Determining Bias to Search Engines from Robots.txt

Determining Bias to Search Engines from Robots.txt
Yang Sun, Ziming Zhuang, Isaac G. Councill, and C. Lee Giles
Information Sciences and Technology
The Pennsylvania State University
University Park, PA 16801, USA
{ysun, zzhuang, icouncill, giles}@ist.psu.edu
Abstract
Search engines largely rely on robots (i.e., crawlers or
spiders) to collect information from the Web. Such crawling
activities can be regulated from the server side by deploy-
ing the Robots Exclusion Protocol in a file called robots.txt.
Ethical robots will follow the rules specified in robots.txt.
Websites can explicitly specify an access preference for each
robot by name. Such biases may lead to a “rich get richer”
situation, in which a few popular search engines ultimately
dominate the Web because they have preferred access to re-
sources that are inaccessible to others. This issue is seldom
addressed, although the robots.txt convention has become a
de facto standard for robot regulation and search engines
have become an indispensable tool for information access.
We propose a metric to evaluate the degree of bias to which
specific robots are subjected. We have investigated 7,593
websites covering education, government, news, and busi-
ness domains, and collected 2,925 distinct robots.txt files.
Results of content and statistical analysis of the data con-
firm that the robots of popular search engines and informa-
tion portals, such as Google, Yahoo, and MSN, are gener-
ally favored by most of the websites we have sampled. The
results also show a strong correlation between the search
engine market share and the bias toward particular search
engine robots.
1 Introduction
1
Without robots, there would probably be no search en-
gines. Web search engines, digital libraries, and many other
web applications such as offline browsers, internet market-
ing software and intelligent searching agents heavily de-
pend on robots to acquire documents. Robots, also called
“spiders”, “crawlers”, “bots” or “harvesters”, are self-acting
agents that navigate around-the-clock through the hyper-
links of the Web, harvesting topical resources at zero costs
1A greatly abbreviated version of this paper appeared as a poster in the
Proceedings of the 16th International World Wide Web Conference, 2007.
of human management [3, 4, 14]. Because of the highly au-
tomated nature of the robots, rules must be made to regulate
such crawling activities in order to prevent undesired impact
to the server workload or access to non-public information.
The Robots Exclusion Protocol has been proposed [12]
to provide advisory regulations for robots to follow. A file
called robots.txt, which contains robot access policies, is
deployed at the root directory of a website and accessible
to all robots. Ethical robots read this file and obey the rules
during their visit to the website. The robots.txt convention
has been adopted by the community since the late 1990s,
and has continued to serve as one of the predominant means
of robot regulation. However, despite the criticality of the
robots.txt convention for both content providers and har-
vesters, little work has been done to investigate its usage
in detail, especially at the scale of the Web.
More importantly, as websites may favor or disfavor cer-
tain robots by assigning to them different access policies,
this bias can lead to a “rich get richer” situation whereby
some popular search engines are granted exclusive access to
certain resources, which in turn could make them even more
popular. Considering the fact that users often prefer a search
engine with broad (if not exhaustive) information coverage,
this “rich get richer” phenomenon may introduce a strong
influence on users’ choice of search engines, which will
eventually be reflected in the search engine market share.
On the other hand, since it is often believed (although this
is an exaggeration) that “what is not searchable does not
exist,” this phenomenon may also introduce a biased view
of the information on the Web.
1.1 Related Work and Contributions
A 1999 study of the usage of robots.txt[10] in UK uni-
versities and colleges investigated 163 websites and 53
robots.txt. Robots.txt files were examined in terms of file
size and the use of Robots Exclusion Protocol within the
UK university domains. In 2002, Drott [7] studied the us-
age of robots.txt as an aid for indexing to protect informa-
tion. 60 samples from Fortune Global 500 company web-

sites were manually examined in this work concluding that
“robots.txt files are not widely used by the sampled group
and for most of the sites on which they appear, they are re-
dundant. ...they exclude robots from directories which are
locked anyway.” Our investigation shows a contrary result
which may be due to the difference in sample size, domain
and time. Other work addresses the legal aspects of obey-
ing robots.txt [2, 8] and an overview of Web robots and
robots.txt usage is given in [5].
None of the aforementioned work investigates the con-
tent of robots.txt in terms of biases towards different robots.
In addition, the sample sizes of previous studies have tended
to be relatively small considering the size of the Web. In this
paper, we present the first quantitative study of such biases,
and conduct a more comprehensive survey of robots.txt us-
age on the Web. By implementing our own specialized
“robots.txt” crawler, we collect real-world data from a con-
siderable amount of unique websites with different func-
tionalities, covering the domains of education, government,
news, and business. We investigate the following questions:
• Does a robot bias exist?
• How should such a bias be measured quantitatively?
• What is the implication for such a bias?
Our contributions are:
• We propose a quantitative metric to automatically mea-
sure robot biases.
• By applying the metric to a large sample of websites,
we present our findings about the most favored and dis-
favored robots.
The rest of the paper is organized as follows. In Section
2 we briefly introduce the Robots Exclusion Protocol. In
Section 3 we present our data collection for this study. In
Section 4, We propose a bias metric and demonstrate how it
is applied to measure the degree of robot bias. In Section 5
we present our observations on robots.txt usage and discuss
the implications. In Section 6 we conclude our paper with
plans for future work.
2 Robots Exclusion Protocol
The Robots Exclusion Protocol2
is a convention that al-
lows website administrators to indicate to visiting robots
which parts of their site should not be visited. If there is
no robots.txt file on a website, robots are free to crawl all
content.
The format of Robots Exclusion Protocol is described
in [12]. A file named “robots.txt” with Internet Media
2http://www.robotstxt.org/wc/norobots.html
Type “text/plain” is placed under the root directory of a
Web server. Each line in the robots.txt file has the for-
mat: < field >:< optionalspace >< value ><
optionalspace >. There are three types of case-insensitive
tags for the < field > to specify the rules: User-Agent,
Allow and Disallow. Another unofficial directive Crawl-
Delay is also used by many websites to limit the frequency
of robot visits.
The robots.txt file starts with one or more User−Agent
fields, specifying which robots the rules apply to, followed
by a number of Disallow : and/or Allow : fields indicat-
ing the actual rules to regulate the robot. Comments are
allowed anywhere in the file, and consist of optional whites-
paces. Comments are started with a comment character ‘#’
and terminated by the linkbreak.
A sample robots.txt is listed below (this robots.txt file is
from BotSeer3
):
User-Agent: *
Disallow: /robots/
Disallow: /src/
Disallow: /botseer
Disallow: /uastring
Disallow: /srcseer
Disallow: /robotstxtanalysis
Disallow: /whois
User-Agent: googlebot
Disallow: /robotstxtanalysis
Disallow: /uastring
User-Agent: botseer
Disallow:
It shows that Googlebot cannot visit “/robotstxtanaly-
sis” and “/uastring”. BotSeer can visit any directory and
file on the server. All the other robots should follow the
rules under User − Agent : ∗ and cannot visit the direc-
tories and files matching “/robots/”, “/src/”, “/botseer”,
“/uastring”, “/srcseer”, “/robotstxtanalysis”, “/whois”.
3 Robot Bias
We propose ∆P(r), a measure of the favoribility of
robots across a sample of robots.txt files, to measure the
degree to which specific robots are favored (or disfavored)
by a set of websites. A formal definition of the robot bias
(favored or disfavored) is described below.
3.1 The GetBias Algorithm
Our definition of a favored robot is a robot allowed to ac-
cess more directories than the universal robot according to
3http://botseer.ist.psu.edu/robots.txt

the robots.txt file in the website. The universal robot is any
robot that has not matched any of the specific User-Agent
names in the robots.txt file. In other words, the universal
robot represents all the robots that do not appear by name in
the robots.txt file.
Let F be the set of robots.txt files in our dataset. Given
a robots.txt file f ∈ F, let R denote the set of named robots
for a given robots.txt file f. For each named robot r ∈ R,
We define the GetBias(r, f) algorithm as specified in Al-
gorithm 1. GetBias measures the degree to which a named
robot r is favored or disfavored in a given robots.txt file f.
Algorithm 1 GetBias(r, f)
1: if r is ∗ then
2: return 0
3: end if
4: Construct DIR for f;
5: bias = 0
6: for all d ∈ DIR do
7: if d is allowed for ∗ then
8: Du ← d
9: end if
10: end for
11: for all d ∈ DIR do
12: if d is allowed for r then
13: Dr ← d
14: end if
15: end for
16: bias = |Dr| − |Du|
17: return bias
Let DIR be the set of all directories that appear in a
robots.txt file f of a specific website. DIR is used as
an estimation of the actual directory structure in the web-
site because the Robot Exclusion Protocol considers any
directory in the website that does not match the directo-
ries in the robots.txt as an allowed directory by default.
Du ∈ DIR is the set of directories that the universal robot
“*” is allowed to visit. If there are no rules specified for
User − Agent : ∗, the universal robot can access every-
thing by default. Dr ∈ DIR is the set of directories that
a given robot r is allowed to visit. |Du| and |Dr| are the
number of directories in Du and Dr.
For a given robot r, the algorithm first counts how many
directories in DIR is allowed for r. Then it calculates the
bias score for robot r as the difference between the number
of directories in DIR that are allowed for the robot r and
the number of directories that are allowed for the universal
robot. In the GetBias algorithm, the bias of the universal
robot is treated as the reference point 0 (GetBias returns 0).
The bias scores of favored robots returned by GetBias are
positive values. Higher score of a robot means the robot is
more favored. On the contrary, the bias scores of disfavored
robots returned by GetBias are negative values, which is
consistent with our bias definition. Thus, the bias of a robot
in a robots.txt file can be represented by a categorical vari-
able with three categories: favored, disfavored, and no bias.
As an example, consider the robots.txt file in
http://BotSeer.ist.psu.edu from Section 2: DIR =
{“/robots/”, “/src/”, “/botseer”, “/uastring”, “/srcseer”,
“/robotstxtanalysis”, “/whois”}. According to the algo-
rithm we have Du = {null}, Dbotseer = {“/robots/”, “/src/”,
“/botseer”, “/uastring”, “/srcseer”, “/robotstxtanalysis”,
“/whois”} and Dgoogle = {“/robots/”, “/src/”, “/botseer”,
“/srcseer”, “/whois”}. Thus, |Du| = 0, |Dbotseer|=7,
and |Dgoogle|=5. According to Algorithm 1, biasu =
|Du| − |Du| = 0, biasbotseer = |Dbotseer| − |Du| = 7
and biasgooglebot = |Dgoogle| − |Du| = 5. Thus, the robots
“googlebot” and “botseer” are favored by this website, and
they are categorized as favored. All other robots will be
categorized as no bias.
3.2 Measuring Overall Bias
Based on the bias score for each file, we propose ∆P(r)
favorability in order to evaluate the degree to which a spe-
cific robot is favored or disfavored on a set of robots.txt files.
Let N = |F| be the total number of robots.txt files in the
dataset. The ∆P(r) favorability of a robot r can be defined
as below:
∆P(r) = Pfavor(r) − Pdisfavor(r)
=
Nfavor(r) − Ndisfavor(r)
N
. (1)
where Nfavor(r) and Ndisfavor(r) are the number
of times a robot is favored and disfavored respectively.
Pfavor(r) is the proportion of the robots.txt files in which
a robot r is favored; Pdisfavor(r) is the proportion of the
robots.txt files in which a robot r is disfavored.
The proportions of robots.txt files that favor or disfavor
a specific robot are simple measures for survey statistics;
however, in our dataset the two proportions in isolation are
not very accurate in reflecting the overall biases in our sam-
ple since there are more than two events (favor, disfavor and
no bias). This means that Pfavor(r) + Pdisfavor(r) < 1.
Each event only reflects one aspect of the bias. For exam-
ple, a robot named “ia archiver” is favored by 0.24% of the
websites in our dataset and the proportion of sites that favor
“momspider” is 0.21%. Alternatively, the proportions of
sites that disfavor “ia archiver” and “momspider” are 1.9%
and 0%, respectively. If we only consider the favored pro-
portion, we will reach the conclusion that “ia archiver” is
more favored than “momspider”.
∆P(r) is the difference of the proportions of sites that
favor and disfavor a specific robot, and thus treats both cases

in unison. For the above example ∆P(ia archiver) is -
1.66% and ∆P(momspider) is 0.21%. Thus, “momspi-
der” is more favored than “ia archiver”. For any no-bias
robot r, ∆P(r) is 0. The bias measure can eliminate the
misleading cases and still be intuitively understandable (fa-
vored robots have positive numbers and disfavored robots
have negative numbers).
3.3 Examining Favorability
The favorability is actually a ranking function of robots.
To evaluate accuracy of this ranking function, we run a
ranking performance test based on Kendall’s rank correla-
tion method [11]. The rank correlation method is briefly
described below. The details of the ranking performance
evaluation using partial order can be found in [9].
For a robots.txt file f, let ma be a bias measure function
for all robots R appearing in f. Let ri and rj ∈ R be two
named robots in f. We denote ri <ma rj if ri is ranked
higher than rj for measure ma. Thus, for any two measure
functions ma and mb, Kendall’s τ can be defined based on
the number Pf of concordant pairs and the number Qf of
discordant pairs. A pair ri = rj is concordant if both ma
and mb agree in how they order ri and rj. It is discordant if
they disagree. In this case, Kendall’s τ can be defined as:
τf (ma, mb) =
Pf − Qf
Pf + Qf
(2)
For any given measure ma and mb, the τf (ma, mb) rep-
resents how well the two ranking measures agree with each
other in a file f. Let ma represent the actual ranking func-
tion of robots. Although we do not know the actual rank-
ing function, we have the partial ranking of robots for each
robots.txt file based on the bias score defined previously.
Thus, computing the τf (ma, mb) for all robots.txt files will
show how well the measure mb agrees with ma for the ac-
tual ranking of robots in file f.
We calculate τf (ma, mb) for each robots.txt files f in
our dataset. If τf (ma, mb) = 1 for a given robots.txt file,
we consider that the file f is a concordant file for ma and
mb. Otherwise, the file f is a discordant file. By counting
the concordant files P and discordant files Q in the dataset,
we can compute the average τ(ma, mb). Note that P +Q =
N, thus,
τ(ma, mb) =
P − Q
P + Q
= 1 −
2Q
N
. (3)
We rank the robots using the δP favorability. The ranked
lists are then compared with the actual ranking using the
method introduced above. The average τ value is 0.957
which we believe is accurate enough to measure the overall
bias of a robot.
4 Data Collection
To observe the potential robot bias on the Web, our work
studies a wide range of websites with different domains and
from different physical locations. The data collection is de-
scribed below in detail.
4.1 Data Sources
The Open Directory Project [6] is the largest and most
comprehensive human-maintained Web directory. Our pri-
mary source to collect the initial URLs to feed our crawler
is DMOZ because the Open Directory Project classifies a
large URL collection into different categories. It enables
us to collect data from different domains and physical loca-
tions. Our collection from the Open Directory Project cov-
ers three domains: education, news, and government. The
university domain is further broken down into the Ameri-
can, European, and Asian university domains.
Since the directory structure of the business domain in
DMOZ is complicated and has a significant amount of over-
laps, we use the 2005 Fortune Top 1000 Company List [1]
as our data source.
There are certain limitations inherent in our data collec-
tion. First, because the website collection in DMOZ is lim-
ited for other countries especially for non-English websites,
the majority of the websites are from the USA. Second, be-
cause the DMOZ entries are organized by human editors,
there might be errors. Finally, we collect business websites
from the Fortune 1000 list which contains data mostly of
large corporations, so that the data in that domain may not
be representative of small businesses. We intend to address
these limitations in future research.
4.2 Crawling for Robots.txt
We have implemented a specialized focused crawler for
this study. The crawler starts by crawling the metadata of a
website obtained from DMOZ including the functional clas-
sification, the name of the website, and the physical location
of its affiliated organization. Then, the crawler checks the
existence of robots.txt for that domain and downloads exist-
ing robots.txt files for offline analysis. A parsing and filter-
ing module is also integrated into our crawler to eliminate
duplicates and for ensuring that retrieved pages are within
the target domain.
Besides the root level directory, our crawler also exam-
ines other possible locations of the robots.txt file. The sub-
directories of a website (up to level 3) are inspected. Re-
sults show that there are few cases where robots.txt is not
placed under the root directory where it should be accord-
ing to the Robots Exclusion Protocol. Misspelled filenames
are also examined by our crawler. In rare cases the filename
“robot.txt” (which will be ignored by robots) is used instead
of “robots.txt”.

In order to observe the temporal properties, the crawler
has performed 5 crawls for the same set of websites from
Dec. 2005 to Oct. 2006. In order to analyze the tempo-
ral properties, the downloaded robots.txt files are archived
according to the date of the crawl.
4.3 Statistics
We crawled and investigated 7,593 unique websites in-
cluding 600 government websites, 2047 newspaper web-
sites, 1487 USA university websites, 1420 European uni-
versity websites, 1039 Asian university websites, and 1000
company websites. The number of websites that have
robots.txt files in each domain from the 5 crawls are shown
in Table 1.
Websites
Collected robots.txt files
Dec.
2005
Jan.
2006
May.
2006
Sep.
2006
Oct.
2006
Government 600 248 257 263 262 264
Newspaper 2047 859 868 876 937 942
USA Univ. 1487 615 634 650 678 683
European
Univ.
1420 497 510 508 524 537
Company 1000 303 306 319 341 339
Asian Univ. 1039 140 248 149 165 160
Total 7593 2662 2823 2765 2907 2925
Table 1. Number of robots.txt found in each
domain for each crawl.
To better describe the usage of the robots.txt in websites
in different domains, Figure 1 illustrates the proportion of
websites having robots.txt in each domain. Overall, except
for in the case of Asian university websites, the usage of
robots.txt has increased. 46.02% of newspaper websites
currently have implemented robots.txt files and the news-
paper domain is the domain in which the Robots Exclusion
Protocol is most frequently adopted. 45.93% of the USA
university websites in our sample adopt the Robots Exclu-
sion Protocol, significantly more than European (37.8%)
and Asian (15.4%) sites. Since search engines and intelli-
gent searching agents become more important for accessing
web information, this result is expected. The Robots Exclu-
sion Protocol is more frequently adopted by government,
newspaper and university websites in the USA. It is used
extensively to protect information not to be offered to the
public and balance workload for these websites. A detailed
robots.txt usage report can be found in [16].
5 Results
There are 1056 named robots found in our dataset. The
universal robot “*” is the most frequently used robot in the
Figure 1. Probability of a website that has
robots.txt in each domain.
User-Agent field and used 2744 times, which means 93.8%
of robots.txt files have rules for the universal robots. 72.4%
of the named robots appeared only once or twice. The most
frequently appearing robots in our dataset are shown in Fig-
ure 2.
Figure 2. Most frequently used robot names
in robots.txt files. The height of the bar rep-
resents the number of times a robot appeared
in our dataset.
5.1 History of Bias
The distribution of how many times a robot is used (see
Figure 3) did not change significantly over the past 11
months. Thus, we show the bias results from the latest crawl
since not much has changed.
Since most of the robots appeared only once or twice
in the dataset, their ranking scores are ranked in the mid-
dle of the list and are almost indistinguishable. We con-
sider only the top ranked (favored) and bottom ranked (dis-
favored) robots. The 10 most favored robots and 10 most

Figure 3. The distribution of a robot being
used.
disfavored robots are listed in Table 2 where N is the sam-
ple size, Nfavor is the number of times the robot is fa-
vored, Ndisfavor is the number of times the robot is dis-
favored and σ = ∆P (r)(1−∆P (r))
N is the categorical stan-
dard deviation[13] of ∆P(r). The categorical standard de-
viation σ gives the variance when using ∆P to estimate the
favorability of robots on the Web.
Our bias measure shows that the most highly favored
robots are from well-known search engines and organi-
zations, e.g., “Google”, “Yahoo” and “MSN” are favored
much more than the remaining robots. Please note that for
some robots in the disfavored category, their ∆P favorabil-
ity does not show a significant difference due to their rare
appearances in the sampled robots.txt files.
On the other hand, most of the disfavored robots are
email collectors (“CherryPicker” and “emailsiphon”) and
off-line browsers (“Wget” and “webzip”). From the pri-
vacy perspective, it is reasonable for webmasters to ex-
clude robots whose major purpose is to collect private
information. Also, webmasters typically do not want
their websites to be copied entirely by others. However,
even robots from well-known companies can be disfa-
vored e.g., “MSIECrawler” (Microsoft) and “ia archiver”
(Alexa). “MSIECrawler” is a robot embedded in Internet
Explorer (IE). When IE users bookmark a page while of-
fline, MSIECrawler downloads the page and all links related
to it, including links, images, JavaScript and Style sheets,
when the user is next online. “ia archiver” is the crawler
from archive.org and Alexa.com. A list of detailed descrip-
tion of known robots appeared in this paper can be found on
the web4
.
We also find that robot biases in different domains vary
4http://botseer.ist.psu.edu/namedrobots.html
Favored Robots (Sample size N = 2925)
robot
name
Nfavor Ndisf ∆P(r) σ
google 877 25 0.2913 0.0084
yahoo 631 34 0.2041 0.0075
msn 349 9 0.1162 0.0059
scooter 341 15 0.1104 0.0058
lycos 91 5 0.0294 0.0031
netmechanic 84 10 0.0253 0.0029
htdig 15 3 0.0041 0.0012
teoma 13 3 0.0034 0.0011
oodlebot* 8 0 0.0027 0.0010
momspider 6 0 0.0021 0.0008
Disfavored Robots (Sample size N = 2925)
robot
name
Nfavor Ndisf ∆P(r) σ
msiecrawler 0 85 -0.0291 0.0031
ia archiver 7 55 -0.0164 0.0023
cherrypicker 0 37 -0.0126 0.0021
emailsiphon 3 34 -0.0106 0.0019
roverbot 2 27 -0.0085 0.0017
psbot 0 23 -0.0079 0.0016
webzip 0 21 -0.0072 0.0016
wget 1 22 -0.0072 0.0016
linkwalker 2 20 -0.0062 0.0015
asterias 0 18 -0.0062 0.0015
Table 2. Top 10 favored and disfavored
robots. σ is the standard deviation of ∆P(r).
significantly. Google is always the most favored robot.
Other top favored robots vary in different domains. Ya-
hoo (“slurp” is a Yahoo robot) and MSN are also favored
in most domains, but they are not significantly favored over
other robots. Other top favored robots are mostly open
source crawlers and crawlers from well-known organiza-
tions. Disfavored robot lists vary widely for different do-
mains. Most of these robots are still email collectors and
offline browsers. The differences could be due to the dif-
ferent behaviors of robots in different domains (e.g., email-
siphon may crawl business websites more often than others
to collect business contacts).
5.2 Search Engine Market vs. Robot Bias
In order to study the impact of the “rich get richer” ef-
fect, we calculate the correlation between the robot bias and
the search engine market share for specific companies. The
market share of Google, Yahoo, MSN and Ask in the past
11 months and the ∆P favorability for the corresponding
robots are considered two independent variables. The Pear-
son product-moment correlation coefficient[15] (PMCC)

between the two variables is a measure of the tendency of
two variables X and Y measured on the same object or or-
ganism to increase or decrease together. For our dataset, the
Pearson correlation of the market share of the four compa-
nies and the ∆P(r) of their corresponding robots is 0.930
with P-Value < 0.001. The search engine market share 5
and robot bias in September, 2006 is shown in Figure 4.
Figure 4. Search engine market share vs.
robot bias.
6 Conclusions
We have presented a comprehensive survey of robot bi-
ases on the Web through careful content and statistical anal-
ysis of a large sample of robots.txt files. Results show that
the robots of popular search engines and information por-
tals, such as Google, Yahoo, and MSN, are generally fa-
vored by most of the websites we have sampled. This im-
plies a “rich get richer” bias toward popular search engines.
We also shows a strong correlation between the search en-
gine market share and the bias toward corresponding robots.
Our study indicates that the usage of robots.txt has in-
creased over the past 11 months in which 2,662 robots.txt
files were found in the first crawl and 2,925 files were found
for the last crawl. We observe that 46.02% of newspaper
websites currently have implemented robots.txt files and
the newspaper domain is the domain in which the Robots
Exclusion Protocol is most frequently adopted. 45.93%
of the USA university websites in our sample adopt the
Robots Exclusion Protocol, significantly more than Euro-
pean (37.8%) and Asian (15.4%) sites. Our future work
will try to further break down the analysis by geographical
region to investigate the robot favorability in each country.
Future work will pursue a deeper and larger scale analy-
sis of robots’ behavior and regulations. We are investigating
other metrics for robot bias. Experimental investigation of
5http://www.netratings.com
web robots’ behavior will be undertaken in order to better
understand how live robots interpret the Robots Exclusion
Protocol.
References
[1] Fortune magazine. http://money.cnn.com/magazines/fortune/-
fortune500, 2005.
[2] M. L. Boonk, D. R. A. d. Groot, F. M. T. Brazier, and A. Os-
kamp. Agent exclusion on websites. In Proceedings of The
4th Workshop on the Law and Electronic Agents, 2005.
[3] S. Chakrabarti, M. Van den Berg, , and B. Dom. Focused
crawling: A new approach to topic-specific web resource
discovery. In Proc. of the 8th WWW Conference, pages 545–
562, 1999.
[4] J. Cho, H. Garcia-Molina, and L. Page. Efficient crawling
through url ordering. In Proceedings of the 7th International
World Wide Web Conference, 1998.
[5] T. Y. Chun. World wide web robots: An overview. Online
Information Review, 23(3):135–142, 1999.
[6] DMOZ. The open directory project. http://dmoz.org, 2005.
[7] M. Drott. Indexing aids at corporate websites: The use of
robots.txt and meta tags. Information Processing and Man-
agement, 38(2):209–219, 2002.
[8] D. Eichmann. Ethical web agents. Computer Networks and
ISDN Systems, 28(1-2):127–136, 1995.
[9] T. Joachims. Optimizing search engines using clickthrough
data. In KDD ’02: Proceedings of the eighth ACM SIGKDD
international conference on Knowledge discovery and data
mining, pages 133–142, New York, NY, USA, 2002. ACM
Press.
[10] B. Kelly and I. Peacock. Webwatching uk web communi-
ties: Final report for the webwatch project. British Library
Research and Innovation Report, 1999.
[11] M. Kendall. Rank Correlation Methods. Hafner, 1955.
[12] M. Koster. A method for web robots control. In the Internet
Draft, The Internet Engineering Task Force (IETF), 1996.
[13] R. L. Ott and M. T. Longnecker. An Introduction to Statisti-
cal Methods and Data Analysis. Duxbury Press; 5th edition,
2000.
[14] G. Pant, P. Srinivasan, and F. Menczer. Crawling the Web,
chapter Web Dynamics. Springer-Verlag, 2004.
[15] R. R. Sokal and F. J. Rohlf. Biometry. Freeman New York,
2001.
[16] Y. Sun, Z. Zhuang, and C. L. Giles. A large-scale study
of robots.txt. In WWW ’07: Proceedings of the 16th inter-
national conference on World Wide Web, pages 1123–1124,
New York, NY, USA, 2007. ACM Press.

Determining Bias to Search Engines from Robots.txt

More Related Content

Viewers also liked (10)

Similar to Determining Bias to Search Engines from Robots.txt (20)

Recently uploaded (20)

Determining Bias to Search Engines from Robots.txt