Research and Advances
Architecture and Hardware

Accessing the Deep Web

Attempting to locate and quantify material on the Web that is hidden from typical search techniques.
  1. Introduction
  2. Results
  3. Conclusion
  4. References
  5. Authors
  6. Footnotes
  7. Figures
  8. Tables

The Web has been rapidly “deepened” by massive databases online and currentsearch engines do not reach most of the data on the Internet [4]. While the surface Web has linked billions of static HTML pages, a far more significant amount of information is believed to be “hidden” in the deep Web, behind the query forms of searchable databases, as Figure 1(a) conceptually illustrates. Such information may not be accessible through static URL links because they are assembled into Web pages as responses to queries submitted through the query interface of an underlying database. Because current search engines cannot effectively crawl databases, such data remains largely hidden from users (thus often also referred to as the invisible or hidden Web). Using overlap analysis between pairs of search engines, it was estimated in [1] that 43,000–96,000 “deep Web sites” and an informal estimate of 7,500 terabytes of data exist—500 times larger than the surface Web.

With its myriad databases and hidden content, this deep Web is an important yet largely unexplored frontier for information search. While we have understood the surface Web relatively well, with various surveys [3, 7]), how is the deep Web different? This article reports our survey of the deep Web, studying the scale, subject distribution, search-engine coverage, and other access characteristics of online databases.

We note that, while the study conducted in 2000 [1] established interest in this area, it focused on only the scale aspect, and its result from overlap analysis tends to underestimate (as acknowledged in [1]). In overlap analysis, the number of deep Web sites is estimated by exploiting two search engines. If we find na deep Web sites in the first search engine, nb in the second, and n0 in both, we can estimate the total number as shown in Equation 1 by assuming the two search engines randomly and independently obtain their data. However, as our survey found, search engines are highly correlated in their coverage of deep Web data (see question Q5 in this survey). Therefore, such an independence assumption seems rather unrealistic, in which case the result is significantly underestimated. In fact, the violation of this assumption and its consequence were also discussed in [1].


Our survey took the IP sampling approach to collect random server samples for estimating the global scale as well as facilitating subsequent analysis. During April 2004, we acquired and analyzed a random sample of Web servers by IP sampling. We randomly sampled 1,000,000 IPs (from the entire space of 2,230,124,544 valid IP addresses, after removing reserved and unused IP ranges according to [8]). For each IP, we used an HTTP client, the GNU free software wget [5], to make an HTTP connection to it and download HTML pages. We then identified and analyzed Web databases in this sample, in order to extrapolate our estimates of the deep Web.

With its myriad databases and hidden content, this deep Web is an important yet largely unexplored frontier for information search.

Our survey distinguishes three related notions for accessing the deep Web: site, database, and interface. A deep Web site is a Web server that provides information maintained in one or more back-end Web databases, each of which is searchable through one or more HTML forms as its query interfaces. For instance, as Figure 1(b) shows, is a deep Web site, providing several Web databases (a book database, a music database, among others) accessed via multiple query interfaces (“simple search” and “advanced search”). Note that our definition of deep Web site did not account for the virtual hosting case, where multiple Web sites can be hosted on the same physical IP address. Since identifying all the virtual hosts within an IP address is rather difficult to conduct in practice, we do not consider such cases in our survey. Our IP sampling-based estimation is thus accurate modulo the effect of virtual hosting.

When conducting the survey, we first find the number of query interfaces for each Web site, then the number of Web databases, and finally the number of deep Web sites.

First, as our survey specifically focuses on online databases, we differentiate and exclude non-query HTML forms (which do not access back-end databases) from query interfaces. In particular, HTML forms for login, subscription, registration, polling, and message posting are not query interfaces. Similarly, we also exclude “site search,” which many Web sites now provide for searching HTML pages on their sites. These pages are statically linked at the “surface” of the sites; they are not dynamically assembled from an underlying database. Note that our survey considered only unique interfaces and removed duplicates; many Web pages contain the same query interfaces repeatedly, for example, in, the simple book search in Figure 1(b) is present in almost all pages.

Second, we survey Web databases and deep Web sites based on the discovered query interfaces. Specifically, we compute the number of Web databases by finding the set of query interfaces (within a site) that refer to the same database. In particular, for any two query interfaces, we randomly choose five objects from one and search them in the other. We judge that the two interfaces are searching the same database if and only if the objects from one interface can always be found in the other one. Finally, the recognition of deep Web site is rather simple: A Web site is a deep Web site if it has at least one query interface.

Back to Top


(Q1) Where to find “entrances” to databases? To access a Web database, we must first find its entrances: the query interfaces. How does an interface (if any) locate in a site, that is, at which depths? For each query interface, we measured the depth as the minimum number of hops from the root page of the site to the interface page.1 As this study required deep crawling of Web sites, we analyzed one-tenth of our total IP samples: a subset of 100,000 IPs. We tested each IP sample by making HTTP connections and found 281 Web servers. Exhaustively crawling these servers to depth 10, we found 24 of them are deep Web sites, which contained a total of 129 query interfaces representing 34 Web databases.

We found that query interfaces tend to locate shallowly in their sites: none of the 129 query interfaces had depth deeper than 5. To begin with, 72% (93 out of 129) interfaces were found within depth 3. Further, since a Web database may be accessed through multiple interfaces, we measured its depth as the minimum depths of all its interfaces: 94% (32 out of 34) Web databases appeared within depth 3; Figure 2(a) reports the depth distribution of the 34 Web databases. Finally, 91.6% (22 out of 24) deep Web sites had their databases within depth 3. (We refer to these ratios as depth-three coverage, which will guide our further larger-scale crawling in Q2.)

(Q2) What is the scale of the deep Web? We then tested and analyzed all of the 1,000,000 IP samples to estimate the scale of the deep Web. As just identified, with the high depth-three coverage, almost all Web databases can be identified within depth 3. We thus crawled to depth 3 for these one million IPs.

The crawling found 2,256 Web servers, among which we identified 126 deep Web sites, which contained a total of 406 query interfaces representing 190 Web databases. Extrapolating from the s =1,000,000 unique IP samples to the entire IP space of t = 2,230,124,544 IPs, and accounting for the depth-three coverage, we estimate the number of deep Web sites as shown in Equation 2, the number of Web databases as shown in Equation 3, and the number of query interfaces as shown in Equation 4 (the results are rounded to 1,000). The second and third columns of Table 1 summarize the sampling and the estimation results respectively. We also compute the confidence interval of each estimated number at 99% level of confidence, as the 4th column of Table 1 shows, which evidently indicates the scale of the deep Web is well on the order of 105 sites. We also observed the multiplicity of access on the deep Web. On average, each deep Web site provides 1.5 databases, and each database supports 2.8 query interfaces.


The earlier survey of [1] estimated 43,000 to 96,000 deep Web sites by overlap analysis between pairs of search engines. Although [1] did not explicitly qualify what it measured as a search site, by comparison, it still indicates that our estimation of the scale of the deep Web (on the order of 105 sites), is quite accurate. Further, it has been expanding, resulting in a 3–7 times increase in the four years from 2000–2004.

(Q3) How “structured” is the deep Web? While information on the surface Web is mostly unstructured HTML text (and images), how is the nature of the deep Web data different? We classified Web databases into two types: unstructured databases, which provide data objects as unstructured media (text, images, audio, and video); and structured databases, which provide data objects as structured “relational” records with attribute-value pairs. For instance, has an unstructured database of news articles, while has a structured database for books, which returns book records (for example, title = “gone with the wind,” format = “paperback,” price = $7.99).

By manual querying and inspection of the 190 Web databases sampled, we found 43 unstructured and 147 structured. We similarly estimate their total numbers to be 102,000 and 348,000 respectively, as summarized in Table 1. Thus, the deep Web features mostly structured data sources, with a dominating ratio of 3.4:1 versus unstructured sources.

(Q4) What is the subject distribution of Web databases? With respect to the top-level categories of the directory as our taxonomy, we manually categorized the sampled 190 Web databases. Figure 2(b) shows the distribution of the 14 categories: Business & Economy (be), Computers & Internet (ci), News & Media (nm), Entertainment (en), Recreation & Sports (rs), Health (he), Government (go), Regional (rg), Society & Culture (sc), Education (ed), Arts & Humanities (ah), Science (si), Reference (re), and Others (ot).

The distribution indicates great subject diversity among Web databases, indicating the emergence and proliferation of Web databases are spanning well across all subject domains. While there seems to be a common perception that the deep Web is driven and dominated by e-commerce (for example, for product search), our survey indicates the contrary. To contrast, we further identify non-commerce categories from Figure 2(b)—he, go, rg, sc, ed, ah, si, re, ot—which together occupy 51% (97 out of 190 databases), leaving only a slight minority of 49% to the rest of commerce sites (broadly defined). In comparison, the subject distribution of the surface Web, as characterized in [7], showed that commerce sites dominated with an 83% share. Thus, the trend of “deepening” emerges not only across all areas, but also relatively more significantly in the non-commerce ones.

(Q5) How do search engines cover the deep Web? Since some deep Web sources also provide “browse” directories with URL links to reach the hidden content, how effective is it to crawl-and-index the deep Web as search engines do for the surface Web? We thus investigated how popular search engines index data on the deep Web. In particular, we chose the three largest search engines Google (, Yahoo (, and MSN (

We randomly selected 20 Web databases from the 190 in our sampling result. For each database, first, we manually sampled five objects (result pages) as test data, by querying the source with some random words. We then, for each object collected, queried every search engine to test whether the page was indexed by formulating queries specifically matching the object page. (For instance, we used distinctive phrases that occurred in the object page as keywords and limited the search to only the source site.)

While there seems to be a common perception that the deep Web is driven and dominated by e-commerce (for example, for product search), our survey indicates the contrary.

Figure 3 reports our finding: Google and Yahoo both indexed 32% of the deep Web objects, and MSN had the smallest coverage of 11%. However, there was significant overlap in what they covered: the combined coverage of the three largest search engines increased only to 37%, indicating they were indexing almost the same objects. In particular, as Figure 3 illustrates, Yahoo and Google overlapped on 27% objects of their 32% coverage: a 84% overlap. Moreover, MSN’s coverage was entirely a subset of Yahoo, and thus a 100% overlap.

The coverage results reveal some interesting phenomena. On one hand, in contrast to the common perception, the deep Web is probably not inherently hidden or invisible: the major search engines were able to each index one-third (32%) of the data. On the other hand, however, the coverage seems bounded by an intrinsic limit. Combined, these major engines covered only marginally more than they did individually, due to their significant overlap. This phenomenon clearly contrasts with the surface Web where, as [7] reports, the overlap between engines is low, and combining them (or metasearch) can greatly improve coverage. In this case, for the deep Web, the fact that 63% objects were not indexed by any engines indicates certain inherent barriers for crawling and indexing data. Most Web databases remain invisible, providing no link-based access, and are thus not indexable by current crawling techniques; and even when crawlable, Web databases are rather dynamic, and thus crawling cannot keep up with their updates.

(Q6) What is the coverage of deep Web directories? Besides traditional search engines, several deep Web portal services have emerged online, providing deep Web directories that classify Web databases in some taxonomies. To measure their coverage, we surveyed four popular deep Web directories, as summarized in Table 2. For each directory service, we recorded the number of databases it claimed to have indexed (on their Web sites). As a result, was the largest such directory, with over 70,000 databases.2 As shown in Table 2, compared to our estimate, it covered only 15.6% of the total 450,000 Web databases. However, other directories covered even less, in the limited range of 0.2%–3.1%. We believe this extremely low coverage suggests that, with their apparently manual classification of Web databases, such directory-based indexing services can hardly scale for the deep Web.

Back to Top


For further discussion, we summarize the findings of this survey for the deep Web in Table 3 and make the following conclusions. While important for information search, the deep Web remains largely unexplored and is currently neither well supported nor well understood. The poor coverage of both its data (by search engines) and databases (by directory services) suggests that access to the deep Web is not adequately supported. In seeking to better understand the deep Web, we’ve determined that in some aspects it resembles the surface Web: it is large, fast-growing, and diverse. However, they differ in other aspects: the deep Web is more diversely distributed, is mostly structured, and suffers an inherent limitation of crawling.

To support effective access to the deep Web, although the crawl-and-index techniques widely used in popular search engines have been quite successful for the surface Web, such an access model may not be appropriate for the deep Web. Crawling will likely encounter the limit of coverage, which seems intrinsic because of the hidden and dynamic nature of Web databases. Further, indexing the crawled data will likely face the barrier of structural heterogeneity across the wide range of deep Web data. The current keyword-based indexing (which all search engines do), while serving the surface Web pages well, will miss the schematic structure available in most Web databases. This situation is analogous to being limited to searching for flight tickets by keywords only; not destinations, dates, and prices.

As traditional access techniques may not be appropriate for the deep Web, it is crucial to develop more effective techniques. We speculate that the deep Web will likely be better served with a database-centered, discover-and-forward access model. A search engine will automatically discover databases on the Web by crawling and indexing their query interfaces (and not their data pages). Upon user querying, the search engine will forward users to the appropriate databases for the actual search of data. Querying the databases will use their data-specific interfaces and thus fully leverage their structures. To use the previous analogy of searching for flight information, we can now query flights with the desired attributes. Several recent research projects, including MetaQuerier [2] and WISE-Integrator [6], are exploring this exciting direction.

Back to Top

Back to Top

Back to Top

Back to Top


F1A Figure 1a. The conceptual view of the deep Web.

F1B Figure 1b. Site, databases, and interface.

F2A Figure 2a. Distribution of Web databases over depth.

F2B Figure 2b. Distribution of Web databases over subject category.

F3 Figure 3. Coverage of search engines.

Back to Top


T1 Table 1. Sampling and estimation of the deep Web scale.

T2 Table 2. Coverage of deep Web directories.

T3 Table 3. Summary of survey findings.

Back to top

    1. The deep Web: Surfacing hidden value;

    2. Chen-Chuan Chang, K., He, B., and Zhang, Z. Toward large scale integration: Building a metaquerier over databases on the Web. In Proceedings of the 2nd CIDR Conference, 2005.

    3. Fetterly, D., Manasse, M., Najork, M., and Wiener, J. A large-scale study of the evolution of Web pages. In Proceedings of the 12th International World Wide Web Conference, 2004, 669–678.

    4. Ghanem, T.M. and Aref, W.G. Databases deepen the Web. IEEE Computer 73, 1 (2004), 116–117.

    5. GNU. wget;

    6. He, H., Meng, W., Yu, C., and Wu, Z. Wise-integrator: An automatic integrator of Web search interfaces for e-commerce. In Proceedings of the 29th VLDB Conference, 2003.

    7. Lawrence, S. and Giles, C.L. Accessibility of information on the Web. Nature 400, 6740 (1999), 107–109.

    8. O'Neill, E., Lavoie, B., and Bennett, R. Web characterization;

    1Such depth information is obtained by a simple revision of the wget software.

    2However, we noticed that also indexed "site search," which we have excluded; thus, its coverage could be overestimated.

Join the Discussion (0)

Become a Member or Sign In to Post a Comment

The Latest from CACM

Shape the Future of Computing

ACM encourages its members to take a direct hand in shaping the future of the association. There are more ways than ever to get involved.

Get Involved

Communications of the ACM (CACM) is now a fully Open Access publication.

By opening CACM to the world, we hope to increase engagement among the broader computer science community and encourage non-members to discover the rich resources ACM has to offer.

Learn More