The deep Web
(also called Deepnet
, the invisible Web
, or the hidden Web
) refers to World Wide Web
content that is not part of the surface Web
, which is indexed
by search engines
. It is estimated that the deep Web is several orders of magnitude
larger than the surface Web.
Michael Bergman mentioned that Jill Ellsworth used the term "invisible Web" in 1994 to refer to websites that are not registered with any search engine. Bergman cited a January 1996 article by Frank Garcia:
"It would be a site that's possibly reasonably designed, but they didn't bother to register it with any of the search engines. So, no one can find them! You're hidden. I call that the invisible Web."
Another early use of the term invisible Web was by Bruce Mount (Director of Product Development) and Matthew B. Koll (CEO/Founder) of Personal Library Software, Inc. (PLS) when describing the @1 deep Web tool. The term was used in a December 1996 press release from PLS.
The first use of the specific term deep Web occurred in that same 2001 Bergman study.
In 2000, it was estimated that the deep Web contained approximately 7,500 terabytes of data and 550 billion individual documents. Estimates – based on extrapolations
from a study done at University of California, Berkeley
– show that the deep Web consists of about 91,000 terabytes
. By contrast, the surface Web (which is easily reached by search engines) is only about 167 terabytes. The Library of Congress
contains about 11 terabytes.
Deep Web resources may be classified into one or more of the following categories:
- Dynamic content – dynamic pages which are returned in response to a submitted query or accessed only through a form, especially if open-domain input elements (such as text fields) are used; such fields are hard to navigate without domain knowledge.
- Unlinked content – pages which are not linked to by other pages, which may prevent Web crawling programs from accessing the content. This content is referred to as pages without backlinks (or inlinks).
- Private Web – sites that require registration and login (password-protected resources).
- Contextual Web – pages with content varying for different access contexts (e.g., ranges of client IP addresses or previous navigation sequence).
- Limited access content – sites that limit access to their pages in a technical way (e.g., using the Robots Exclusion Standard, CAPTCHAs or pragma:no-cache/cache-control:no-cache [header]s), prohibiting search engines from browsing them and creating cached copies.
- Non-HTML/text content – textual content encoded in multimedia (image or video) files or specific file formats not handled by search engines.
To discover content on the Web, search engines use web crawlers
that follow hyperlinks
. This technique is ideal for discovering resources on the surface Web
but is often ineffective at finding deep Web resources. For example, these crawlers do not attempt to find dynamic pages that are the result of database queries due to the infinite number of queries that are possible. It has been noted that this can be (partially) overcome by providing links to query results, but this could unintentionally inflate the popularity (e.g., PageRank
) for a member of the deep Web.
One way to access the deep Web is via federated search based search engines. Search tools such as Science.gov are being designed to retrieve information from the deep Web. These tools identify and interact with searchable databases, aiming to provide access to deep Web content.
Another way to explore the deep Web is by using human crawlers instead of algorithmic crawlers. In this paradigm referred to as Web harvesting, humans find interesting links of the deep Web that algorithmic crawlers can't find. This human-based computation technique to discover the deep Web has been used by the StumbleUpon service since February 2002.
In 2005, Yahoo! made a small part of the deep Web searchable by releasing Yahoo! Subscriptions. This search engine searches through a few subscription-only Web sites. Some subscription websites display their full content to search engine robots so they will show up in user searches, but then show users a login or subscription page when they click a link from the search engine results page.
Crawling the deep Web
Researchers have been exploring how the deep Web can be crawled in an automatic fashion. Raghavan and Garcia-Molina (2001) presented an architectural model for a hidden-Web crawler that used key terms provided by users or collected from the query interfaces to query a Web form and crawl the deep Web resources. Ntoulas et al. (2005) created a hidden-Web crawler that automatically generated meaningful queries to issue against search forms. Their crawler generated promising results, but the problem is far from being solved.
Since a large amount of useful data and information resides in the deep Web, search engines have begun exploring alternative methods to crawl the deep Web. Google’s Sitemap Protocol and mod oai are mechanisms that allow search engines and other interested parties to discover deep Web resources on particular Web servers. Both mechanisms allow Web servers to advertise the URLs that are accessible on them, thereby allowing automatic discovery of resources that are not directly linked to the surface Web.
Federated search by subject category or vertical is an alternative mechanism to crawling the deep Web. Traditional engines have difficulty crawling and indexing deep Web pages and their content, but deep Web search engines like CloserLookSearch, Science.gov and Northern Light create specialty engines by topic to search the deep Web. Because these engines are narrow in their data focus, they are built to access specified deep Web content by topic. These engines can search dynamic or password protected databases that are otherwise closed to search engines.
It is difficult to automatically determine if a Web resource is a member of the surface Web or the deep Web. If a resource is indexed by a search engine, it is not necessarily a member of the surface Web, because the resource could have been found using another method (e.g., the Sitemap Protocol
, mod oai
) instead of traditional crawling. If a search engine provides a backlink
for a resource, one may assume that the resource is in the surface Web. Unfortunately, search engines do not always provide all backlinks to resources. Even if a backlink does exist, there is no way to determine if the resource providing the link is itself in the surface Web without crawling all of the Web. Furthermore, a resource may reside in the surface Web, but it has not yet been found by a search engine. Therefore, if we have an arbitrary resource, we cannot know for sure if the resource resides in the surface Web or deep Web without a complete crawl of the Web.
The concept of classifying search results by topic was pioneered by Yahoo! Directory search and is gaining importance as search becomes more relevant in day-to-day decisions. However, most of the work here has been in categorizing the surface Web by topic. This classification poses a challenge while searching the deep Web whereby two levels of categorization are required. The first level is to categorize sites into vertical topics (e.g., health, travel, automobiles) and sub-topics according to the nature of the content underlying their databases. Several deep Web directories are under development such as OAIster by the University of Michigan, INFOMINE at the University of California at Riverside and DirectSearch by Gary Price to name a few.
The more difficult challenge is to categorize and map the information extracted from multiple deep Web sources according to end-user needs. Deep Web search reports cannot display URLs like traditional search reports. End users expect their search tools to not only find what they are looking for quickly, but to be intuitive and user-friendly. In order to be meaningful, the search reports have to offer some depth to the nature of content that underlie the sources or else the end-user will be lost in the sea of URLs that do not indicate what content lies underneath them. The format in which search results are to be presented varies widely by the particular topic of the search and the type of content being exposed. The challenge is to find and map similar data elements from multiple disparate sources so that search results may be exposed in a unified format on the search report irrespective of their source.
The first commercial deep Web tool was @1 from Personal Library Software (PLS), announced December 12th, 1996 in partnership with large content providers. According to a December 12th, 1996 press release, @1 started with 5.7 terabytes of content which was estimated to be 30 times the size of the nascent World Wide Web. PLS was acquired by AOL in 1998 and @1 was abandoned.
- Barker, Joe (Jan 2004). Invisible Web: What it is, Why it exists, How to find it, and Its inherent ambiguity UC Berkeley - Teaching Library Internet Workshops.
- Bergman, Michael K. (2001). "The Deep Web: Surfacing Hidden Value". The Journal of Electronic Publishing 7 (1):
- Gruchawka, Steve (June 2006). How-To Guide to the Deep Web TechDeepWeb.com, http://TechDeepWeb.com
- Hamilton, Nigel (2003). The Mechanics of a Deep Net Metasearch Engine - 12th World Wide Web Conference poster.
- He, Bin; Chang, Kevin Chen-Chuan (2003). "Statistical Schema Matching across Web Query Interfaces". Proceedings of the 2003 ACM SIGMOD International Conference on Management of Data .
- He, Bin; Patel, Mitesh; Zhang, Zhen; Chang, Kevin Chen-Chuan (2007). "Accessing the Deep Web: A Survey". Communications of the ACM (CACM) 50 (2): 94–101.
- Ipeirotis, Panagiotis G.; Gravano, Luis; Sahami, Mehran (2001). "Probe, Count, and Classify: Categorizing Hidden-Web Databases". Proceedings of the 2001 ACM SIGMOD International Conference on Management of Data, 67-78. .
- King, John D.; Li, Yuefeng; Tao, Daniel; Nayak, Richi (2007). "Mining World Knowledge for Analysis of Search Engine Content". Web Intelligence and Agent Systems: An International Journal 5 (3): 233–253.
- McCown, Frank; Liu, Xiaoming; Nelson, Michael L.; Zubair, Mohammad (2006). "Search Engine Coverage of the OAI-PMH Corpus". IEEE Internet Computing 10 (2): 66–73.
- Ntoulas, Alexandros; Zerfos, Petros; Cho, Junghoo (2005). "Downloading Textual Hidden Web Content Through Keyword Queries". Proceedings of the Joint Conference on Digital Libraries (JCDL), 100-109. . Extended version
- Personal Library Software (Dec 1996) press release announcing @1 as an "Invisible Web" search service
- Price, Gary; Sherman, Chris (2001). The Invisible Web : Uncovering Information Sources Search Engines Can't See. CyberAge Books.
- Raghavan, Sriram; Garcia-Molina, Hector (2001). "Crawling the Hidden Web". Proceedings of the 27th International Conference on Very Large Data Bases (VLDB), 129-138. .
- Shestakov, Denis (June 2008). Search Interfaces on the Web: Querying and Characterizing. TUCS Doctoral Dissertations 104, University of Turku
- Wright, Alex (Mar 2004). In Search of the Deep Web, Salon.com, http://www.salon.com/tech/feature/2004/03/09/deep_web/