Wikipedia:Search engine test
|This page in a nutshell: Measuring is easy. What's hard is knowing what it is you're measuring and what your measurement can mean. Put simply, web searches test the understanding of the WP:Five pillars of Wikipedia.|
A search engine lists web pages on the Internet. This facilitates research by offering an immediate variety of applicable options. Possibly useful items on the results list include the source material or the electronic tools that a web site can provide, such as a dictionary, but the list itself, as a whole, can also indicate important information. However, discerning that information may require insight.
Referencing search engine results is a quick way to either present (what is notable) or delete (what is not verifiable) source material, depending on their reliability. There is a high demand for reliability on Wikipedia. Discerning the reliability of the source material is an especially core skill for using the web, while the wiki itself only facilitates the creation of multiple drafts. As presentations and deletions progress, this variety of choices for input tend to produce the desired objective—a neutral viewpoint. Depending on the type of query and kind of search engine, this variety can open up to a single author.
Some search engine tests
- Popularity – See Google's trending tool below.
- Usage – Identify a term's notability. (See for example Google's ngram tool.)
- Genuineness – Identify a spurious hoax or an urban legend.
- Notability – Decide whether a page should be nominated for deletion.
- Existence – Discover what sources (including websites) actually exist for possible presentation.
- Information – Review the reliability of facts and citations.
- Names and terminology – Identify the names used for things (including alternative names and terminology).
- Copyrighting – Identify whether material is copied, and if so, check the licensing.
This page describes both these web search tests and the web search tools that can help develop Wikipedia, and it describes their biases and their limitations.
The advantages of a specific search engine can be distinguished by using a variety of common search engines. The distinct advantages of each are their user interface and, less obviously, their algorithms for compiling and searching their own indexes. Because a web crawler can be blocked—specific ones or just in general—different search engines can list different web sites, and there are more web sites available by URL than are indexed in any database.
The most common search engines are at Google, Bing, and Yahoo. Specialized search engines exist for medicine, science, news and law amongst others. Several generalized search engines exist. These adapt your query to many search engines. See Common search engines below. This page mostly uses Google instead of Bing or Yahoo, but aims for generality where it can. For example, it describes Google Groups (usenet groups), Google scholar (academia), Google news, and Google books.
Good-faith Googling, a rule of thumb
If an unsourced addition to an article appears plausible, consider taking a moment to use Google (or some other search engine) to find a reliable source before deciding whether to revert.
Search engine tests
Depending on the subject matter, and how carefully it is used, a search engine test can be very effective and helpful, or produce misleading or non-useful results. In most cases, a search engine test is a first-pass heuristic or "rule of thumb".
What a search test can do, and what it can't
A search engine can index pages and text which others have placed on the internet, just like a big index at the back of a book.
Search engines can:
- Provide information and lead to pages that assist with the above goals
- Confirm "who's reported to have said what" according to sources (useful for neutral citing)
- Often provide full cited copies of source documents
- Confirm roughly how popularly referenced an expression is. Note, however, that Google searches may report vastly more hits than will ever be returned to the user, especially for exact quoted expressions. For example, a Google search for "the green goldfish", with quotes, currently initially reports around 22,700 results, yet on paging through to the last search results page shows the returned number of hits to be 370. See also here to calculate statistical significance.
- Search more specifically within certain websites, or for combined and alternative phrases (or excluding certain words and phrases that would otherwise confuse the results).
Search engines cannot:
- Guarantee the results are reliable or "true" (search engines index whatever text people choose to put online, true or false).
- Guarantee why something is mentioned a lot, and that it isn't due to marketing, reposting as an internet meme, spamming, or self-promotion, rather than importance.
- Guarantee that the results reflect the uses you mean, rather than other uses. (E.g., a search for a specific John Smith may pick up many "John Smiths" who aren't the one meant, many pages containing "John" and "Smith" separately, and also miss out all the useful references indexed under "J. Smith" or, if the term is put in quotes, "John Michael Smith" and "Smith, John")
- Guarantee you aren't missing crucial references through choice of search expression.
- Guarantee that little-mentioned or unmentioned items are automatically unimportant.
- Guarantee that a particular result is the original instance of a piece of text and not a reprint, excerpt, quotation, misquotation, or copyright violation.
and search engines often will not:
- Provide the latest research in depth to the same extent as journals and books, for rapidly developing subjects.
- Be neutral.
A search engine test cannot help you avoid the work of interpreting your results and deciding what they really show. Appearance in an index alone is not usually proof of anything.
Search engine tests and Wikipedia policies
Search engine tests may return results that are fictitious, biased, hoaxes or similar. It is important to consider whether the information used derives from reliable sources before using or citing it. Less reliable sources may be unhelpful, or need their status and basis clarified, so that other readers gain a neutral and informed understanding to judge how reliable the sources are.
Google (and other search systems) do not aim for a neutral point of view. Wikipedia does. Google indexes self-created pages and media pages which do not have a neutrality policy. Wikipedia has a neutrality policy that is mandatory and applies to all articles, and all article-related editorial activity.
As such, Google is specifically not a source of neutral titles – only of popular ones. Neutrality is mandatory on Wikipedia (including deciding what things are called) even if not elsewhere, and specifically, neutrality trumps popularity.
- (See WP:NPOV#Neutrality and Verifiability for information on balancing the policies on verifiability and neutrality, and WP:NPOV#Article naming on how articles should be named)
Raw "hit" (search result) count is a very crude measure of importance. Some unimportant subjects have many "hits", some notable ones have few or none, for reasons discussed further down this page.
Hit count numbers alone can only rarely "prove" anything about notability, without further discussion of the type of hits, what's been searched for, how it was searched, and what interpretation to give the results. On the other hand, examining the types of hit arising (or their lack) often does provide useful information related to notability.
Additionally, search engines do not disambiguate, and tend to match partial searches. (However, as described below, you can eliminate partial matches by quoting the phrase to be matched): While Madonna of the Rocks is certainly an encyclopedic and notable entry, it's not a pop culture icon. However, due to Madonna matching as a partial match, as well as other Madonna references not related to the painting, the results of a Google or Bing search result count will be disproportionate as compared to any equally notable Renaissance painting. To exclude partial matches when Googling for the phrase, quote the phrase to be matched as follows: "Madonna of the Rocks".
Using search engines
Search engine expressions (examples and tutorial)
- This section explains some search expressions used in Google web search. Similar approaches will work in many other search engines, and other Google searches, but always read their help pages for further information as search engines' capabilities and operation often differ. Note that if you are signed in to a Google account when searching on Google then this may affect the results that you get, based on your search history. Also be sure to check "Languages for Displaying (Search) Results" in "Search Settings".)
The single most useful search engine tool may be the use of quotation marks to find an exact match for a phrase. However, a search engine such as Google has both an easy, and an advanced search with further search options. The advanced search makes it easier to enter advanced options, that may help your searching. The following collapsible sections cover basic examples and help for using search engines with Wikipedia.
Specialized search engines such as medical paper archives have their own specialized search structure not covered here.
Specific uses of search engines in Wikipedia
- Google Books has a pattern of coverage that is in closer accord with traditional encyclopedia content than is the Web, taken as a whole; if it has systemic bias, it is a very different systemic bias from Google Web searches. Multiple hits on an exact phrase in Google Book search provide convincing evidence for the real use of the phrase or concept. You can compare usage of terms, such as "Tidal wave" vs. "Tsunami". Google Book search can locate print-published testimony to the importance of a person, event, or concept. It can also be used to replace an unsourced "common knowledge" fact with a print-sourced version of the same fact.
- Google News can help assess whether something is newsworthy. Google News used to be less susceptible to manipulation by self-promoters, but with the advent of pseudo-news sites designed to collect ad revenues or to promote specific agendas, this test is often no more reliable than others in areas of popular interest, and indexes many "news" sources that reflect specific points of view. The news archive goes back many years but may not be free beyond a limited period. News results often include press releases, which are not neutral, independent sources.
- Topics alleged to be notable by popular reference can have the type of reference, and popularity, checked. An alleged notable issue that only has a few hundred references on the Internet may not be very notable; truly popular Internet memes can have millions or even tens of millions of references. However note that in some areas, a notable subject may have very few references; for example, one might only expect a handful of references to some archaeological matter, and some matters will not be reflected online at all.
- Topics alleged to be genuine can be checked to test if they are referenced by reliable independent sources; this is a good test for hoaxes and the like.
- Copyright violations from websites can often be identified (as described above).
- Alternative spellings and usages can have their relative frequencies checked (eg, for a debate which is the more common of two equally neutral and acceptable terms). Google Trends can compare usage in the "News" category ("Tidal wave" vs "Tsunami" example), but this may not be reliable for older news.
- Google Groups (USENET newsgroups) is a significantly different sample from websites, and represents, for the most part, conversations in English conducted by people on various topics. Because the sources are very different, hit numbers are not comparable, however Group searches are particularly helpful in identifying matters which might be discussed, or whose presence may have been artificially inflated by promotional techniques; it is suspicious if a phrase gets, say, 100,000 Web hits but only 10 Groups hits.
A raw hit count should never be relied upon to prove notability. Attention should instead be paid to what (the books, news articles, scholarly articles, and web pages) is found, and whether they actually do demonstrate notability or non-notability, case by case. Hit counts have always been, and very likely always will remain, an extremely erroneous tool for measuring notability, and should not be considered either definitive or conclusive. A manageable sample of results found should be opened individually and read, to actually verify their relevance.
In the case of Google (and other search engines such as Bing and Yahoo!), the hit count at the top of the page is unreliable and should usually not be reported. The hit count reported on the penultimate (second-to-last) page of results may be slightly more accurate. For searches with few reported hits (less than 1000) the actual count of hits needed to reach the bottom of the last page of results may be more accurate, but even this is not a sure thing. Google returns different search results depending on factors such as your previous search history and on which Google server you happen to hit.
Other useful considerations in interpreting results are:
- Article scope: If narrow, fewer references are required. Try to categorize the point of view, whether it is NPOV, or other; e.g., notice the difference between Ontology and Ontology (computer science).
- Article subject: If it's about some historical person, one or two mentions in reliable texts might be enough; if it's some Internet neologism or a pop song, it may be on 700 pages and might still not be considered 'existing' enough to show any notability, for Wikipedia's purposes.
Biases to be aware of
In most cases, search results should be reviewed with an awareness and careful skepticism before relying upon them. Common biases include:
- General (the Internet or people as a whole)
- Personal bias – Tendency to be slightly more receptive to beliefs that one is familiar with, agrees with, or are common in one's daily culture, and also to be more doubtful about beliefs and views that contradict one's preferred views.
- Cultural and computer-usage bias – Biased towards information from Internet-using developed countries and affluent parts of society (internet access). Countries where computer use is not so common will often have lower rates of reference to equally notable material, which may therefore appear (mistakenly) non-notable.
- Undue weight – May disproportionally represent some matters, especially related to popular culture (some matters may be given far more space and others far less, than fairly represents their standing): popularity is not notability.
- Sources not readily accessible – Some sources are accessible to all, but many are payment only, or not reported online.
- General web search engines (Google, Bing web search etc.)
- Dark net – Search engines exclude a vast number of pages, and this may include systematic bias so that some matters are excluded disproportionately (for example, because they are commonly visible on sites that do not allow Google indexing, or the content for technical reasons cannot be indexed (Flash- or image-based websites etc)
- Search engines as promotion tool – An industry exists seeking to influence site position, popularity, and ratings in such searches, or sell advertising space related to searches and search positions. Some subjects, such as pornographic actors, are so dominated by these that searches cannot be reliably used to establish popularity.
- Review process varies; some sites accept any information, while others have some form of review or checking system in place.
- Self-mirroring – Sometimes other sites clone Wikipedia content, which is then passed around the Internet, and more pages built up based upon it (and often not cited), meaning that in reality the source of much of the search engine's findings are actually just copies of Wikipedia's own previous text, not genuine sources.
- Popular usage bias – Popular usage and urban legend is often reported over correctness
- A search for the incorrect Charles Windsor gives 10 times more results than the correct Charles Mountbatten-Windsor.
- A search for the most common spelling of El Niño will often report it spelt "El Nino", without the diacritic.
- Urban legends are often reported widely, for example hundreds of sites report that the USS Constitution set sail in 1779, although the correct date is 1797.
- Popular views and perceptions are likely to be more reported. For example, there may be many references to acupuncture and confirming that people are often allergic to animal fur, but it may only be with careful research that it is revealed there are medical peer-reviewed assessments of the former, and that people are usually not allergic to fur, but to the sticky skin particles ("dander") within the fur.
- Language selection bias For example, an Arabic speaker searching for information on homosexuality in Arabic, will likely find pages which reflect a different bias than an English speaker searching in English on the same subject, since popular and media views and beliefs about homosexuality can differ widely between English speaking countries (USA, UK, Australasia) that tend to include a higher proportion of homosexuality-accepting groups, and Arabic speaking countries (Middle East) that tend to include a lower proportion.
- Note that other Google searches, particularly Google Book Search, have a different systemic bias from Google Web searches and give an interesting cross-check and a somewhat independent view.
In some cases, it is helpful to estimate the relative popularity of a website. Alexa Internet is a tool for this (Hitwise and Quantcast are others). To test Alexa's ranking for a particular web site, visit alexa.com and enter the URL.
The Alexa measuring system is based on a toolbar that users must choose to install, which can be installed on several browsers including Internet Explorer and Mozilla Firefox, across different operating systems. Sources of bias include both websites whose users disproportionately do not install such toolbars, as well as webmasters who install Alexa Toolbar for the sole purpose of enhancing their ratings. Specifically, Alexa rankings are not part of the notability guidelines for web sites for several reasons:
- Below a certain level, Alexa rankings are essentially meaningless, because of the limited sample size. Alexa itself says that ranks lower than 100,000 are not reliable.
- Alexa rankings vary and include significant systematic bias which means the ratings often do not reflect popularity, but only popularity amongst certain groups of users (See Alexa Internet#Concerns). Broadly, Alexa rates based upon measurements by a user-installed toolbar, but this is a highly variable tool, and there are large parts of the Internet user community (especially corporate users, many advanced users, many open-source and non-Windows users) who do not use it and whose Internet reference use is therefore ignored.
- Alexa rankings do not reflect encyclopedic notability and existence of reliable source material if so. A highly ranked web site may well have nothing written about it, or a poorly ranked web site may well have a lot written about it.
- A number of unquestionably notable topics have web sites with poor Alexa rankings.
- To obtain statistics, visit http://quantcast.com, enter url, click "Search."
- For entities which subscribe to Quantcast's service, Quantcast declares that their traffic measurements are "verified." This may provide better reliability than Alexa results, as it does not depend on user installation of a plugin.
- For entities which do not subscribe to be "quantified", Quantcast declares their traffic measurements to be "estimates."
- The same reliability and notability provisions listed under #Alexa ratings apply here.
Foreign languages, non-Latin scripts, and old names
Often for items of non-English origin, or in non-Latin scripts, a considerably larger number of hits result from searching in the correct script or for various transcriptions—be sure to check "Languages for Displaying (Search) Results" in "Search Settings". An Arabic name, for instance, needs to be searched for in the original script, which is easily done with Google (provided one knows what to search for), but problems may arise if – for example – English, French and German webpages transcribe the name using different conventions. Even for English-only webpages there may be many variants of the same Arabic or Russian name. Personal names in other languages (Russian, Anglo-Saxon) may have to be searched for both including and excluding the patronymic, and searches for names and other words in strongly inflected languages should take into account that arriving at the total number of hits may require searching for forms with varying case-endings or other grammatical variations not obvious for someone who does not know the language. Names from many cultures are traditionally given together with titles that are considered part of the name, but may also be omitted (as in Gazi Mustafa Kemal Pasha).
Even in Old English, the spelling and rendering of older names may allow dozens of variations for the same person. A simplistic search for one particular variant may underrepresent the web presence by an order of magnitude.
A search like this requires a certain linguistic competence which not every individual Wikipedian possesses, but the Wikipedia community as a whole includes many bilingual and multilingual people and it is important for nominators and voters on AfD at least to be aware of their own limitations and not make untoward assumptions when language or transcription bias may be a factor.
Google distinct page count issues
Note also, that the number of search string matches reported by search engines is only an estimate. For example, Google will only calculate the actual number of matches once the user navigates through all result pages, to the last one, and even then it places restrictions on the figure. At times, the "match" count estimate can be significantly different (by one or more orders of magnitude) to the total count of results shown on the last results page.
A site-specific search may help determine if most of the matches are coming from the same web site; a single web site can account for hundreds of thousands of hits.
For search terms that return many results, Google uses a process that eliminates results which are "very similar" to other results listed, both by disregarding pages with substantially similar content and by limiting the number of pages that can be returned from any given domain. For example, a search on "Taco Bell" will give only a couple of pages from tacobell.com even though many in that domain will certainly match. Further, Google's list of distinct results is constructed by first selecting the top 1000 results and then eliminating duplicates without replacements. Hence the list of distinct results will always contain fewer than 1000 results regardless of how many webpages actually matched the search terms. For example, from the about 742 million pages related to "Microsoft", Google presently returns 572 "distinct" results (as of 14 December 2010[update]). Caution must be used in judging the relative importance of websites yielding well over 1000 search results.
Search engine limitations – technical notes
Many, probably most, of the publicly available web pages in existence are not indexed. Each search engine captures a different percentage of the total. Nobody can tell exactly what portion is captured.
The estimated size of the World Wide Web is at least 11.5 billion pages, but a much deeper (and larger) Web, estimated at over 3 trillion pages, exists within databases whose contents the search engines do not index. These dynamic web pages are formatted by a Web server when a user requests them and as such cannot be indexed by conventional search engines. The United States Patent and Trademark Office website is an example; although a search engine can find its main page, one can only search its database of individual patents by entering queries into the site itself.
Google, like all Internet search engines can only find information that has actually been made available on the Internet. There is still a sizable amount of information that is not on the Internet.
Google, as all search engines should, follows the robots.txt protocol and can be blocked by sites that do not wish their content to be indexed or cached by Google. Sites that contain large amounts of copyrighted content (Image galleries, subscription newspapers, webcomics, movies, video, help desks), usually involving membership, will block Google and other search engines. Other sites may also block Google due to the stress or bandwidth concerns on the server hosting the content.
Search engines also might not be able to read links or metadata that normally requires a browser plugin, Adobe PDF, or Macromedia Flash, or where a website is displayed as part of an image. Search engines also can not listen to podcasts or other audio streams, or even video mentioning a search term. Similarly search engines cannot read PDF files consisting of photoscans or look inside compressed (.zip) files.
Google and other popular search engines are also a target for search engine "search result enhancement", also known as search engine optimizers, so there may also be many results returned that lead to a page that only serves as an advertisement. Sometimes pages contain hundreds of keywords designed specifically to attract search engine users to that page, but in fact serve an advertisement instead of a page with content related to the keyword.
Hit counts reported by Google are only estimates, which in some cases have been shown to necessarily be off by nearly an order of magnitude, especially for hit counts above a few thousands. For such common words as to yield several thousand Google hits, freely available text corpora such as the British National Corpus (for British English) and the Corpus of Contemporary American English (for American English) can provide a more accurate estimate of the relative frequencies of two words.
Example of the limitations
The Economic Crime Summit site is a rather Google- and Internet Archive-unfriendly site. It is very graphics heavy, providing Google with little to nothing to look for and many missing pages in the Internet Archive version. So while you can bring up the 2002 Economic Crime Summit Conference, the overview link that would tell you who presented what does not work. The 2004 Economic Crime Summit Conference archive is even worse as that was in three places and none of the archived links tells you anything about the papers presented.
Via Internet Archive you have proof that some information regarding "Impact of Advances in Computer Technology in Evidence Processing" existed on the Internet. Yet today Google cannot find that information! A program known to be part of the 2002 Economic Crime Summit Conference and at one time was listed on a website on the Internet currently cannot be found by Google.
Common search engines
The most common search engines are Google, Bing, and Yahoo, but the most useful search engine, which depend on a context, may not be the most common ones.
Type Examples General search engines Google, Bing, Yahoo! etc. Website popularity indexes Alexa, Hitwise General information About.com Professional research indexes Medline (medical), science, law, Google Scholar News and media Google News archives search Historical archives of web pages Archive.org, Web cache (how web pages looked and their contents, at different times or if deleted) Books and historical literature Project Gutenberg, Google Books, Amazon.com and a9.com (for book info) Universities and higher education organisations 4icu.org (University websites search engine)
Google groups archives Usenet. Because it covers over twenty years, it one of the oldest archives on record, going back to the beginning of the web.
Specialized search engines
Google Scholar works well for fields that are paper-oriented and have an online presence in all (or nearly all) respected venues. This search engine is a good complement for the commercially available Thompson ISI Web of Knowledge, especially in the areas which are not well covered in the latter, including books, conference papers, non-American journals, the general journals in the field of strategy, management, international business, English language education and educational technology. The analysis of the PageRank algorithm utilised by Google Scholar demonstrated that this search engine, as well as its commercial analogs, provides an adequate information about popularity of some concrete source, although that does not automatically reflect the real scientific contribution of concrete publication.
Medline, now part of Pubmed, is the original broadly based search engine, originating over four decades ago and indexing even earlier papers. Thus, especially in biology and medicine, Pubmed "associated articles" is a Google Scholar proxy for older papers with no on-line presence. E.g., The journal Stroke puts papers on-line back through 1970s. For this 1978 paper , Google Scholar lists 100 citing articles, while Pubmed lists 89 associated articles
There are a large number of law libraries online, in many countries, including: Library of Congress, Library of Congress (THOMAS), Indiana Supreme Court, FindLaw (USA); Kent University Law Library and sources (UK).
See also this list of search engines.
Generalized search engines
Several generalized search engines exist. These adapt your query to many search engines. Web browsers offer a choice of search engines to choose to employ for the search box, and these can be used one at a time to experiment with search results. Meta-search engines use several search engines at once. Ten popular ones from About.com offer reviews. A web browser plugin can add a search engine or a meta-search engine to your list of choices.
- For example, if there are 16 hits at Google Books under one name, and 24 under another, there is only a 70% confidence that the second name is actually more common.
- Google Search Operators and more search help
- Search history personalization
- Google Search Settings
- Avoid inauthor:"Books, LLC", as LLC 'publishes' raw printouts of Wikipedia articles.
- Google search for: AYB OR AYBABTU OR "All your base"
- Google Answers question on word frequency in news sources
- Takuya, Funahashi; Hayato, Yamana (2010). "Reliability Verification of Search Engines' Hit Counts" (PDF). Proceedings of the 10th international conference on Current trends in web engineering. Computer Science and Engineering Division, Waseda University. Retrieved 5 May 2015.
- Sullivan, Danny (21 October 2010). "Why Google Can't Count Results Properly". SearchEngineLand.com. Retrieved 5 May 2015.
- Google search for "Microsoft"
- Antonio Gulli and Alessio Signorini (2005-08-29). "The Indexable Web is more than 11.5 billion pages".
- Alvin More and Brian H. Murray (2000). "Sizing the Internet" (PDF). Cyveillance Inc.
- Mark Liberman (2009), "Quotes with and without quotes", Language Log.
- Mark Liberman (2005), "Questioning reality", Language Log, and other Language Log posts linked from there.
- Harzing, A.W.K. & Wal, R. van der (2008). Google Scholar as a new source for citation analysis? Ethics in Science and Environmental Politics, vol. 8, no. 1, pp. 62-71
- Jan van Aalst. (2010) Using Google Scholar to Estimate the Impact of Journal Articles in Education. Educational Researcher 39: 387.
- Maslov, S., & Redner, S. (2008). Promise and pitfalls of extending Google’s PageRank algorithm to citation networks. Journal of Neuroscience, 28, 11103–11105
- Joe Meert (2006-04-30). "Argumentum ad Googlum". Science, AntiScience and Geology.—Meert observes that "The temptation to find a quick retort means that, many times, people don't bother to check the source carefully." and that "people will look for a specific phrase that may be taken out-of-context to support their argument". He states that it is "dangerous and irresponsible to think that we can Google away a complex discussion" and that he has "learned long ago that there is no substitute for detailed research on a topic".
- Rich Turner (2004-02-29). "Argumentum ad Googlum; Why Getting a Million Hits on Google Doesn't Prove Anything". Grumbles.—Turner points out that "that something gets hits on Google does not make it correct" and gives several examples of things that are incorrect that garner thousands of hits on Google search results.
- Thelwall, M. (2008). Quantitative comparisons of search engine results, Journal of the American Society for Information Science and Technology, 59(11), 1702–1710. http://www.scit.wlv.ac.uk/~cm1993/papers/SearchEngineComparisons_preprint.doc
- Thelwall, M. (2008). Extracting accurate and complete results from search engines: Case study Windows Live. Journal of the American Society for Information Science and Technology, 59(1), 38–50. http://www.scit.wlv.ac.uk/~cm1993/papers/2007_Accurate_Complete_preprint.doc
- Gomes, et al. (2000). Detecting query-specific duplicate documents. http://patft.uspto.gov/netacgi/nph-Parser?Sect1=PTO2&Sect2=HITOFF&u=%2Fnetahtml%2FPTO%2Fsearch-adv.htm&r=1&p=1&f=G&l=50&d=PTXT&S1=6615209.PN.&OS=pn/6615209&RS=PN/6615209
- Thelwall, M. (2008). Quantitative comparisons of search engine results, Journal of the American Society for Information Science and Technology, 59(11), 1702–1710. http://www.scit.wlv.ac.uk/~cm1993/papers/SearchEngineComparisons_preprint.doc
- Nakov, Preslav and Hearst, Marti (2005). A Study of Using Search Engine Page Hits as a Proxy for n-gram Frequencies, Proceedings of Recent Advances in Natural Language Processing 2005 http://biotext.berkeley.edu/papers/nakov_ranlp2005.pdf
- Baroni, Marco and Ueyama, Motoko (2006) Building general- and special-purpose corpora by Web crawling, Proceedings of the 13th NIJL International Symposium Language Corpora Their Compilation and Application. http://tokuteicorpus.jp./result/pdf/2006_004.pdf