Most commentators when referring to web server market share link to Netcraft's Web Server Survey. However, the IIS plugin provider Port80 Software suggest that this survey should be considered carefully. There is a large disconnect between NetCraft's headline figures (last month putting Apache on 67.4% of all domain names with IIS on 21.0%) and Port80's survey of Fortune 1000 companies (IIS 53.8%, Netscape/SunONE 18.0%, Apache 15.4%).
Firstly, discard Netcraft's headline figures. The fact is that one domain name does not equal one machine or one installation of a web server. Many of the domain names on the web are hosted on hosting providers - indeed, Netcraft offers a hosting provider league table. The average in Netcraft's analysis is for one physical server to have 3.4 IP addresses, each hosting 10.6 hosts (with 2.4 children :-D). Many of these sites are merely 'parked' - basically, they're the 'Buy this domain now!' pages that you sometimes see when guessing a domain name incorrectly. Netcraft try to eliminate this effect to produce the Active Sites data.
Looking at the data for last month, we see that Apache's share goes up to 68.3%, while IIS' climbs to 23.4% (the big loser is SunONE, dropping from 3.39% to 1.01% - about 1.3m sites (over 80%!) hosted on SunONE are parked, according to these statistics). Now, IIS' share has still fallen since last month, but not by as much.
We still don't know how many distinct machines or distinct organisations are in each grouping.
Port80's analysis is quite different: instead of trying to get the largest number of servers, they're trying to consider the 'most significant' servers (my quotes) - the ones run by the largest companies. There don't appear to be any government departments or voluntary organisations in the analysis, though, which could skew the results somewhat.
The difference is striking, and worth considering, though.
I'm a little annoyed that Netcraft don't release the statistics on underlying operating systems, although here the methodology becomes more suspect, and can be swayed by intervening filter/load balancing devices - for example, earlier this year there were a number of misinformed articles about www.microsoft.com running on Linux. Microsoft seem to have recently reverted to hosting the site directly. Anyway, I'd like to see how many of the Apache sites run on Linux - and how many run on a BSD.