Arthur, B. and Lane, D. Information contagion. Structural Change and
Economic Dynamics 4, 1 (1993), 81–103.
5. Barabasi, A. and Albert, R. Emergence of scaling in random networks.
Science 286 (Oct. 1999), 509–512.
6. Fujita, M., Krugman, P., and Venables, A. The Spatial Economy: Cities,
Regions, and International Trade. MIT Press, Cambridge, MA, 1999.
7. Gordon, J. Does the ‘New Economy’ measure up to the great inven-
tions of the past? J. Econom. Perspect. 4, 14 (Fall 2000), 49–74.
8. Kleinberg, J. and Lawrence, S. The structure of the Web. Science 294
(Nov. 2001), 1849–1850.
9. Krugman, P. Confronting the mystery of urban hierarchy. J. Japanese
and Int. Econ. 10 (1996), 399–418.
10. Meeker, M., Mahaney, M., Joseph, D., Trowbridge, M., Cascianelli,
F., and Brown, M. Morgan Stanley Dean Witter. Global IU3, Brand
Value, and Customer Monetization for AOL, Yahoo, eBay, Amazon. Mor-
gan Stanley, 2001.
11. Stanley, H., Amaral, A., Buldyrev, V., Halvin, S., Leschhorn, H.,
Maass, P., Salinger, A., and Stanley, H. Scaling behavior in the growth
of companies. Nature 379 (Feb. 1996) 804–806.
12. Watts, D. and Strogatz, S. Collective dynamics of ‘small-world’ net-
works. Nature 393 (June 1998), 440–442.
Petros Kavassalis (petros@itc.mit.edu) is the director of the
Atlantis Group at the University of Crete in Greece.
Stelios Lelis (slelis@csd.uoc.gr) is a Ph.D. candidate in the Depart-
ment of Computer and Communication Engineering at the University
of Thessaly, Greece, and a research fellow in the Atlantis Group at the
University of Crete in Greece.
Mahmoud Rafea (mahmoud@mail.claes.sci.eg) is the director of the
Department of Knowledge Engineering and Expert System Building
Tools at the Central Laboratory for Agricultural Expert Systems in Egypt.
The work described here was done while he was a senior developer at the
Swedish Institute of Computer Science in Kista, Sweden.
Seif Haridi (seif@sics.se) is scientific leader of the Distributed
Systems Laboratory and Chief Scientific Advisor at the Swedish
Institute of Computer Science in Kista, Sweden. He is also a professor
of computer systems in the Department of Microelectronics and
Information Technology at the Royal Institute of Technology in Stock-
holm, Sweden.
This research covers the findings of the iCities project funded by the European Com-
mission (Information Cities Project: IST-1999-11337, Future and Emerging Tech-
nologies). We are particularly grateful to Hervé Tanguy (l’Ecole Polytechnique,
France) and Konstantin Popov (Swedish Institute of Computer Science, Sweden) for
their collaboration in this research.
© 2004 ACM 0002-0782/04/0200 $5.00
c
COMMUNICATIONS OF THE ACM February 2004/Vol. 47, No. 2 55