Numerous studies dating as far back as 1994 have shown that the distribution of web requests follows the Zipf-law.  More precisely, the probability that a request is for the i'th most popular page is C/i (where C is a normalizing constant).  However, other studies have refuted this claim.  Breslau et al. conduct their own study in this paper (using much larger data sets) and find that the distribution of web requests follows a "Zipf-like" law: C/(i^a) where  0 < a <= 1.  They further show that this result supports empirical observations regarding cache hit-ratios and request inter-arrival times.

This paper presents a more thorough and convincing argument in support of the "Zipf-like" distribution theory than previous papers.  First, because the authors have performed their study using a much larger data set (6 organizations, over 1 million requests in each trace).  Second, the authors use the theoretical model to extrapolate and support additional empirical observations of web request behaviour.  Finally, they apply simulations of different cache algorithms to verify the expected relative performances of each algorithm.  The thoroughness of the authors in gathering data and performing validations gives strength to their claims.

However, this study (and previous similar studies) only considers web pages (retrieved from HTTP GETs) to study "web requests".  As such, there is no insight into the behaviour of data-heavy (FTP, online gaming) or data-streaming (multimedia) web requests, which composes a great deal of Internet traffic. 

More interestingly, none of the charts shown in the paper follow "Zipf-like" behaviour completely.  More precisely, the most popular and least popular pages do NOT exhibit Zipf-like behaviour.  In fact, the most and least popular pages consistently lack Zipf-like behaviour.  An interesting question remains, do these particular pages follow a different distribution that is statistically significant?  This could potentially lead to an interesting categorization of web requests:

Most Popular Pages - Distribution 'X'
Least Popular Pages - Distribution 'Y'
"Gray area" Pages - Zipf-like Distribution