MAMA: The URL Set
- MAMA's initial URL source—Yahoo!'s random URL generators
- DMoz (The Open Directory Project)
- Other URL sets used by MAMA: The W3C Member Company list and the Alexa 500
- DMoz overlap with the other URL sets
- Limiting domain over-representation in a URL set (domain-capping)
- Surface vs. Deep URLs and the MAMA URL set
- The domain-parking problem
- Protocols: To use or not to use, that is the question
- Web server error conditions
- Getting to know our set of URLs
How big is the Web?
In Sisyphean fashion, we will start this section by trying to calculate the size of the ever-evolving Web. As a shortcut, we will consider another attempt at this estimate from the WWW FAQ. It does this by roughly estimating the number of Web pages on the Web, based on some Yahoo! and Netcraft numbers from August 2005. The methodology used is coarse, but it is an effective, general handwaving assessment. The FAQ comes up with an "average number of pages per Web site" number (based on the Yahoo! and Netcraft data) of 273.
Assuming the average number of pages per Web site has not changed appreciably over time, the FAQ then extrapolates Feb 2007 numbers. Taking this number and adding on Apr 2008 site-reachability numbers from Netcraft (165,719,150), the total number of pages on the WWW becomes 45.2 billion (although Netcraft also lists an additional, unexplained quantity of "active" sites that would lower the rough total to about 18 billion). By comparison, MAMA ended up analyzing only 3.5 million URLs—a grain of sand on the beach, or a drop in the ocean, you could say.
Note: Since I originally wrote this, Google released a blog entry in July 2008 stating that their Web exploration system had reached a milestone of exposing 1 trillion unique URLs.
How to choose a URL set
MAMA's analysis total is a mere fraction of even a single percent of such a daunting total. It seems odd to say that 3.5 million of anything is insignificant. So let us assume for a moment that it is not. We are just not able to look at every Web page, so we must choose a smaller group of URLs to look at and justify that this is representative of the whole Web. One option is to choose a set of URLs selected at random. I had some conversations with Rene Saarsoo (author of an excellent previous study on coding practices), and he brought up many excellent points about the structure of the Web and choices in URL sets—some of which I have tried to paraphrase here. He definitely hit the nail on the head about random URL samples when he said:
"We can raise different problems that the truly random sample has, but it's hard to come up with some other methodology that wouldn't be even more flawed."
MAMA's URL set actually started life as an attempt at a random grouping, but it
has evolved into its current iteration to be decidedly non-random. How then do
we justify that any non-random group of URLs is sufficiently random enough
and/or representative of the larger Web? Can we just settle it by saying
looks like the Web because the sample is big—JUST LIKE THE WEB?
If we could look at the entire Web, the discussion about establishing the legitimacy of the URL set would be moot, surely? No, even that idea is a fallacy. If one could somehow look at all URLs in the publicly-accessible Web, it still doesn't include any URLs from the so-called private Web of intranets and other private URLs hidden behind passwords, forms, and other barriers to open access. These URLs cannot be lightly dismissed either. Added together, the overall URL space becomes many times bigger than our previously-stated WWW FAQ estimate. This hidden Dark Matter of the Web may also look very different to the public Web that MAMA samples data from, so there is no guarantee that these pages we can view are truly representative of what all pages look like. We will next take an in-depth look in this section at the non-random URL set that MAMA ended up using and examine the strategies, advantages and shortcomings inherent and employed where possible. We will put the URLs through the wringer to examine the usefulness of our set in representing the larger Web.
MAMA's initial URL source—Yahoo!'s random URL generators
In the early days of MAMA, a different source for URLs was used. Another tool on which the author was working polled the RYL service as part of its process to browse URLs automatically. That tool would log the URLs that it used, so they could be referred to later to reproduce any problems encountered. The US RYL service returned URLs from all over the world, whereas the Japanese RYL service (URL no longer active) returned Japanese-specific URLs that were useful when testing Asian language issues.
Over time, this automated browsing tool amassed a rather large number of logs. When MAMA was being developed, it was a simple task to write a script to comb these logs and create a unique list of URLs that could be used to test the nascent MAMA system. This list, consisting of ~750,000 URLs, became the core set of URLs used for MAMA analysis. The size of the list was manageable enough to crawl fully in just a few days of straight processing, and it represented a sizeable chunk of Web, useful for finding URLs using specific technologies for which MAMA searched. Given MAMA's resources and database query speed at the time, this was really about the most it could handle.
Note: The RYL links are only listed for posterity—sadly, the Japanese generator has been defunct for several years, and Yahoo!'s "regular" Random Link generator has been up and down over the years. Perhaps Yahoo! will someday choose to bring this useful tool back for good, but, for now, one can only be referred to it in the past-tense. Nevertheless, it served its purpose well for MAMA.
MAMA used this URL set repeatedly and continues to do so, but the efficacy of it is waning over time. The list of Yahoo! random URLs was a one-time, static snapshot, gathered between 2002-2004. Many factors, including link-rot and domain expiration, have reduced the usable URL space to less than 500,000 in the latest analysis. I discuss the Yahoo! URL set here only to note the evolution of MAMA's processes—URLs from this set are not included in this research unless they are also present in one of the URL sources that MAMA did use (DMoz, Alexa or W3C companies).
DMoz (The Open Directory Project)
"More input. MORE INPUT!"
--Johnny 5, Short Circuit
The RYL tool worked fine in its time and was a great progenitor to this present research. However, MAMA was ready to cover a wider URL space and needed a new source for URLs—where to get a sufficiently large set of URLs for MAMA to use?
The two feasible options were:
- To acquire a URL set
- To return to the Web-crawler idea
Writing a Web spider was still not that appealing at the time, because all of MAMA's planned development time was going to be devoted to the analysis tool for the foreseeable future. Acquiring a URL set had its own issues; companies that track or mine URLs in some manner (search engines such as Google or Yahoo!, or trend-tracking services such as Alexa) generally consider their URL databases to be either direct or indirect intellectual property. These companies protect their intellectual property either by providing limited access to their URL set, or charging for access. For example, if you search for "boat" on google.com, the results (at the time of writing) proudly say that there are 225 million search results—but you can only access the first 1,000 of those results. Unless you build your own Web crawler—the infrastructure for which can be quite an undertaking in itself—where can you get URLs of the necessary diversity and quantity to make an "analysis of the Web" interesting?
DMoz: Just what the doctor ordered
I discovered the publicly-available Open Directory Project (aka DMoz) list of URLs in 2004, between the first and second generations of MAMA development. I knew this would be MAMA's next eventual analysis goal, even though MAMA did not have the resources to analyze a set of that size at the time. DMoz had many advantages in its favor, including diversity, human selection, low repetition, and full categorization! It was also freely usable by anyone, so the results could be more openly verified. With ~5 million URLs, the DMoz set was about ten times MAMA's then-current sample size, so it seemed to be an ideal next step. Reinforcing the decision was the later discovery that Dagfinn Parnas had used DMoz in his university thesis research involving markup validation in 2001, and Rene Saarsoo had also used it in his more comprehensive university thesis research on Web page coding practices. MAMA would be treading on familiar ground.
About the DMoz URL set
DMoz is a great project that can best be summed up in its own words:
"The Open Directory Project is the largest, most comprehensive human-edited directory of the Web. It is constructed and maintained by a vast, global community of volunteer editors...The Open Directory powers the core directory services for the Web's largest and most popular search engines and portals, including Netscape Search, AOL Search, Google, Lycos, HotBot, DirectHit, and hundreds of others."
As of 24 March 2008, DMoz harnessed,
4,599,644 sites—79,311 editors—over
590,000 categories, but these numbers can fluctuate fairly wildly (on 27
January, 2008 it had 4,830,584 sites and 75,151 editors). Any analysis of DMoz
must take into account that DMoz grows and changes over time as editors add,
freshen, or delete URLs from its roster. URLs can grow stale or obsolete through
removal, and domains can and do die on a distressingly regular basis. The
aggregation source of these URLs remains the same, but the set itself is an
evolving, dynamic entity.
Dmoz has a liberal Free Use License
and the entire URL set is available for easy
download. Beware, however—the ~200MB gzipped RDF data file expands
to over 2GB of URLs and metadata! To pull the URLs out of the DMoz list, it
was a fairly trivial matter of separating out the content of any and all
<link r:resource > attribute values from the RDF data file
with a small Perl script. The initial DMoz snapshot for MAMA was taken in May
2007 and was used throughout the summer and into the fall during initial
testing. After the big full analysis took place, a new DMoz snapshot was
downloaded in late November 2007. Any new DMoz URLs not previously in MAMA
to that point were then added, which makes our final DMoz URL analysis set
effectively a superset of the May and November 2007 snapshots. This resulting
LONG list of URLs became the new basis for MAMA's analysis,
but a lot of work needed to be done to it before it was entirely usable for
our needs. We will discuss the ways that this URL set needed to be distilled
before examination could begin as well as some post processing that was also
Other URL sets used by MAMA: The W3C Member Company list and the Alexa 500
The W3C-member companies
The W3C is the organization that creates the CSS and markup standards and the markup validator used in this study. It is instructive to look at how URLs from these companies compare with URLs that are not as closely connected to the standards process. One would expect that the member company URLs [January 2008 snapshot list, latest live version] would adhere to the standards a little more closely than the general population—if for no other reason than they have an active stake in promoting them. Other studies have also looked at the W3C-member-company URLs to measure various standards compliance factors, so it was a natural fit for MAMA to include it as well.
The Alexa Global Top 500 list
Alexa is a service from Amazon. It utilizes Web crawling and user-installed browser toolbars to track the popularity of sites among the users of those toolbars. Alexa maintains, among many other useful measurements, a global "Top 500" [January 2008 snapshot list, latest live version]. The Alexa list was chosen primarily because the size of the list was similar in size to the W3C member company list—so even though MAMA might be comparing apples to oranges, at least it compared a fairly equal number of apples and oranges. The W3C company list skews toward academic and "big money" commercial computer sites, whereas the Alexa list is representative of what people actually use/experience on the Web on a day-to-day basis.
While few could argue that Alexa's "Top 500" list is relevant and popular, it has some definite biases:
- It is prejudiced toward big, popular sites with many country-specific variants, such as Google, Yahoo!, and eBay. This ends up reducing the breadth of the list. Google is the most extreme example of this, with 63 of the 487 URLs in the analyzed set being various regional Google sites.
- It includes the top pages of domain aggregators with varied user content, such as LiveJournal, Facebook, and fc2.com. These top pages are not representative of the wide variety of the user-created content they contain.
- The list consists entirely of top-level, entrance, or surface pages of a site. There is no intentional deep URL representation.
DMoz overlap with the other URL sets
DMoz simply does not cover all URLs in the world—as a human-filtered system it cannot do so. Nor does it even include all URLs that some of MAMA's other, much smaller sources consider important. Despite this, DMoz does overlap well with all the other sets MAMA has. The only facet of these lists that might generate any caution is that roughly half of the Alexa 500 set is NOT included in DMoz. Remember that these Alexa URLs are all considered to be "very important" by its own measure. However, from a quick perusal, one notices that many of the URLs from the Alexa list that are missing from DMoz fall into a few broad and possibly questionable categories that DMoz might avoid—namely, pornography, file sharing, and warez. That does not mean that the average user shies away from them though.
|URL Set||Total Set Size||Shared With DMoz||Percentage|
|Yahoo! Random initial set||462,970||171,248||36.99%|
|Alexa Global Top 500||487||261||53.59%|
|W3C Member Companies||413||227||54.96%|
Limiting domain over-representation in a URL set (domain-capping)
Not a new problem—MAMA's "Wikipedia Effect" becomes the "CNN Effect"
In the early days of MAMA's research, a problem surfaced that skewed some of the statistics returned. Searching for a specific technology via MAMA might return, say, 1,500 URLs matching the criteria, but in some instances over 1,000 of them might be from the same domain. At the time, I dubbed this the "Wikipedia Effect" because it was often the most frequent example of this problem. Wikipedia had several thousand URLs present in the URL set at that point and it used many exotic features that MAMA looked for.
After doing a first examination of DMoz, I found that my "Wikipedia Effect" could better be called the "CNN Effect"—the initial DMoz set from May 2007 contained 223,616 URLs from cnn.com domains out of 4,464,396 total URLs—over 5 percent of all its URLs! CNN was (by far) the worst offender of this issue. The solution to this was to keep track of the domains used in MAMA and how many URLs are analyzed from each domain. When a domain reached a specific number of URLs analyzed, MAMA would no longer analyze any more from that domain. Various domain caps were tried, but I settled for 30 for this analysis—it seems to strike the right balance between content diversity and over-representation. In MAMA's URL set, 682,058 URLs were rejected from analysis because their domains had exceeded the domain cap. Some 4,413 domains had URLs with a full complement of 30 URLs (see also the full frequency table).
Issues noticed with domain-capping
There are a number of issues that could arise when contemplating the use of domain-capping in URL analysis. MAMA's attempt was definitely first-generation, but it is a much-needed step in the right direction. Using a domain-capping system seems to be a new feature in this type of study—to my knowledge, this approach was not taken before (by Parnas or Saarsoo). The method used is not foolproof, and more could have been done with this. MAMA only did a simple, exact-string comparison to judge whether a domain was the same. This method does not allow for the common practice of having similar sub-domains on large Web sites. Looking again at CNN's site, although the specific domain cnn.com has only 30 URLs in MAMA, there are 35 CNN sub-domains in all represented in MAMA (such as edition.cnn.com, etc.). So, instead of the initial goal to have no more than 30 URLs from CNN, we now have 346! Future studies attempting to limit per-domain representation should try to take sub-domains into account.
Setting a conservative domain cap can be helpful if you are not sure of what
your URL population looks like yet. If bias or skew is later exposed in the
distribution of the URL population, the domain cap can be readjusted to help
remedy the situation. A final, fundamental question must be asked that challenges
the basic assumption about the need for domain capping:
Is having too many URLs
from the same domain inherently a BAD thing? That will not
necessarily always be the case. With MAMA, we wanted to stress breadth rather
than depth because the URL set is already quite breadth heavy (almost 94% of
the domains have only 1 URL representative).
Another incorrect assumption when adopting domain capping is that authoring patterns in URLs from the same domain will follow repeating patterns—this does not always hold true either. For some very large sites, the opposite is true, such as community hosting sites with user-generated content such as Earthlink.com or Geocities.com. In such cases, foreknowledge of important prominent hosting sites could lead one to use the equivalent of a whitelist to circumvent domain-capping mechanisms.
Surface vs. Deep URLs and the MAMA URL set
We can divide HTTP URLs into two types:Surface and Deep.
Surface URLs are the top-level/home page of a given domain and take the
[protocol]://[domain] (with an optional
trailing slash). An additional type of URL would also fall into this
category—a pseudo-Surface URL if you will, where the bare
Surface format also carries a default system filename, which is a moral
equivalent to the bare Surface version, such as
All other HTTP URLs fall into the flip-side category where there are extra
query arguments, path and/or other file information at the end of the
Surface URL form. The rough numbers from the WWW FAQ mentioned
before gave us 275 as a value for the average number of pages per site.
It is obvious that Deep URLs will almost always outnumber Surface URLs by
a very large margin (a domain will only have one Surface URL). If we picked
a page truly at random from a given domain, that page would almost never be
a Surface URL.
Hickson's research at Google is heavily skewed toward Deep URLs, to be more in keeping with the nature of the Web itself. Whatever other problems that URL set might or might not have, at least it can be said that it is fairly representative of the real world skew of Surface vs. Deep URLs. In choosing the DMoz URL set as the main candidate for analysis, both Saarsoo's and MAMA's research follow a different path. DMoz is VERY heavily biased towards Surface URLs—approximately 75% of the domains in MAMA have only Surface URLs represented! This is a direct result of the nature of the URL set itself and the set's human selection factor.
Surface URLs usually serve as the locus and focal point of the entire site, where access to the rest of the site is most easily achieved. They usually have more traffic than any of the site's individual Deep URLs, so of any single page to look at on a site, the Surface URL will often be the most interesting. The Surface page is most likely to have the most development scrutiny and effort; the reasonable side-effect of this being that the quality of these pages will likely be higher than many of the Deep URLs on the same site. On the other hand, many sites use templates to maintain uniform design, so the Surface URLs can often bear a great resemblance to other Deep URLs. Both of these reasons are some good arguments towards excusing the heavy weighting of Surface URLs in the DMoz set.
We are interested in an average user browsing experience, but such behavior will vary from site to site. How can we justify that the URLs we are examining represent the average user's browsing experience? We can start this process by inferring that a typical user will likely visit more than one page on a site and will also likely visit the "Home"/Surface URL of the site. To take steps to balance the DMoz URL set in the future, one might choose to examine each Surface URL looking for all the Deep URL hyperlinks contained within. One would then add a random Deep URL from these candidates to the URL set to be analyzed. This should at least bring the Surface to Deep URL ratio roughly into balance—a step in the right direction. Using such a balance, one could even classify URLs by their Surface/Deep nature and do comparisons between them to see how the general composition of each type of page fares.
Note: The equivalent of the following Perl regular expression was used on MAMA's URL set to detect pseudo-Surface URLs:
the following table shows the number of URLs MAMA looked at that were surface URLs, versus the number that were deep URLs.
|Description of domain condition||Quantity|
|Domain has a Surface URL||2,018,936|
|Domain only has Surface URLs||1,969,092|
|Domain has a pseudo-Surface 'Default filename' URL||352,963|
|Domain only has pseudo-Surface 'Default filename' URLs||342,535|
|Domain has Surface or pseudo-Surface URL but not Deep||2,312,533|
|Domain has a Deep URL||1,119,404|
|Domain only has Deep URLs||699,135|
The domain-parking problem
One issue to watch when dealing with a URL set of this size is URL atrophy. In addition to URLs going dead, the domain names under which they operate can come up for renewal and switch ownership. The too-common practice of "domain parking" occurs when an entity buys a domain name for the express purpose of trawling for click-throughs from users looking for something relevant to their original needs that brought them to that page. There are a number of very large-scale domain-parking companies that do this and it is BIG business, with several sites like www.dailydomainer.com and parkquick.com devoted entirely to the topic. Domain parking becomes a problem for a research project like this because the URL set becomes a potential random minefield of unwitting links to the same exact type of templated markup. In MAMA's case, a redirect to a new domain should eventually hit the domain cap and limit the "damage" of over-representation. Problem solved?
Not quite. The whole domain-parking issue came to light after noticing a common trend in the URLs that MAMA was processing. A URL that might have originally looked something like this:
was redirected to a domain-parking site with a URL like this:
This particular domain parker's way of using the original domain's name as
a starting point easily circumvented MAMA's domain capping mechanism. Other
domain parkers may utilize different mechanisms that complicate strategies
to detect legitimate URLs. Even though all 12,514 URLs matching
domainsponsor.com's URL pattern have been removed from consideration in this
research, this is a seemingly never-ending problem. One factor that can help
weed out domain parkers is that big parking companies will probably be
well-represented in ANY sizeable URL set, so odd authoring practices will
tend to stand out. For example, another domain parker (ActiveAudience) was
recently detected this way via unusual values for the
Name attribute. Sadly, this pattern was discovered too late and
the URLs were not excluded from MAMA.
Some well-known large scale domain-parking companies (there are many more):
Ian Hickson's Google research analyzed only URLs that ranked high in Google's PageRank technology. One hopes that Google's PageRank ranking system would rate domain-parking URLs much lower and the research would avoid the issue altogether. If the research being done does not have the benefit and backing of a major search-engine player though, vetting of URLs needs to be done from scratch. Future studies should be aware of the domain parking issue and take as many steps as possible to detect or eliminate domain parkers from their URL rosters. Otherwise, they end up skewing the results.
Protocols: To use or not to use, that is the question
DMoz contains some types of URLs that MAMA did not want to concern itself with. These non-markup URLs were removed from the list of URLs to be analyzed. MAMA only looked at HTTP or HTTPS protocol documents. We'll look more at the breakdown of HTTP protocol versions and other criteria detected by MAMA in a later section to be published in the future.
Web-server error conditions
An error page thus created is not something we want to count towards MAMA's statistics. There are certain repeating patterns that many URLs of this response type have—primarily the keywords "error" and "404". MAMA was able to leverage this to filter them out from consideration. This keyword detection approach works, but it is not without its problems. False positives are easy to find; for example, it is still not very long since terrorism became a popular topic of debate and discussion. Eliminating URLs with the "error" substring eliminates all URLs that discuss "terror" and "terrorism". So, there is more work to be done.
I never came up with a foolproof method for this, so I settled for a good enough approach. I eliminated all URLs from consideration that had the substrings "404" and "error" (but not "terror"). This eliminated 4,388 URLs from our final analysis set. It is not a perfect solution, but doing a manual scan of the excised URLs showed that less than 100 of these URLs are not error pages. Restoring these non-error, error-like URLs is not really that interesting considering the overall size of our URL set—100 pages out of several million seems an acceptable loss.
Getting to know our set of URLs
Now that we have deflated our initial exuberance over having a nicely-sized URL analysis set, let us take a look at some hard numbers about MAMA's URLs. There are 3,509,180 URLs in 3,011,668 domains, from 217 identified countries. 4,994 of those URLs use the HTTPS protocol. We will discuss more of the details of the set in other sections appropriate to their topics.
|Total analyzed URLs||3,509,180|
|Total analyzed domains||3,011,668|
|Domains having only 1 URL||2,826,740|
This article is licensed under a Creative Commons Attribution, Non Commercial - Share Alike 2.5 license.