Data Harvesting 2.0: from the Visible to the Invisible Web - Inria

14 Pages·2013·0.75 MB·English

Data Harvesting 2.0:
from the Visible to the Invisible Web
ClaudeCastelluccia Ste´phaneGrumbach LukaszOlejnik
INRIA INRIA INRIA
[email protected] [email protected] [email protected]
Abstract It might at first sight seem irrelevant to know which cor-
porations are handling and concentrating data, and in what
Personal data are fuelling a fast emerging industry which
countriestheyarebased,butithastremendousconsequences
transform them into added value. Harvesting these data is
related in particular to the jurisdiction that applies, not to
therefore of the outermost importance for the economy. In
mentiontheeconomicimpact,directorindirect.Somecom-
this paper, we study the flows of personal data at a global
plexfiscalissueshavealreadybeenraisedinEuropeonthe
level, and distinguish countries based on their capacity to
economyofthedata,whichatthisstageismostlyinvisible
harvestdata.Weestablishacartographyofinternationaldata
toEuropeanstates1.
channelsonthevisibleandinvisibleWeb.ThevisibleWebis
Personaldatahavebecomeanessentialresourceforthenew
composedofthesitesthatareavailabletothegeneralpublic
economy of the information society, much like iron ore or
and are typically indexed by search engines. The invisible
crudeoilwerefortheindustrialeconomy.Developingtools
Webreferstotags,Webbugs,pixelsandbeaconsthatappear
toharvestpersonaldataisthereforeofstrategicimportance
onWebsitestotrackandprofileusers.
to catch up with the digital revolution. Harvesting data is
It is well known that the US dominate the visible Web
mostlydonebysystems,suchassearchengines,socialnet-
with more than 70% of the top 100 sites in the world. We
works,clouds,etc.wherepeopleprovidetheirpersonaldata
show that this domination is even stronger on the invisible
in exchange for a service, which most of the time is acces-
Web.Thelargestproportionoftrackersinmostcountriesare
sible essentially for free 2. More precisely, the private data
indeed from the US. Apart from the US, two countries ex-
givenbyusers,constitutethemeansbywhichuserspayfor
hibitanoriginalstrategy.China,whichdominatesitsvisible
services.Itisthereforenotonlyaresource,butcanbeseenin
Web with a majority of local sites, but surprisingly these
asenseasavirtualcurrency.Harvestingdatacanbedoneas
sites still contain a majority of US trackers. Russia, which
wellinmoresubtleways,onthe”invisibleWeb”,bytrack-
alsodominatesitsvisibleWeb,andistheonlycountrywith
ingpeople,whichiscarriedonmostlybythirdpartyentities.
morelocaltrackersthanUSones.
The main objective of this paper is to analyze the current
1. Introduction
worldsituation,andperformageographicalanalysisofdata
SunTzufamouslywroteinthe6thcenturyBC:”Ifyouknow harvestingonboththevisibleandtheinvisibleWeb.Thevis-
yourenemiesandknowyourself,youwillnotbeimperilled
ible Web is composed of the sites that are available to the
in a hundred battles; if you do not know your enemies but
generalpublicandaretypicallyindexedbysearchengines.
do know yourself, you will win one and lose one; if you
TheinvisibleWebreferstotags,Webbugs,pixelsandbea-
do not know your enemies nor yourself, you will be imper-
consthatappearonWebsitestotrackandprofileusers.
illed in every single battle” [1]. This statement might be-
comemorerelevantthaneverwiththeexplosionofinforma- Thispaperaimsatansweringthetwofollowingquestions:
tion now available from the Internet and Web 2.0 systems.
Some countries are widely collecting data on the visible or 1. What are the most influential Web systems? In which
the invisible Web about their citizens as well as sometimes countries are they based? What are their regional influ-
about citizens of other countries. Some other countries on ence?
the other hand are relying mostly on foreign systems, thus
lettingalargeamountoftheirdatabehandledoutsidetheir 1Somecountries,suchasFrance[19],actuallyintendtotaxthetracking
borders.Thisdiscrepancybetweenthesetwoapproachesto and data-mining providers. The premise behind this assumption is that
consumerspayforservicewiththeirprivatedataandthisgoodisexchanged
the management of personal data could result in informa-
withoutissuingatax.
tionasymmetriesbetweentheplayers,whetherindustrialor 2Notethatmostpeoplepayabout1000dollars/year(ISP,3G,...)toaccess
governmental,intheircapacitytoaccessstrategicdata[24]. theInternet.Itisarguablewhethertheseservicesarereallyfree.
2. WhicharethebiggesttrackersontheInternet?Inwhich cently certain countries, notably the UK, passed laws re-
countries are they based? What are their regional influ- quiring Websites to gain their users consent for the cookie
ence? usage [14]. Examples include the site of the BBC, recently
extendedwithaninformationpop-up.Usersareexpectedto
In other words, this paper aims at analyzing the cyber-
consentthoughtheymostlymechanicallydoso,withoutany
strategies of different countries in terms of data harvesting
specialconsiderationorunderstanding.
onboththevisibleandtheinvisibleWeb.
Itisimportanttonotetheexistenceofothertrackingmech-
Interestingly, the answers to the previous questions reveal anisms. Of particular importance, in addition to Evercook-
very different patterns, not necessarily correlated with the ies[11],arenon-standardbrowserfingerprintingtechniques
penetrationrate.Numerousstudieshavemeasuredthelevel such as browser configurations [4], history [18], host iden-
ofpenetrationoftheinformationsocietysuchastheWorld tifiers [26], pixels [16], allowing tracking of the browser
EconomicForum[3].Recently,theWebFoundation3,ledby acrossvarioussites.Thediscussionoftherisksandprotec-
Tim Berners-Lee, launched its Web Index, which as previ- tions against potential tracking by social buttons (i.e. Face-
ousstudiesrankstopofthelistNorthAmericaandNorthern book’s Like) is covered in the work of [22]. According to
Europe, as well as some countries of Asia. The index mea- [18], most popular sites can be leveraged to basic history-
sures three key attributes of the Web: ”Web readiness”, for based fingerprinting and using this intuition, we assumed
thecommunicationsinfrastructure;”Webuse”,forthepop- thiscanalsobethecaseofthemostimportantsitesfortrack-
ulationonlineandthecontentsavailable;and”theimpactof ing.
theWeb”,foritseconomic,socialandpoliticalimpact.The New risks resulted in new reactions. Quite recently, a new
Web Foundation gives a high weight to the political open- and bold initiative attempts at limiting the tracking on the
ness. China ranks 29th out of 61 countries in this index, a Internet. This initiative, known as Do Not Track [15], pro-
ratherlowrank,whichrevealsthoughanincreasedWebuse, motes the consensual and easy solution of opting out from
butarelativelyslowevolutionofWebcontent. tracking on the Web. It is technically achieved by a simple
Thereisonemeasurethatislittletakenintoaccountinthese additionofananotherHTTPheaderinthebrowser’srequest:
rankings, namely the local development of the Web 2.0 in- DNT[25].AccordingtoMozilla,morethan11%ofFirefox
dustry,whichoffersanindicatorofthepotentialstrategicac- usershaveactivatedtheDNT.TheDNTinitiativeleadtodif-
tivityonBigDatainthecountry4.TheUSwouldberanked ficultnegotiationswiththeadvertisementindustryintheUS
atthetoppositionforsuchacriteria.Theyhaveindeedde- [9].
veloped the strongest industry worldwide, with most of the Trackingcanalsobelimitedbyusingdedicatedbrowserex-
first online social systems accessed in the world, such as tensions, which can block unwanted tracker scripts and/or
Google, Facebook, YouTube, Yahoo!, Wikipedia, Windows ads.Weselectedtwoofthemostprominentones,Ghostery
Live,Twitter,Amazon,tonamethemostpopular.Withthese and AdBlock Plus, and compared them to present actual
corporations,USAharvestsprivatedataofpeopleallaround metrics of performance. For a global analysis, we refer to
the World that can be analysed for an unpredictable set of [12],whichshowshowtrackinghaschangedwithtimeand
purposes,withconsiderableeconomicimpact.Ofcoursethe acquisitions of various companies by others. Our work fo-
UShaveadominantpositioninanumberofstrategicsectors cusesonaglobalapproach,incontrastwith[18],wherethe
oftheinformationsociety,includingtheoperatingsystems, situation of individuals was addressed. Geographic differ-
the browsers, or the clouds that support the systems of the encesintheTwitternetworkhavebeenstudiedaswell[13].
Web. Whilesomeanalogiestoourresearchcanbefound,trackers
informationarefundamentallydifferentfrominformationon
Technically speaking, tracking is made possible by the de-
Twitter’susers.
sign of the HTTP protocol [17] itself. Third-party scripts
Paper organisation: The paper is organized as follows. In
can indeed be used for tracking purposes. Every inclusion
the next section, we present the top harvesting sites of the
ofathird-partyscriptonafirst-party,i.e.onthevisitedsite,
visible Web, and their geographical influence. Section 3 is
requires the browser to execute a request to this third-party
devotedbothtothetrackeesandthetrackersoftheinvisible
server, download the script and execute it into the user’s
Web. In Section 4, we analyse the correlations between the
browser.
differentharvestingtechniques.
Browser cookies [2] are traditionally used to maintain a
browser-state of a Web user. They allow to tie a given
browser with the internal profile in a third-party database.
This motivates to limit the use of cookies on the Web. Re-
2. ThevisibleWeb
3http://www.Webfoundation.org/
Our investigation focuses on the top Websites of 55 coun-
4Inthesequel,weconsiderthemostpopularWebsites,globallyorinone
tries,whichhavebeenidentifiedusingthestatisticsprovided
specificcountrybasedonAlexa’sranking.AlexaisasubsidiaryofAmazon.
alexa.com by Alexa. For simplicity, in the sequel we use the country
code top-level domain in CcTLD format5. Alexa maintains with the american systems. Both China and Russia have
lists of over 500 sites per country, but we restricted our at- developedaverypowerfulindustrywhichharvestsmostof
tentiontothe25to100mostpopularsitesintheselists. the data produced by their citizens. In China most of the
first 50 sites are Chinese [10]. As shown in the infography
2.1 Topsitesbycountry
produced by Ogilvy8, there is no area of the social media
Data on the Web 2.0 are produced by users everywhere in where a Chinese company cannot be found. Moreover, in
theWorld,buttheyareaccumulatedbycorporations,which some areas, several very large systems coexist, while only
formostusersworldwidearenotintheirowncountry.Afirst one dominates in the US, not to mention the rest of the
measureofthisphenomenoncanbeestimatedbymeasuring World.Itisthecaseformicrobloggingplatformforinstance,
foreachcountrythepercentageofnationalWebcorporations where Sina Weibo, and Tencent Weibo coexist with both
amongthetop25sitesusedinthatcountry6.Theresultsare around 300 million users, and both ahead of Twitter. The
striking.Table1presentsforafewrepresentativecountries ratio of local sites in Russia is lower than in China. Most
thepercentageofnationalWebcorporationsamongthetop ofthetopsitesoftheUShavepredominantpositions(e.g.,
25ineachcountry. Facebook) in Russia, while they are blocked in China. The
relativesizeofthetwocountriesthoughimpactonthesizeof
CC Nat.ratio ForeignSites
theirfirstsystems,butbothalsohavetheirrespectivesphere
US 100% noforeignsite
ofinfluenceabroad.
CN 92% onlyforeignsite:Google
SouthKoreahasanextremelyinterestingpatternofdiversity.
RU 68% mostlyamericansites
Amongthetop25sites,thereareonly24%ofsiteswhichare
JP 36% mostlyamericansites
national,whilethereare36%ofbothAmericanandChinese
KR 24% halfAmerican
sites,aremarkablesituation.Amongolianportal(zaluu)also
halfChinese
belongtothelist.
FR 36% Onlyamericansites
NG 24% mostlyamericansites Letusconsidernowthetop100sites.Whenlookingbeyond
thetop25sites,theratioofnationalsitesincreases,inpar-
Table1:PercentageofnationalWebsitesamongtop25 ticularwithmostofthelocalnewspapersandmagazines,as
wellassomeservicessuchasbankinginstitutions.
In the US, there are no foreign sites among the top 25
Websites.Forallothercountriesweconsidered,apartfrom
China and Russia, the ratio of national sites amounts at
best to around a third of the Websites. Both in Japan and
France,only36%ofthetop25Websitesarenational,butthis
number hides very different realities in the two countries.
First,whileinFranceall64%offoreignsitesareAmerican,
Figure1:Ratio forsitesinAlexalists,byorigin.
inJapan,thereismorediversity.TwoChinesesites(search us
engine Baidu, instant messaging QQ) and one Korean site
Figure1showsforaselectedlistofcountries,theproportion
(search portal Naver) belong to the top 25 sites in Japan.
ofsitesfromtheUS(inblack),thecountryitself(inred),as
Second, and more importantly, the French sites are mostly
well as the two others locally most active foreign countries
sites7, such as newspapers, which do not gather as much
(inblue),inthetop100sites.Theresultsarenormalizedto
personal data, while in Japan, national sites include very
thenumberofsitesfromtheUS,thatistheycorrespondtoa
dataintensiveones,suchasWebportals,e-commerce,blogs,
ratio,wheretheper-countrycountisdividedbythenumber
etc. Similar patterns would be found for other European
oftheUSbasedsites.
countries. Italy for instance has only 28% of national sites.
Differentpatternscanbeobserved.ForEuropeancountries,
InAfrica,Nigeria,hasonly24%ofnationalsites,mostlyin
Hungary, Poland, Norway and France, the number of local
thePress.
sites is about twice the number of US sites in the top 100.
Chinaistheonlycountrywhichhasdevelopedsystemswith Russiahasasimilarpattern.Chinaontheotherhandmain-
a number of users, in the hundreds of million, comparable tainsmorethan80%oflocalsitesinthetop100list.Nigeria
showsaverydifferentsituationwithaverysmallproportion
5AT, AU, BE, BR, CA, CH, CN, CY, CZ, DE, DK, DZ, EC, ES,
oflocalsitesinthetop100.
FI, FR, GB, GR, HK, HN, HU, IE, IL, IN, IT, JP, KR, KW,
KZ, LY, MA, MY, NG, NL, NO, NZ, OM, PA, PE, PL, PT, PY, NotethatthepercentageofusersthatvisitspecificWebsites
QA, RO, RU, SA, SE, SG, SI, SN, TH, TN, US, UY, VN. decreases very fast from the top sites in a list to the subse-
6Unlessotherwisespecified,thenumberspresentedinthefollowingtables
areextractedfromAlexa’srankingasofmidseptember2012. 8China social media equivalents: a new info-
7National sites among the top 25 in France: leboncoin, Orange, Free, graphic http://www.asiadigitalmap.com/2011/02/
commentcamarche,lemonde,lequipe,lefigaro,pagesjaunes,sfr china-social-media-equivalents-a-new-infographic/
quent sites. The first ten sites generally concentrate an im- Facebook.com and Youtube.com, we show the number of
portant share of the total traffic. Countries whose ratio of countriesinwhichthesiteisamongthetop2or4sites(as
websites is low in the first segment and increases then for indicatedinthetable).
the top 100, have in general a relatively small share of the It is important to note that the traffic decreases very fast in
globaltraffic. thetop50list.Google.com,whichoccupiesthetopposition
for instance drives more than 40% of global users (not to
2.2 Topsitesglobally
mention the local sites such as google.de, google.com.hk,
WenowconsidertheglobalimpactofWebcorporations. etc.),andisamongthefirst4sitesinmorethan30countries,
Table2presentstheproportionsoftop50sitesintheworld while Sohu.com, which occupies the 50th position, drives
thatareownedbycompaniesinagivencountry. about 2% of users and is among the top ten sites only in
TheUShavemorethantwothirdsofthetop50sitesworld- China.
wide. These sites have a real prominence worldwide as we
Rank Website %users #countries
haveseenontheprevioustable.Theonlytwocountriesthat
1 Google.com 43.75 30(1st-4th)
have more than one site in this club, are China and Rus-
2 Facebook.com 42.65 34(1st-2nd)
sia, which have both developed their own industry for fun-
3 YouTube.com 33.43 35(1st-4th)
damental tools such as search engines, blogs, e-commerce,
4 Yahoo.com 20.11 29(1st-10th)
etc. Three countries, in the European sphere, have one site
5 Baidu.com 12.19 3(1st-10th)
amongthetop50.
8 Amazon.com 8.26 15(1st-10th)
Here again, China occupies a remarkable position after the
18 Yandex.com 2.97 6(1st-10th)
US,whichhavetheabsolutesupremacy.Chinahaseightof
the first fifty sites worldwide according to Alexa’s ranking. 32 Tumblr.com 2.57 0(1st-10th)
IfthesizeoftheChinesepopulationimpactsofcourseonthe 50 Sohu.com 1.84 1(1st-10th)
numberofusersoftheChinesesystems,andthereforeulti-
Table3:Percentageofglobalusersandtopcountries
matelyontherankingofthesesystems,themostimportant
reasonfortheirsuccessistheassociationofaclearpolitical
ambition, a strong appetence for social networking, and a
dynamicindustry.Thesizeofthepopulationisbynomeans 2.3 Searchengines
an explanation by itself. India for instance has only a few Letusconsidermorecarefullyparticularsegments,suchas
national sites among its top 25 sites, which are almost all thesearchengine,whichplaysanessentialroleintheway
American. people access knowledge. Here again, distinct patterns can
Unlike their American counterparts, Chinese systems have befound.Table4presentsthetopsearchenginesforafew
currently most of their users in China. Most of them are countries.
of course widely used in Hong Kong and Taiwan as well,
CC 1stSE share 2ndSE share
while some are also used in South Korea, and in Russia
suchasTaobaopopularfore-commerce.Theirinternational US Google 65% Bing/Yahoo 15%
ambitionwillmostprobablygrowinthecomingyears. CN Baidu 73% Qihoo/Sogou 8-9%
JP Yahoo!Japan 51% Google 36%
CC Ratio Topsiteswiththeir(rank) RU Yandex 60% Google 25%
US 72% Google(1);Facebook(2);YouTube(3); UK Google 91% Bing 5%
Yahoo(4);... FR Google 92% Bing 3%
CN 16% Baidu(5);QQ(8);Taobao(13);Sina(17); CZ Google 53% Seznam 37%
163.com(28);Soso(29);Sinaweibo(31);
Sohu(43) Table4:Thetoptwosearchenginesbycountry
RU 6% Yandex(21);kontakte(30);Mail.ru(33)
IL 2% Babylon(22)
The US have developed major search engines. The three
UK 2% BBC(46)
which dominate the American market, Google, Yahoo and
NL 2% AVG(47) Bing,areamongthemostpopularworldwide.Googlehasa
dominant position with 65% market share, while Bing and
Table2:TheTop50Websitesworldwidebycountry
Yahoo have both 15% share in USA. Globally Google has
65%marketshare,Baidu,8.2%marketshare,Yahoo,4.9%
Table 3 shows the percentage of global users who visited a market share, Yandex, 2.8% market share, and Microsoft,
site in the last three months. It also displays the number of 2.5%marketshare9.
countriesinwhichthesiteappearsinthe10topsiteslocally. 9http://searchengineland.com/google-worlds-most-
Note that for the most popular sites, namely Google.com, popular-search-engine-148089
China10 and Russia11 are in a very similar situation, where otherAsiancountriessuchasJapanandKoreaforinstance
the dominant search engine is the local one, Google be- shows their appetence for local systems. The strong focus
ing the next most widely used engine. Baidu has a rela- inWesternmediaonthecensorshipimposedontheInternet
tivelybiggershareinChinathanYandexinRussia.Morere- hasoftenledtounderestimatethestrategyofChinatowards
cently,GooglelostsharesinChina,withthesuddenraiseof ITandtheinformationsociety,andoverestimatetheimpor-
twootherlocalsearchengines,whichareapproaching10% tanceofthecontrolofthecontent.
sharesoftheChinesemarket,Qihoo360,andSogou.
In Europe12, there are no local search engines with strong 3. TheinvisibleWeb
positions,andthemarketisdominatedbyGoogle,whichhas
While Section 2 deals with the cartography of the visible
aquasimonopolisticposition.OnlySeznamhasareasonable
Web, this section analyses what is often referred to as the
shareforczech,butwhichisnowdecreasingwithrespectto
invisible Web. The invisible Web refers to tags, Web bugs,
Google’sshare.
pixels and beacons that appear on Websites to track and
2.4 Socialnetworks profileusers.Wefirstpresentthemethodologyusedtotrack
the trackers. We then consider the invisible Web from the
Otherdomainsoftheinformationsocietysuchassocialnet-
trackeespointofview,weaimtoshowhowInternetusersare
works would lead to very similar conclusion, with Face-
tracked across the world. Finally, we analyze the trackers,
book largely dominating in Europe, while alternative sys-
andconsiderwhetherthetrackersaredistributeduniformly
tems have been developed in Asia. The size of the Chinese
ontheplanet,orwhethersomecountriesaredominatingthe
social networks deserve some attention. The ranking of the
trackingbusiness.
GlobalWebIndexbasedonthepercentageofglobalInternet
usersarestriking.Table5showsthat6outof10ofthemost
3.1 Trackingthetrackers
widelyusedsocialsystems13areChinese.
In order to establish a cartography of global third-party re-
CC Corporation share source utilization on the Web, we used PlanetLab’s infras-
US Facebook 41% tructure [20]. PlanetLab connects many servers in different
countries.Inourexperiment,37proxyserversfromdistinct
US Google+ 21%
countrieshavebeenused14.
CN Qzone 19%
To retrieve information on trackers, we created dynamic
CN SinaWeibo 18%
tunnelstotherelevantPlanetLab’sservers.Subsequently,all
CN TencentWeibo 16%
sites from the respective lists were visited and the trackers
US Twitter 16%
detectedonthesesitesweresavedforfurtheranalysis.This
CN Renren 11%
processwasautomatedwiththeuseofaWebDrivertogether
CN Kaixin 8%
withaFirefoxbrowser,equippedwithmodifiedpluginsand
US LinkedIn 7%
Flash enabled. All our data has been obtained between the
CN 51.com 6%
endofoctoberandthebeginningofdecember2012.
Forourtestswehavechosentwopopulartoolsenablingthe
Table5:PercentageofglobalInternetusers
detection and blocking of third-party resources: Ghostery
andAdBlockPlus(ABP).Theybothworkinasimilarman-
China has developed a large industry on the net, with es- nerrequiringthescanningofthevisitedWebsite,andsearch-
sentially all the usual services initially proposed by mostly ing for an offending resource or connection. If a resource
American companies, such as online search engines, social is found to be present on the respective list of blocked re-
networks, news, business, instant messaging, etc. Chinese sources (filter lists), this may either be reported to the user
companieshavetakenadvantageintheirdevelopmentofthe orblockedbytheplugin.
difficulties to access their foreign counterparts from Main-
Ghostery. Ghostery is a popular extension which detects
land China, but they would most certainly have succeeded
trackers and display their names (such as “DoubleClick”,
withoutthecensorshipofforeignsites.Thediversityinsome
“Omniture”) [6]. Ghostery analyzes the requests made by
10http://www.chinainternetwatch.com/1444/ a browser and compares them against a database of known
china-search-engine-market-share-by-revenue-q1-2012/ trackers. It is important to note that Ghostery maintains a
11http://www.bloomberg.com/news/2012-04-02/ listofconfirmed trackers:atrackerisnotonlyaddedtothe
yandex-internet-search-share-gains-google
database, but also included on the project’s Webpage (e.g.
-steady-liveinternet.html
12http://theeword.co.uk/seo-manchester/google_tops_july_ http://www.ghostery.com/apps/omniture),withtheir
2012_search_engine_market.html respective privacy policy. In our experiment, we saved the
13http://globalWebindex.net/thinking/
social-platform-report-series-september-2012 14Ifaserverfromaspecificcountrywasunavailable,weusedalocalIP
-facebook-on-track-to-hit-2bn/ fromInria.
namesofthetrackersfoundforeveryvisitedsiteforfurther
analysis.
AdBlock Plus (ABP). For comparison purposes, we also
used the AdBlock Plus extension [21], in the same envi-
ronment as described previously. AdBlock is able to block
third-partyresources,mostlyunwantedadvertisements,and
it maintains a dedicated trackers list: EasyPrivacy15, which
includes many known trackers as well as Web bugs, which
allows AdBlock Plus to block these resources. We use this
laterlistinouranalysis.
Although these two tools are different, they provide results
that are consistent. A more detailled comparaison of these
toolsandresultsareprovidedinAppendixA.
Country of origin. Determining the country of origin of
Webcorporationsisachallengingproblem.Onesolutionis
toidentifythelocationoftheirheadquarters,butthisisnot
Figure2:Averagenumberoftrackers(Ghostery)
alwaysrelevant.ItisalsopossibletouseWhoisdatabasesto
identifythelocationofthesiteholdingtheirdomainname.
However,domainregistrarsaresometimelocatedelsewhere
than reported in Whois databases. A third possibility could are accessed from a local IP, and the trackers are detected
be,inthecaseofABPforinstance,tousethetrackers’do- usingGhosterydataforthe55selectedcountries.Theresults
mainnameresolutionandperformgeolocationusingappro- clearlyshowthatusersaretrackeddifferentlyontheInternet
priate databases. None of these methods are fully satisfac- accordingtotheircountry.Thedifferentcolorsonthefigure
tory.Weproposeinsteadanalternativetechnique,whichcan refertothedifferentcontinents16.AsshowninAppendixB,
beautomated. these results do not depend in fact on the visitor locations.
ForABP,aWebresourceisassociatedtotrackers.Thetop- Inotherwords,inmostcases,agivensitetracksitsvisitors
level domain name can be extracted, a domain name reso- independentlyofthevisitor’sIPaddress.
lution can be performed, and subsequently a lookup using While results do not show big differences between conti-
geolocationdatabase,toidentifyitslocation.Thisapproach nents, some countries seem to be much more tracked than
can correctly identify the country of origin in most of the others.Forexample,USInternetusersaretracked5times
cases. For Ghostery, we performed a geolocation on the morethanChineseusers17.
URLs in the trackers description from the Website http:
//www.ghostery.com/apps/, where all the trackers are
described by Ghostery themselves. The geolocation has
beendoneusingPythonGeoIPandgeoiplookupcommand-
line utility tools, which query geolocation databases. The
results were then cross-checked manually. For example
http://www.ghostery.com/apps/digilant mentions
Digilant company’s Website www.digilant.com, which
resolves to United States. The ”Privacy Contact“ tab on
Ghostery’s Website confirms that Diligant headquarters are
locatedinBoston,USA.
3.2 Thetrackees
Wefirstconsidertheaveragenumberoftrackerspersite
foreachofthe55countriesconsidered.Theaverageiscom-
puted by connecting to the top 100 most popular sites of a
country, summing the number of trackers on each of them,
Figure3:Averagenumberoftrackers(Ghostery,mobile)
anddividingthefinalresultbythenumberofretrievedsites.
Figure 2 displays the average number of trackers (Y axis) 16Europeisingreen,Asiainblack,NorthAmericaindarkblue,Africain
for each considered country (X axis). The top 100 sites lightblue,SouthAmericainpurple,Australiainwhite
17WedefineasaUS(resp.Chinese)user,auserthatvisitstheAlexa100
15copiedon30/11/2012 top,i.e.the100mostpopular,sitesintheUS(resp.inChina)
Figure3displaysthesametypeofresultsasFigure2,with
thebrowser’s User-Agentstring setto amobile device.Al-
thoughtheresultsaredifferent,withasmalleraveragenum-
beroftrackers,thetrendsaresimilar.Somecountriesarestill
muchmoretrackedthanothers.Thisdifferenceislikelydue
to the fact that many Websites redirect the mobile browser
toaspecialversionofthesite,tailoredtowardsdeviceswith
asmallerdisplay.Thesesites,asitseems,includelessthird-
partyresources,probablytospeeduploading.Furthermore
manypopularsiteshaveadedicatedmobileapplicationany-
way,sothepotentiallossesfromtrackingand/oradvertising
canbebalancedwiththeuseofin-appadvertisements.
Figure5:Distributionoftrackersforselectedcountries
much larger in the US than in other countries, which auto-
matically increases the average number of trackers per site
intheUS.
Finally,weconsideredthedistributionofdifferenttracker
types. Ghostery divides its detected scripts into five cate-
gories[7],whichwerecallbelow:
1. Ad:advertisementsprovidedbythead-networks;
2. Tracker: scripts which actually perform tracking (often
usingverysophisticatedbehavioralanalysis);
Figure4:Averagenumberoftrackers(ABP)
3. Analytics: utility scripts for Website creators allowing
them to discover various statistical details about their
Figure 4 exhibits results from experiments similar to those visitors;
of Figure 2, but while using ABP instead of Ghostery. The
4. Widget: small Web applications such as clocks, weather
absolutenumbersareslightlysmaller,butthetrendsarevery
tables,andothers.OtherexamplesincludeFacebookSo-
similar.AppendixApresentsmoreexperimentsthatanalyse
cialPlugins,Google+1,etc.;
theconsistencyoftheresultsobtainedbyABPandGhostery.
5. Privacy:typicallyascriptdisclosingprivacypoliciesand
We then considered the distribution of trackers for se- practicesrelatedtoads,suchasEvidonNotice18.
lectedcountries,whichisshownonFigure5.Thedistribu-
Figure 6 displays the distribution of trackers for the five
tion isobtained by counting, fora given country,all occur-
categoriesof[7]forthe55countries,basedonGhostery.It
rencesofaparticulartracker,e.g.,DoubleClick.Thesenum-
isinterestingtonotethatthedistributionofeachtypeseems
bers are then ordered from the largest to the smallest, and
tobequitesimilarindifferentcountries.TheAdtrackersare
plotted. The first value of a given curve shows the number
themostcommon,followedbytheanalyticsones.
ofoccurrencesofthemostpopulartracker,thesecondvalue
showsthenumberofoccurrencesofthesecondmostpopular
3.3 Thetrackers
tracker,andsoon.TheanalysisisbasedonGhostery.
Wenextconsiderthegeographicoriginofthetrackers,and
The results clearly show that in China there are only 10
thewaytrackersproliferatedependinguponthecountrythey
different trackers on the top 100 Chinese sites, and these
come from. We first start by analyzing the origin of the
trackers are not very active. Moreover, the most popular,
trackers on the top 100 sites of the global list, that is the
CNZZ, only appears in 10 of the top 100 Chinese sites. In
listofthe100mostpopularWebsitesworldwide.
contrastothercountries,suchastheUS,havealargenumber
ofdifferenttrackers(about90fortheUS),andsomeofthese
Table6showsthedistributionofthedetectedtrackersusing
trackers,forexampleGoogleAnalytics,areverypopularin Ghostery. We computed the number of trackers, T, on the
alargenumberofsites.
Does it mean that the US market is more fragmented? In- 18http://www.evidon.com/about/
deed, it seems that the number of tracking companies is 19ROW:RestOfWorld
Nigeria,andHungary,andanalyzetheoriginofthetrackers
ofthetop100sitesforeachofthesecountries20.
CC Tracker(countryoforigin) Count
GoogleAnalytics(US) 53
LiveInternet(RU) 51
RU Yandex.Metrics(RU) 31
TNS(GB) 30
Rambler(RU) 27
GoogleAnalytics(US) 20
MarkMonitor(US) 16
CN CNZZ(CN) 10
ScoreCardResearchBeacon(US) 6
DoubleClick(US) 5
GoogleAnalytics(US) 39
Figure6:Distributionoftrackersbytypes
DoubleClick(US) 32
US ScoreCardResearchBeacon(US) 32
CC Share Omniture(US) 21
US 87% FacebookConnect(US) 14
CN 3% GoogleAnalytics(US) 62
FacebookSocialPlugins(US) 39
GB 3%
NG DoubleClick(US) 32
RU 2%
ROW19 4.2% FacebookConnect(US) 30
GoogleAdsense(US) 23
Table6:Distributionoftrackersontop100globalsites GoogleAnalytics(US) 70
Median(HU) 48
HU Gemius(PL) 47
Adverticum(HU) 37
top100sitesworldwide.Wethencountedforeachcountry
FacebookSocialPlugins(US) 31
i,thenumberoftrackersoforigini,C ,andthencomputed,
i
foreachcountry,thepercentagep =C /T.
i i
Table8:Top5trackersforselectedcountries
The results show a clear domination of the US, with more
than80%coverageofthetop100sitesworldwide,withonly
afewcountries,suchasChina,GB,andRussia,whichhave
Table8displaysthetop5mostpopulartrackingcompanies
asmallpercentageoftrackers.
onthe100mostpopularWebsitesofthese6selectedcoun-
tries. Once again, the results clearly indicate that the track-
Tracker(countryoforigin) Count ing business is largely dominated by US companies. Only
GoogleAnalytics(US) 25 fewcountriesseemtoresistthisdominationsuchasRussia,
DoubleClick(US) 20 China,andHungary.Interestingly,Russiaistheonlycoun-
ScoreCardResearchBeacon(US) 19 trywhosetrackersaremostlylocal.
FacebookSocialPlugins(US) 11
We next consider the origin of the most prominent trackers
Omniture(US) 10
in the following eight countries: USA, Hungary, Poland,
Norway,France,Russia,China,andNigeria.Weperformed
Table7:Top5trackersontop100globalsites
thefollowinganalysis.Wefirstcollectedallthetrackerson
thetop100sitesofacountry,classifyingthemaccordingto
Table7displaysthetop5mostpopulartrackingcorporations their origin, and counting their occurrences. Since the raw
on the top 100 most popular sites worldwide. The results numbers,evenaverages,maynotbethemostinformativein
clearly confirm that the tracking business is largely domi- thisanalysis,wedecidedtoplottheinformationwithrespect
natedbyUScompanies. to the detected number of US-based trackers to ease the
presentation.Morespecifically,wecomputedforalltrackers
Letuspushfurthertheanalysisoftheoriginofthetrackers,
ofanoriginN detectedonthetop100sitesofacountryC,
andconsideranin-depthAnalysisforSelectedCountries.
For the sake of readability, we decided to display here the 20Theanalysisofall55countrieslistedatthebeginningofthispaperis
results for only 5 countries, namely Russia, China, USA, presentedinAppendixC.
4. SitesvsTrackersAnalysis
In this section, we compare the harvesting activity on the
visibleandtheinvisibleWeb.Ourobjectiveistounderstand
whether the predominance of the US is larger on the visi-
bleorontheinvisibleWeb.Inordertoachievethisgoal,we
compute and compare the proportion of US sites (resp. US
Figure7:Ratio fortrackersbyorigin(Ghostery) trackers),withtheproportionoflocalsites(resp.localtrack-
us
ers)foreachofthetop100sitesofeachofthe55considered
countries.
Theseproportions,P andP ,arecomputedasfollows:
US local
#UStrackers(resp.sites)incountryC
P (C)=
US
#Alltrackers(resp.sites)incountryC
P (C) is the number of US trackers (resp. US sites) di-
US
videdbythetotalnumberoftrackers(resp.sites)inthetop
Figure8:Ratio fortrackersbyorigin(ABP) 100sitesincountryC.Quantitatively,itshowshowtrackers
us
ofadominatingcountry(e.g.US)tracktheworld.
#Localtrackers(resp.sites)incountryC
thefollowingratio: Plocal(C)=
#Alltrackers(resp.sites)incountryC
#trackersoforiginN P (C) is the number of local trackers (resp. local sites)
Ratio (N)= local
us
#UStrackers dividedbythetotalnumberoftrackers(resp.sites)inthetop
100sitesincountryC.Quantitatively,itshowsthestrength
TheUShasbeenchosenasareferencebecauseoftheglobal
oflocaltrackers.
prevalenceofthetrackersofthisorigin.Indeed,UStrackers
arepresentonalmosteverysites.
Figures 7 and 8 show for each of these selected countries
theoriginofthetop4mostprominenttrackingcountriesfor
both detection methods, Ghostery and ABP. In parenthesis
for each country, the ratio of US trackers, and the ratio
of local trackers. These figures clearly show that the US
havealmostalwaysthelargestratiooftrackers.Thesecond
largest tracker country, is apart from some exceptions, the
country itself. Austria constitutes for instance an exception
tothisrule,withGermanyassecondtrackerwitharatioof
0.24.
China and even more so Russia constitute the two excep-
tions,withexceptionallystrongratiosoflocaltrackersover
US trackers. In fact, China has a ratio of 0.78, and Russia
aratioof2.3,thatindicatesthatRussiansitescontainmore
RussiantrackersthanUSones. Figure9:RatioUSvslocalsitesandtrackers
Wealsoobservedsomevariationinthenumberandthedis-
tributionofcountriesinvolvedintrackinginagivencountry, The scatterplot for P and P of all the 55 countries
US local
beyondtheUSandthecountryitself.ForexampleinFrance, consideredforsitesandtrackersareshownonFigure9.Ev-
there are third-party resource providers from 11 countries, ery black square (resp. red circle) point corresponds to a
whereasinChinaonly6countriesare”represented“. specific country C, with coordinates P (C) (X axis) and
US
Inaddition,weobservedthatthesetrackersoftencomefrom P (C) (Y axis), for sites (resp. trackers). The trackers
local
neighboringcountries.Forexample,Danishsitesoftencon- have been obtained using Ghostery, as for previous mea-
taintrackersfromSwedenorFinland.Thesameobservation sures.
applies to Austria as we noticed already, as well as other TheresultsclearlyshowthattheUSareevenmorepresent
countriesinCentralEurope,suchasHungary,Slovakiaand on the invisible than on the visible Web. Most of the
theCzechRepublic.Therefore,withtheexceptionoftheUS, tracker points (in red) are located on the right lower corner
mosttrackersareregional. oftheplot.ThisindicatesthatthepercentageofUStrackers
is large in most considered countries, while the percentage CC P P
US local
oflocaltrackersisusuallysmaller,whilethedistributionof RU 0.39/0.36 0.49/0.52
theproportionofthesitesismorebalancedbetweentheUS CN 0.69/0.68 0.29/0.31
andlocalsites. HU 0.57/0.53 0.21/0.23
PL 0.65/0.63 0.17/0.17
NO 0.68/0.62 0.03/0.04
FR 0.71/0.67 0.19/0.22
NG 0.9/0.86 0/0
Table10:P andP fortrackers
US local
sites are excluded. The results contrast with the results of
the previous table, and clearly show that the percentage of
US trackers is larger that the percentage of local trackers,
exceptforRussia.
Finally it is interesting to compare Russia and China with
respecttotheproportionoflocalsitesandtrackers.
Figure10:SitesvsTrackers(Ghostery)
This finding is further confirmed with the results shown on
Figure 10, which presents the scatterplots for sites against
trackers:therelationshipsbetweenUSsites(blacktriangles)
Figure11:SitesandtrackersinRussia
(resp.localsites(redstars))andthecorrespondingtrackers
forGhostery.Theredstarsareonthelowerpartofthefigure,
corresponding to a small percentage of local trackers for Figure 11 shows how Russia manages to have both more
most considered countries even with a large proportion of russian sites and trackers at home, although somehow with
local sites. Whereas, the black triangles are on the higher thesameorderofmagnitudesasUSsitesandtrackers,while
partofthefigure,correspondingtoalargepercentageofUS Figure 12 shows that China has mostly Chinese sites, but
trackers. mostlyUStrackers.
Letusnowconsideragainanin-depthAnalysisofSelected
Countries.WeconsidertheUSandlocalproportionofsites
and trackers for seven selected countries, namely Russia,
China,Hungary,Poland,Norway,FranceandNigeria.
CC P P
US local
RU 0.23 0.54 Figure12:SitesandtrackersinChina
CN 0.12 0.86
HU 0.26 0.56
PL 0.21 0.69 5. Conclusion
NO 0.37 0.51
Westudiedtheglobaldistributionandproliferationofthird-
FR 0.34 0.61
party resources on the most popular sites in various coun-
NG 0.72 0.04
tries. Our research reveals very different strategies and/or
capacities to harvest local as well as global data, both on
Table9:P andP forsites
US local
thevisibleandtheinvisibleWeb.
• InEurope,mostcountrieshavetwiceasmanylocalsites
Table9displaystherespectivepercentagesofUSandlocal
as US sites amongst their top 100 most popular sites,
sites.Itshowsthatlocalsitesareoftendominant,exceptfor
although the top sites are mostly US, and the trackers
somedevelopingcountriessuchasNigeria.
are mostly US as well, followed by local and regional
Table10displaystherespectivepercentagesofUSandlocal trackers, thus leading to an important flow of data from
trackers. The second value shows the results when the US EuropetotheUS.