A recent discussion between some colleagues on the utility (or  
otherwise) of subject classification in repositories prompted me to  
undertake a brief investigation whose results I present here. (I'll  
also send this to AMSCI, so apologies for any duplicate copies that  
you see.) The discussion has broadly been between computer scientists  
and librarians over whether subject classification schemes offer  
advantages over Google-style text retrieval; the study below looks at  
the evidence as demonstrated in the usage of one particular  
repository. As such it doesn't address the intrinsic value of  
classification, but it does offer some insight into the effectiveness  
of navigational tools (including subject classification) in the  
context of a repository.

The University of Southampton Institutional Repository has been in  
operation for a number of years and an official (rather than  
experimental or pilot) part of its infrastructure for just over a  
year. As part of its capabilities, it includes lists of most recently  
deposited material, various kinds of searches, a subject tree based  
on the upper levels of the Library of Congress Classification scheme  
and an organisational tree listing the various Faculties, Schools and  
Research Groups in the University and a list of articles broken down  
by year of publication. These all provide what we hope are useful  
facilities for helping researchers find papers (ie by time, subject,  
affiliation or content).

Over a period of some 29.5 hours from 0400 GMT on March 7th 2006,  
1978 "abstract" pages (ie eprints records) were downloaded from the  
repository (ignoring all crawlers, bots and spiders).

Of the 1978 downloaded pages, the following URL sources (referrers,  
in web log speak) were responsible:
   439  - (direct URL, perhaps cut and paste into a browser or  
clicked on from an email client)
   225  EPRINTS SOTON pages
     25  OTHER SOTON WEB pages

ie the local repository facilities, including subject views and  
searches, led to only 225/1978 = 11% of all downloads.