arXiv:1907.12860v2 [cs.CR] 2 Jun 2020
Clash of the Trackers: Measuring the Evolution of
the Online Tracking Ecosystem
Konstantinos Solomos∗, Panagiotis Ilia∗, Sotiris Ioannidis†, Nicolas Kourtellis‡
∗University of Illinois at Chicago, USA; †FORTH,Greece; ‡Telefonica Research, Spain
∗{ksolom6,pilia}@uic.edu; †sotiris@ics.forth.gr; ‡nicolas.kourtellis@telefonica.com
Abstract—Websites are constantly adapting the methods used,
and intensity with which they track online visitors. However, the
wide-range enforcements of regulations such us GDPR and e-
Privacy has forced websites serving EU-based online visitors to
eliminate, or at least reduce, such tracking activity, given they
receive proper user consent. Thus, it is important to analyze
the aftermath of such policies, record the evolution of this
tracking activity, and assess the overall “privacy health” of the
Web ecosystem. This work makes a significant step towards
this direction. In this paper, we analyze the ecosystem of 3rd-
parties embedded in top websites, which amass the majority
of online tracking, through six time snapshots taken every few
months apart, in the duration of the last few years. We perform
this analysis in three ways: 1) by looking into the network
activity that 3rd-parties impose on each publisher hosting them,
2) by constructing a bipartite graph of “publisher-to-tracker”,
connecting 3rd-parties with their publishers, 3) by constructing
a “tracker-to-tracker” graph connecting 3rd-parties who are
commonly found in publishers. We record significant changes
through time in number of trackers, traffic induced in publishers,
embeddedness of trackers in publishers, popularity and mixture
of trackers across publishers. In the last level of analysis, we dig
deeper and look into the interconnectivity of trackers, and how
this relates to potential cookie synchronization activity.
I. INTRO DUC TIO N
Online users’ privacy is constantly violated by leaks of
their PII to unauthorized parties. Users lose their anonymity
due to web tracking via cookies [1], [2], device or browser
fingerprinting [3]–[5] and cookie synchronization [6]–[10]. In
the last few years, different legislations and directives such
as the e-Privacy [11], the General Data Protection Regulation
(GDPR) [12] in EU, and the California Consumer Privacy Act
(CCPA) [13] in USA, were introduced in an effort to increase
transparency in user tracking and help users with personal
data management and privacy protection. In particular, GDPR
forces websites to stop such activity or receive informed
consent from their online visitors for any potential tracking,
data collection and processing they may do, and also for any
data sharing they may do with 3rd-parties.
Recent studies [14]–[18] have investigated the aftermath
of GDPR and its effects on the online tracking ecosystem,
and how websites may have reduced their tracking activity.
However, the web tracking ecosystem is constantly evolving
and adapting to new blocking methods. Trackers continue
their aggressive activity across multiple domains, at times
unbalanced depending on the type of website [19], and even
employ cookie-less, machine learning-based methods to track
users across their different devices [20]–[22], all in the name
of “more effective ad-campaigns”.
In this work, we build on previous studies and methods on
web tracking, and perform a first of its kind longitudinal study
to measure this ecosystem’s changes over the last few years.
We perform this analysis in three levels through time, using six
crawls of top Alexa websites in time snapshots of a few months
apart. First, we look into the network activity that 3rd-parties
impose on each publisher (1st-party) hosting them. With this
first-level analysis, we confirm existing reports that claim
reduction in tracking, by measuring generic HTTP network
activity from 3rd-parties. Second, we construct “publisher-to-
tracker” (P T ) bipartite graphs, connecting 3rd-parties with
their publishers. With this second level of analysis, we employ
graph mining tools and metrics such as clustering coefficient,
density, degree centrality, coreness, etc., to study the graph
properties of the six bipartite graphs. We find that the structure
of the tracking ecosystem with respect to embeddedness in
publishers has not changed significantly through time, and that
top degree centrality trackers such as Google’s suite (google-
analytics, doubleclick, etc.), Facebook, AppNexus, Criteo, etc.,
dominate the ecosystem in all time snapshots, without losing
their market share of publishers. Moreover, we also identify
top betweenness trackers such as Twitter and Adobe, which are
not in the typical top degree list but have embedded themselves
in central positions in the web ecosystem.
Third, we construct a “tracker-to-tracker” graph (T T ), con-
necting 3rd-parties that are commonly found in publishers.
With this analysis, we construct T T pairs from the P T graphs
that can reveal potential collaborations between 3rd-parties.
We compare these pairs with ground truth data from confirmed
data sharing flows of cookie syncing (CS) pairs, in two
different CS datasets from past studies. Proper cookie syncing
flows between trackers are not easy to get as they require
activity from real users or persona-based automated browsing
to trigger the CS mechanism. Interestingly, we identify a
high overlap between CS and T T pairs (∼47%-81% when
compared to previous ground truth CS datasets). We propose
that such information flows can be inferred from the T T
graphs with reduced cost in deployment and measurements,
as they require only web crawling of 1st-parties.978-3-903176-27-0 c
2020 IFIP
II. BACKGROUND AND RELATED WOR K
A. Web Tracking & Graph Modeling
Many works have focused on analyzing the web tracking
ecosystem, its internal mechanisms and their impact on user’s
privacy. One of the first studies on web tracking, by Mayer
and Mitchell [23], investigated which information is collected
by 3rd-parties and how users can be identified. Roesner et
al. [1] studied the various tracking behaviors and measured
the prevalence of trackers while Falahrastegar et al. [24]
measured the existence of cookie synchronization trackers.
Papadopoulos et al. [10] used a heuristic-based mechanism to
detect information exchanged between advertisers and found
that 97% of the users are exposed to cookie syncing at least
once, and that ad-related entities participate in more than 75%
of the overall cookie synchronization.
A plethora of studies investigate stateful tracking techniques
(e.g., [1], [2], [6], [25]), and stateless techniques such as
browser fingerprinting [3]–[5], [8], [26], [27]. Acar et al. [8]
investigated the prevalence of “evercookies” and the effects of
cookie respawning in combination with cookie syncing. Engle-
hardt and Narayanan [2] conducted a large scale measurement
study to quantify both stateful and stateless tracking in the
web, and Lerner et al. [28] conducted a longitudinal study of
3rd-party behaviors and found that tracking has increased in
prevalence and complexity over time.
In general, by studying the graph network properties, one is
able to understand the characteristics of the tracking entities,
and dissect the ecosystem and its inner mechanisms. In that
direction, Kalavri et al. [29] built a 2-mode bipartite graph
based on real user traffic logs, and focused their analysis on
the communities formed by the graph vertices. Their analysis
showed that trackers are well connected to each other, since
94% of them are in the largest connected component. Urban et
al. [30] collected behavioral data from emulated users located
in 20 EU countries and created a cookie synchronization
graph that connects 3rd-parties that share information. They
reported that the number of trackers and the number of direct
syncing connections decreased through time, since fewer 3rd-
parties are present in the publisher domains (40% less syncing
connections). Also, based on the properties of their graph,
they found that the structure of the ecosystem did not change
significantly. Similarly, Bashir et al. [7] constructed a cookie
syncing graph and associated the graph metrics with the
existence of different tracking domains.
B. GDPR Enforcement and Web Privacy
The GDPR [12] is a regulatory initiative by the EU to
harmonize data protection laws between its member states,
and specifies under which circumstances personal data may be
processed. Since the regulation directly affects the web and
the online advertising and tracking ecosystem, recent works
have focused on investigating the state of the ecosystem, the
evolution of the privacy policies and their impact on user’s
privacy from different perspectives [16], [18], [31]. Iordanou
et al. [14] collected data from users across EU and identified
TABLE I: Timeline of web crawls using top Alexa lists, and
number of unique 1st- and 3rd-party domains we detected in
each crawl. The (*) shows which snapshots were collected
after the enforcement of GDPR. The overlap row indicates
the intersection of common domains across all snapshots.
Dataset Alexa Ranks 1st-parties 3rd-parties
September 2017 10K 8311 848
January 2018 30K 29444 1036
May 2018 80K 73493 1096
(*) June 2018 80K 73813 1068
(*) November 2018 65K 61287 1002
(*) April 2019 65K 59662 819
Overlap 10K 5100 527
the directions of tracking flows inside EU. They reported that
85% of the tracking flows terminate in servers located inside
the EU, and identified that the most sensitive types of user
information based on GDPR that is being tracked is health,
sex orientation and politics. Degeling et al. [15] quantified the
changes of privacy policies on the Top-500 sites of the 28 EU
countries. They found that, in total, 85% of the websites have
a privacy policy, and that GDPR did not significantly change
the way 3rd-party cookies are used. In their most recent work,
Sorenser et. al. [17] measured the changes on the presence
of 3rd-parties, before and after GDPR enforcement, in 1200
popular websites across EU. Their study shows that there were
not significant changes in the general state of the web, and that
GDPR had a potential effect only on specific types of websites.
III. DATA COLLE CTI ON
Since our purpose is to conduct a longitudinal study on
the evolution of tracking, and how it was affected by the
recent EU regulations, we collected historical data covering a
period of almost 2 years (September 2017 to April 2019). For
collecting these data we used the OpenWPM framework [2] to
crawl multiple websites via scripted browsers, and we stored
all the HTTP(s) incoming and outgoing requests. We also
logged the cookies set by JavaScript, and stored various other
crawl-related data (i.e., time of visit, HTML files, etc.). During
the crawling, we did not set the “Do Not Track” flag, we
configured our browser to accept all connections, and we
empirically set the timeout for a website to respond to 30
seconds. For not introducing any unnecessary complexity and
overhead during the crawls, we do not perform any browsing to
visit deeper than the main page of each website. We deployed
the framework on a single computer at an EU academic
institution, having a unique IP address, thus, avoiding any
content biases or any type of location based discrimination. As
a website input corpus, we used the Alexa Top 100K list [32],
and based on our available resources, we crawled each time
the top websites covering different subsets of the list. Starting
from September 2017, we repeated our crawls approximately
every 5 to 6 months. We also performed two consecutive
crawls just before and after the GDPR enforcement, i.e., mid
May 2018 and beginning of June 2018.
0
0.25
0.5
0.75
1
0 20 40 60 80 100
CDF
Trackers per Publisher
Sep17
Jan18
May18
Jun18
Nov18
Apr19
0.7
0.8
0.9
20 30
0.7
0.8
0.9
20 30
(a)
0
0.25
0.5
0.75
1
0 100 200 300 400 500 600
CDF
3rd-Party HTTP(s) Requests per Publisher
Sep17
Jan18
May18
Jun18
Nov18
Apr19
0.7
0.8
0.9
80 100 120 140
0.7
0.8
0.9
80 100 120 140
(b)
0
0.25
0.5
0.75
1
0 10 20 30 40 50 60
CDF
3rd-Party HTTP(s) Requests per Publisher
Sep17
Jan18
May18
Jun18
Nov18
Apr19
0.7
0.8
0.9
2 4 6 8
0.7
0.8
0.9
2 4 6 8
(c)
0
0.25
0.5
0.75
1
0 5 10 15 20 25
CDF
Unique 3rd-Party Libraries per Publisher
Sep17
Jan18
May18
Jun18
Nov18
Apr19
0.6
0.7
0.8
6 8 10
0.6
0.7
0.8
6 8 10
(d)
Fig. 1: These figures, from left to right, show (a) the number of unique 3rd-party trackers, (b) the total number of 3rd-party
HTTP(s) requests, (c) the average number of 3rd-party HTTP(s) requests, and (d) the number of embedded 3rd-party libraries.
In each snapshot of our dataset, we define two different
entities: (i) Publishers, which are the websites that the users
explicitly visit (i.e., 1st-parties), and (ii) Trackers, which refer
to the 3rd-party domains that are embedded within the visited
pages (i.e., domains from which resources are fetched, that
set cookies, serve content, etc.). We use the most recent, by
the time of each crawling, Disconnect List [33] (a popular
browser ad-blocking list), in order to identify which requests
are directed towards trackers. In this way, we were able to
accurately detect the presence of trackers in each page and
to identify their behaviors. Since tracking domains change
frequently, and only a subset of them might remain constant
through time, the list is constantly updated to capture such
changes. Furthermore, we only consider publishers that embed
at least one tracker in their page. For simplicity, we inter-
changeably refer to 1st-parties or publishers, and to 3rd-parties
or trackers in the rest of the paper. A detailed description of
our datasets is presented in Table I.
IV. NET WOR K ACT I VI T Y OF TR AC KE RS
Using the aforementioned six time snapshots, we perform an
analysis on the network activity of trackers that are embedded
in the publishers crawled. As shown in Table I, the snapshots
in our datasets include different subsets of the top websites.
Also, considering that the Alexa list is volatile and that it is
constantly changing [34], we focus our analysis on a subset
of 5100 publishers that are common across all time snapshots.
HTTPS Adoption. Before digging deeper into the patterns
and the effects of the online tracking ecosystem, here we
make a preliminary measurement of the HTTPS adoption by
the publishers. In our first snapshot (i.e., September 2017)
we found that 3144 (61.64%) of the publishers were using
HTTPS. This percentage is reasonable considering that our
corpus was formed by publishers that rank among the Top-10K
of the Alexa list, and that those tend to be more regulation
compliant and probably privacy-preserving, as also reported
in previous work [35]. In the snapshot of January 2018, we
found 944 additional publishers (80.15% of our dataset) that
adopted HTTPS, while ∼92% of the publishers adopted it by
April 2019, validating the shift towards more secure practices.
Tracking Activity. Figure 1 summarizes different dimensions
of the tracking activity of 3rd-parties that are present in the
publishers of our datasets. The unique number of trackers
embedded in each publisher (i.e., based on the outgoing HTTP
requests observed) is shown in Figure 1(a). We observe that,
in general, there is a clear decrease in the number of trackers
through time. Specifically, until January 2018, 50% of the
publishers communicated with almost 10 individual trackers
in every visit, but this changed by the end of April 2019,
when this number dropped to almost half, reaching the lowest
measured value. This might not be a typical trend of the
ecosystem, and this “momentary” elimination of trackers could
be caused by the enforcement of GDPR in May 2018, but
further investigation is needed to see whether publishers will
go back to the same levels of tracking as before GDPR.
In Figure 1(b) we report the total number of 3rd-party
requests per publisher. We observe that the trend on the
distributions is similar: through time there is a decline in the
average number of 3rd-party requests per publisher. Summa-
rizing the previous two measures, in Figure 1(c) we show the
average number of 3rd-party requests per publisher. In theory,
since both the number of trackers and the total requests are
in decline, their ratio (i.e., average) should also be reduced
through time. This trend is observed for the majority of
websites, but for a small set of the publishers in the last
snapshot (∼10%), this number increases.
Since 3rd-party requests may be issued for different reasons
(e.g., cookie delivery, tracking pixels, content, ad-libraries),
in Figure 1(d), we report the unique number of trackers
that deliver 3rd-party libraries (i.e., JS) to publishers. For
this measurement, we consider that connections to different
subdomains, e.g., subdomain1.domain.com/ad-library.js and
subdomain2.domain.com/ad-library.js, result in a connection
to a unique 3rd-party provider. From the results, we note that
by 2017, up to 50% of the publishers communicated with at
least 5 different providers, and by 2019 this number decreased
to 4. Also, there is a reduction in the plurality of the embedded
providers (e.g., Amazon and Tapad were not present in the
latest snapshots), which potentially results in higher tracking
activity by the remaining entities, or increases the chances of
cookie sharing and synchronization. Previous works on this
topic (e.g., [14], [30], [36]) also report similar decline in
the number and frequency of trackers. Specifically, in [36]
they reported that fewer 3rd-parties are present in specific
categories of websites. These facts might be a side effect of the
GDPR being enforced, or other effects of the general evolution
of the Web tracking ecosystem.
V. PU BL ISH ERS & TRACK ERS : TH E PT GR APH
Graph Construction. We follow a similar approach to Kalavri
et al.’s [29] for constructing our graphs. We create a set of 2-
mode graphs of the publishers and their associated trackers,
where the edges of each graph connect vertices of different
modes. In this graph, a publisher can connect to multiple
trackers, and a tracker can connect to multiple publishers.
We represent all the domains that a browser requests as a 2-
mode graph, by creating a set of mappings. The VPrepresents
the set of websites (publishers) a user visits, and accordingly
VTis the set of trackers embedded in publishers. We also
define Ewas the set of weighted edges connecting vertices of
the two different modes, and w= (i, j)as the weight of the
edge connecting tracker iwith publisher j. Furthermore, we
go beyond the state-of-art (i.e., [29]) and add weights on the
edges, to represent the number of HTTP(s) requests between
a publisher and a tracker. The weight w=(i, j), encodes the
number of times that a tracker icommunicated via HTTP
requests with a publisher j.
Data Filtering & Graph Metrics. As reported in Section III,
in each snapshot, we crawled a subset of the top Alexa list.
Since we want to create a connected representation of the
bipartite graphs, we use the Largest Connected Component
(LCC) of each graph. We found that there are some isolated
groups of nodes, that include websites that communicated with
one or two different, but not popular trackers; we exclude such
isolated groups from the graphs. In each of our final graphs,
the LCCs contain on average ∼99% of the publishers and
∼95% of trackers of the originally crawled lists.
This type of connected graphs allows us to apply various
graph metrics to quantify their properties, and compare them
across time. In our analysis, we use graph metrics similar
to those of previous works on this topic (e.g., [7], [29],
[30]). Specifically we compute the Density and Diameter to
study the inner structure of each network and the connectivity
between their nodes. Other metrics like the Average Clustering
Coefficient, Degree Centrality and Betweenness Centrality
reveal the properties of each node in terms of the shortest
paths that pass through the specific edge, and the number of
connected neighbors. Finally, we also use Coreness-Periphery,
a metric not studied before in the literature, that measures the
importance of each node in terms of how “involved or core”
it is inside the network. Detailed definitions of these metrics
can be found in [37]–[39]. The generated graph models and
their properties for each snapshots are given in Table II.
A. Stability of PT graph properties over time
The differences in the number of nodes and edges between
these graphs reflect the different number of visited publishers
in each snapshot. In general, the average clustering coefficient
measures the degree with which the nodes of a graph tend to
cluster together (i.e., tend to close triangles between triplets of
TABLE II: PT graph characteristics: number of vertices (N);
edges (E); normalized average weight per edge (W); average
clustering coefficient (CC); density (DE); diameter (D).
Dataset N E W CC DE D
Sep17 5710 74037 0.0013 0.024 0.022 7
Jan18 5688 86875 0.0012 0.024 0.027 7
May18 5678 81717 0.0013 0.021 0.026 8
Jun18 5654 76077 0.0013 0.023 0.025 7
Nov18 5636 71481 0.0013 0.026 0.024 8
Apr19 5602 72722 0.0011 0.022 0.025 4
nodes, or quadruples in bipartite graphs). The low value of this
metric on each of the graphs is associated with the bipartite
connectivity between the sets of nodes [40]. This metric, in
conjunction with the low density, reveals sparse connections
between the different groups of nodes.
Regarding the number of trackers, there is a stable reduction
through time, matching the measured decrease in the average
number of trackers, as discussed in Section IV. In general, the
characteristics and distance metrics computed on the graphs
reveal a consistent structure of the ecosystem during the focal
period of our analysis, except for the most recent snapshot
where the diameter reaches its lowest value, while the number
of edges increases. This trend captures an increase in the
connectivity of the graph nodes, pointing to tracker nodes
being closer in the graph, as also discussed next.
Stability of Publishers vs. Trackers. We measure the degree,
betweenness centrality and coreness-periphery of each node
in our bipartite graphs, in order to quantify the backbone
structure of each graph in terms of connectivity and centrality
of nodes. Figure 2, plots the distribution for each metric, as
computed on each snapshot for publishers and trackers.
Focusing our analysis on Figures 2a and 2d, the degree
centrality for publishers is an order of magnitude lower than
the one for trackers. Also, 50% of the publishers’ degree
centrality is ≤0.025, with only ∼5% of the publishers having
more than 0.01. Conversely, tracker nodes have higher degree
centrality scores, with ∼5% measured at ≥0.15, hinting to the
fact of the well-known trackers that cover approximately all
the publishers (e.g., Google and Facebook). Some examples
of publishers with high degree centrality are telegraph.co.uk,
newyorker.com and rollingstone.com, and examples of track-
ers with high degree centrality include google-analytics.com,
criteo.com, and facebook.com. By definition, since we are
analyzing a bipartite network constructed by two sets of nodes
with different sizes and edge weights, it is reasonable for the
tracking nodes to be more central in the network structure.
The betweenness centrality for publishers and trackers are
given in Figures 2b and 2e, respectively. Publishers’ low
scores are expected, as betweenness centrality measures the
extent to which a node lies on paths between other nodes.
Obviously, publishers are not connected to each other but
only with trackers. Trackers that are highly active, central
and well known, are found on the tail of the distribution,
0
0.25
0.5
0.75
1
0 0.005 0.01 0.015
CDF
Degree Centrality
Sep17
Jan18
May18
Jun18
Nov18
Apr19
(a)
0
0.25
0.5
0.75
1
0 0.0002 0.0004 0.0006 0.0008 0.001
CDF
Betweenness Centrality
Sep17
Jan18
May18
Jun18
Nov18
Apr19
(b)
0
0.25
0.5
0.75
1
0 0.25 0.5 0.75 1
CDF
Coreness Periphery Score
All Snapshots
(c)
0
0.25
0.5
0.75
1
0 0.1 0.2 0.3 0.4 0.5
CDF
Degree Centrality
Sep17
Jan18
May18
Jun18
Nov18
Apr19
0.6
0.7
0.8
0.9
0.01 0.02 0.03
0.6
0.7
0.8
0.9
0.01 0.02 0.03
(d)
0
0.25
0.5
0.75
1
0 0.01 0.02 0.03 0.04 0.05
CDF
Betweenness Centrality
Sep17
Jan18
May18
Jun18
Nov18
Apr19
0.9
1
0 0.01 0.02
0.9
1
0 0.01 0.02
(e)
0
0.25
0.5
0.75
1
0 0.25 0.5 0.75 1
CDF
Coreness Periphery Score
Sep17
Jan18
May18
Jun18
Nov18
Apr19
(f)
Fig. 2: Top row shows metrics for Publishers and bottom row for Trackers: (a) and (d) Normalized degree centrality; (b) and
(e) Betweenness Centrality; (c) and (f) Coreness periphery for each PT Graph.
with scores ≥0.02. Example of publishers with high be-
tweenness are starbucks.com,livescore.com and cnn.com, and
accordingly trackers with high betweenness are moatads.com,
instagram.com and scorecardresearch.com.
Finally, when comparing coreness scores for publishers and
trackers, in Figures 2c and 2f, respectively, trackers tend
to occupy positions in the graph with higher coreness than
publishers. Interestingly, we observe two classes of nodes in
each graph, with periphery nodes being in the beginning of
each CDF (up to ∼0.25 coreness score), and core nodes being
in the end of each CDF, with coreness score ≥0.75. Well-
known publishers such as sfgate.com,sport.es and indian-
express.com, and trackers such as yandex.ru,adroll.com and
amazon-adsystem.com. have high coreness-periphery scores.
B. How do these centrality metrics correlate?
After studying the inner structure of the bipartite networks,
we evaluate the relationship between the centrality of nodes
for these three metrics. This study will help us understand if
the nodes (trackers or publishers) tend to be top (or bottom) in
all metrics at the same time, or if there is some disassociation
between these metrics. Such disassociation can reveal outliers
of nodes who are high in one metric but low in another.
For this reason, we compute the Pearson Correlation score
between the distributions of Degree Centrality, Betweenness
Centrality and Coreness-Periphery for the trackers and pub-
lishers, independently. A detailed report on the correlation
scores is given in Table III. In general, we find a strong
association between all the distributions with high confidence
TABLE III: Pearson Correlation Coefficient between the distri-
bution of: Degree & Betweenness Centrality (DC-BC), Degree
Centrality & Coreness Periphery (DC-CP), and Betweenness
Centrality & Coreness Periphery (BC-CP), for each set of
the PT graph for common publishers across snapshots. All
results reported are statistically significant at p−value<0.009.
The publishers are ranked according to the Alexa list, while
trackers according to their Degree Centrality.
Dataset Publishers Trackers
∼DC-BC ∼DC-CP ∼BC-CP ∼DC-BC ∼DC-CP ∼BC-CP
Sep17 0.50 0.69 0.26 0.75 0.47 0.26
Jan18 0.43 0.83 0.3 0.73 0.50 0.24
May18 0.47 0.75 0.20 0.73 0.50 0.24
Jun18 0.46 0.74 0.24 0.75 0.48 0.25
Nov18 0.42 0.72 0.19 0.77 0.49 0.26
Apr19 0.43 0.71 0.23 0.54 0.47 0.18
level. Trackers have higher correlation score in the Degree-
Betweenness comparison, hinting their importance and central
role on the network, regardless of the metric used. On the
contrary, the correlation scores for Degree-Coreness, for the
publishers, are measured higher than trackers, again validating
the importance of the publishers who tend to have high and im-
portant position in the network structure. Finally, across both
types of nodes, Betweenness-Coreness correlation scores are
lower, pointing to a disassociation between the two measures.
TABLE IV: Top-25 Trackers ranked by Degree Centrality and
labeled under the umbrella of company/organization and the
average percentage of coverage in publishers through time.
(*): Trackers who were Top-25 trackers across all snapshots.
(+/-): Trackers who were Top-25 in one or more snapshots,
but their rank decreased through time below Top-25.
(+): Set of trackers that were not part of the Top-25 in the first
snapshot, and they climbed in the Top-25 through time.
Tracker Organization Publishers(%)
(*) google-analytics.com Google 81.0
(*) doubleclick.net 70.0
(*) google.com 51.0
(*) googleapis.com 57.5
(*) googletagmanager.com 36.5
(*) facebook.com Facebook 44.5
(*) facebook.net 41.5
(*) googletagservices.com Google 28.0
(*) gstatic.com 44
(*) googlesyndication.com 28.3
(*) googleadservices.com 19.0
(*) cloudfront.net Amazon 18.0
(*) adnxs.com App Nexus 18.0
(*) criteo.com 13.0
(*) criteo.net Criteo 13
(*) scorecardresearch.com comScore 12.5
(*) twitter.com Twitter 20.0
(*) rubiconproject.com Google 12.5
(*) pubmatic.com Pubmatic 11
(*) openx.net OpenX 8.5
(*) casalemedia.com Casale Media 9.0
(*) advertising.com Verizon Media 7.0
(+/-) quantserve.com Quantcast 9.0
(+/-) adsrvr.org The Trade Desk 9.0
(+/-) taboola.com Taboola, Inc 7.0
(+/-) nr-data.net New Relic 8.0
(+/-) 2mdn.net Google 6.0
(+/-) bluekai.com BlueKai 8.0
(+) alexametrics.com Amazon 4.0
(+) demdex.net Adobe 7.0
(+) newrelic.com New Relic 8.0
C. What is the ecosystem’s current state?
After analyzing the graphs’ inner properties, we want to
quantify the importance of the top tracking nodes for the
ecosystem and their impact to user privacy. To have a clear
view of the most important trackers at each time of our crawl,
we measured their degree centrality and ranked them accord-
ingly. A complete report about the Top-25 trackers across
time is given in Table IV. Interestingly, the Top-22 trackers
remained the same across all snapshots with minor fluctuations
on the internal ranking. The most important trackers contain
the “big”, well-known entities of the ad-industry, such us
Google,Facebook,Twitter and Criteo, as well as some smaller,
but well established such as Bluekai and Taboola. This table
also reports those trackers that gradually climbed into the Top
list, illustrating the plurality of the tracking ecosystem.
TABLE V: Top-25 Trackers ranked by Betweenness Centrality
(BC) and labeled under the umbrella of company/organization,
with their BC score and coverage in publishers by April 2019.
We highlight the trackers not present in Table IV.
Tracker Organization BC % Publishers
googletagmanager.com Google 0.077 46.55
doubleclick.net 0.012 72.15
googleadservices.com 0.008 19.59
googletagservices.com 0.006 28.82
gstatic.com 0.005 44.75
cloudfront.net Amazon 0.005 14.52
newrelic.com New Relic 0.004 7.63
rlcdn.com Live Ramp 0.003 8.66
pubmatic.com Pubmatic 0.003 13.1
google.com Google 0.030 59.09
nr-data.net New Relic 0.002 7.61
facebook.com Facebook 0.002 44.01
everesttech.net Adobe 0.002 3.52
casalemedia.com Casale Media 0.002 7.48
alexametrics.com Amazon 0.002 4.46
ads-twitter.com Twitter 0.002 4.72
adsrvr.org Trade Desk 0.002 5.67
adnxs.com App Nexus 0.002 13.83
twitter.com Twitter 0.001 13.32
rubiconproject.com Google 0.001 10.15
quantcount.com Quantcast 0.001 4.11
openx.net OpenX 0.001 8.59
linkedin.com Microsoft 0.001 4.71
advertising.com Verizon Media 0.001 7.23
Furthermore, in Table V, we make a similar investigation for
trackers’ betweenness centrality. We note that the top of the list
is populated by similar trackers. However, new entities such
as linkedin.com,ads-twitter.com and everesttech.net emerge,
demonstrating a central position in the ecosystem with respect
to mediating flows between distant parts in the ecosystem.
Overall, the almost immutable list of top trackers in either of
the two metrics points to the fact that the GDPR enforcement
had no effect on them, either in their importance in the web
tracking ecosystem, or their coverage across websites. Also,
in the previous sections, we found that trackers are present in
more websites (as time passed by), but at the same time, web
requests have been reduced. We can conclude that there may
be a “shift” of publishers on the type of business relationships
they make with the well-known and GDPR-compliant trackers.
VI. TRAC KE R T O TRACK ER: THE TT GRA P H
Graph Construction. We build tracker-to-tracker graphs that
are undirected but weighted, T T = (VT T , ET T ), and originate
from their corresponding PT graphs. Similarly to Section V,
we create different sets of nodes that have specific properties.
In our TT graphs, VT T represents the set of trackers embedded
in publishers, and ET T is the set of weighted edges connecting
two trackers, if and only if both trackers coexist in at least
two different publishers. Also, the weight w=(i,j), encodes the
number of publishers that tracker iand jcoexisted. A detailed
description of our T T graphs is presented in Table VI.
TABLE VI: Characteristics of TT graphs produced from the
PT graphs with common publishers across all snapshots.
Dataset N E W CC DE DD
Sep17 815 63177 0.004 0.69 0.19 4
Jan18 774 53325 0.006 0.67 0.17 4
May18 846 76686 0.005 0.72 0.21 4
Jun18 824 69145 0.005 0.70 0.20 4
Nov18 834 72265 0.005 0.71 0.20 4
Apr19 841 74012 0.005 0.71 0.20 4
The T T graph of each snapshot has a fairly dense structure
(average Density 0.19 −0.21). The number of nodes, as well
as the number of edges, are comparable through snapshots,
which is reasonable since we focused only on the common
publishers in the dataset and extracted the trackers present in
each snapshot. Interestingly, in May 2018 (before GDPR) the
number of edges reached a maximum which was not surpassed
in subsequent snapshots. In general, in all T T graphs, trackers
are well connected and clustered with each other (average
clustering coefficient 0.67 −0.72). These properties of the T T
graphs highlight the dense structure of the tracking ecosystem,
and how 3rd-parties potentially share user’s information.
A. Are cookie synchronization pairs present in TT graphs?
Since the purpose of the T T graphs is to study potential data
sharing among trackers, we need to compare the constructed
pairs with existing data that already measure data sharing flows
(i.e., ground truth datasets). We received access to two CS
datasets provided by Papadopoulos et. al. [10] and Bashir et
al. [7]. These datasets contain pairs of 3rd-parties performing
CS while real users [10], or crawlers [7] browsed the web.
Interestingly, the dataset from [10] also includes a normal-
ized frequency on each pair, encoding the number of times the
two entities of the pair shared information. Following a similar
representation as with the T T graphs, we create 2 undirected
CS graphs, CS=(VCS , EC S ), with weighted edges for the data
from [10], i.e., w(i,j) is the number of times that a pair of
trackers (i, j)performed information exchange. The first CS
graph from [10] has 4656 trackers and 8582 edges connecting
them, whereas the second CS graph from [7] has 59 trackers
and 200 edges connecting them. To investigate the existence
of CS pairs into T T pairs, we define the following sets:
•ECS for the set of edges in a CS graph; ET T for the set
of edges in a T T graph.
•¬CS for the set of non-edges in a CS graph; ¬T T the
set of non-edges in a T T graph.
We also define the following overlaps of the above sets:
Ocommon =ECS ∩ET T ,O¬C S =¬CS ∩ET T and O¬T T
=ECS ∩ ¬T T . To have an accurate measurement between the
different overlaps, we filter the edges and store only those that
are part of the common trackers between each CS and each
T T graph. A detailed report on the percentages of overlap
between each of the two CS graphs is given in Tables VII
and VIII.
TABLE VII: Percentage of overlap between the different sets
of trackers for the CS edges extracted from [10]. We refer to
the common number of trackers as |N|.
Dataset |N| |ECS | |ET T |Ocommon O¬C S O¬T T
Sep17 226 3015 28631 59.70 49.10 1.70
Jan18 226 3015 28631 59.70 49.10 1.70
May18 226 3024 27353 58.10 46.80 1.90
Jun18 222 3003 24943 55.30 44.70 2.30
Nov18 214 2820 20976 52.20 39.90 2.20
Apr19 210 2929 18669 47.30 37.90 3.30
TABLE VIII: Percentage of overlap between the different sets
of trackers for the CS edges extracted from [7]. We refer to
the common number of trackers as |N|. The O¬T T value was
measured 0% across all TT graphs.
Dataset |N| |ECS | |ET T |Ocommon O¬C S
Sep17 42 104 7321 80.80 76.40
Jan18 41 102 7169 80.40 75.90
May18 41 104 6860 78.90 71.80
Jun18 41 104 6389 73.10 69.20
Nov18 40 104 5896 69.30 68.50
Apr19 39 104 5309 64.40 63.20
According to Table VII, the overlap between CS and T T
edges across snapshots is 47 −60%. In the smaller CS dataset
(Table VIII), this overlap is even higher, ranging to 64 −81%.
We remind the reader that this overlap is very high, considering
all possible ways that such a graph could have been wired
just by chance. Considering that the T T graphs were built
artificially, this high overlap gives us an indication about the
“nature” of CS pairs, and how such data sharing flows can
be found in a T T graph. This is a crucial finding: we can
detect potentially collaborating pairs of trackers, without the
need to deploy infrastructure to collect real users’ data, or
train artificial personas to collect CS activity. Also, the overlap
is reduced over time possibly because of two reasons. First,
the ground truth datasets were collected closer to our initial
crawl (2017). Second, as reported by [30], there has been a
reduction in CSs after GDPR. Therefore, the CS ecosystem
is changing, and 3rd-parties that engaged in CS in earlier
snapshots may not be doing so later, affecting the measured
pair overlap.s Moreover, since the first CS graph was weighted,
we investigated how well T T edges that overlap with CS edges
cover the weights distribution. That is, how representative T T
edges are of CS edges, with respect to weights. We found that
the common T T edges cover well (∼75%) of the distribution
of the CS weights. Finally, 2% of the overlapping T T edges
are the 5% most frequent edges on the CS graph, hinting again
that synthetic data can capture confirmed real-world CS cases.
VII. DI SC USS ION & LIMI TATIO NS
We presented a first of its kind longitudinal study on the
changes of the web tracking ecosystem during the last 2-3
years. The analysis through time was performed using six
crawls of the top Alexa websites, with snapshots collected
a few months apart. In the first level of analysis, we focused
on network-level traffic between publishers and trackers. We
found that there are fewer trackers embedded in the websites
through time, with a reduction of 9% for the median site, and
10% for the 90% percentile site. Moreover, there are fewer
3rd-party HTTP(s) requests, with a reduction of 17% for the
median and 25% for the 90% percentile site.
In the second level of analysis, we constructed bipartite
graphs of publishers connected with their trackers (P T ) and
studied the P T graph properties through time. In summary, the
network structure of the tracking ecosystem and how trackers
are embedded in publishers remained the same through time.
The same 3rd-parties that existed through time have been
forced to cover more websites, and especially the top central
trackers. Also, top trackers in terms of publisher coverage and
centrality in P T graphs (e.g., google-analytics,doubleclick,
facebook,criteo,appnexus, etc.), remained top through time.
In terms of node importance, tracker sets appeared to be central
in many of the examined centrality metrics.
Finally, we constructed tracker-to-tracker graphs (T T ) for
trackers who coexisted in the same websites. We compared
these models with confirmed cookie synchronization (CS)
pairs, which we use as ground truth flows between tracking
entities. We found high overlap between the T T edges and the
CS pairs, which means that we can detect potential CS activity
and data sharing flows between trackers with a practical and
cheaper alternative than collecting data from real users. Fur-
thermore, the investigation of T T graph properties uncovered
“hidden”, and not studied so far, relationships between CS and
T T pairs.
Limitations. The measurements presented in this paper consti-
tute initial steps, along with related works, towards exploring
the evolution of online tracking and the ad-ecosystem. Mea-
suring such changes and identifying the tracking patterns and
behaviors of the Web is a non-trivial process, as it can involve
various vantage points and types of network connections,
devices, user demographics, emulated and/or real users, etc.
We believe that our study provides a lower bound on the
intensity of the tracking ecosystem’s state, but more work is
needed towards establishing a full and representative picture
of the online tracking, and its changes through time. Thus, our
work can be extended in the following directions:
•Main Page Crawls. We performed multiple, consistent,
easily replicable and lightweight- in terms of computing
power and network usage- crawls on each website’s main
page. Indeed different tracking may be happening in
subdomains and in nested website’s layers, compared to
the main domain page. Although it is important to go
beyond the main page, it is also challenging, as it requires
more resources and careful planning on browsing behav-
ior and activity exhibited by the crawlers (e.g., random vs.
consistent, visit duration, etc.). Interestingly, this aspect
of the problem was covered by recent work of Urban
et. al. [41], where they measured the 3rd-party tracking
mechanisms in subdomains of 10k popular pages. How-
ever, they did not investigated the problem in a time-
longitudinal fashion. Thus, future work could study how
subdomains’ tracking ecosystem evolves through time.
•IP Address & Location. We measured the ecosystem’s
tracking state using an EU-based IP address, in order to
highlight the potential GDPR effect on the EU space.
Measuring biases, discrimination and other differences
in tracking between locations inside and outside the EU
(e.g., EU vs. USA, or EU vs. Asia) and the type of IP
address used (e.g., residential, cellular, etc.) is a complex
research problem in itself with several parameters at play,
and it was out of the scope of this paper. In fact, recent
works (e.g., [16], [42]) have investigated this problem
in different ways, but not in a time-longitudinal fashion.
Therefore, future work could study this tracker evolution
while changing the crawling location.
•Cookie Synchronization Detection. One could argue
that our crawling method is not able to capture the
full scale of the cookie synchronization ecosystem. In-
deed, measuring this kind of information exchange is
challenging and requires specialized methodology such
as using user synthetic personas [22], [43]. Instead,
our approach is simpler and relies on automated crawls
without personas, which captures much of the cookie
synchronization, but not all. Thus, it measures a lower
bound of such data flows, as would be expected for real
users [10]. Future work could focus on utilizing more
elaborated techniques on top of ours, to more accurately
capture and study Cookie Synchronization flows.
VIII. CON CLU SIO N
In this paper, we presented a time-longitudinal study of the
online tracking ecosystem, in order to analyze its changing
state over a period of two years. We revealed patterns of the
Web tracking, and effects of privacy laws such as GDPR on
the tracking’s intensity in over 5000 popular websites. Apart
from studying such trends, we also presented graph models
easily built over our automated crawls. These graphs capture
different aspects of the tracking ecosystem, including key
or central trackers with high embeddedness with respect to
connectivity with other trackers, their influence on websites,
as well as changes in their ranking of importance. We also
studied how information flows between trackers, due to cookie
synchronization in traffic of real users, are reflected and match
flows found in our automated (non-real user) crawls. Our
study points to the future opportunity for building accurate and
representative graph models of the Web tracking ecosystem,
without the overhead of collecting real user data.
ACK NO WL E DG E ME N TS
We thank our shepherd Ruben Cuevas and the anonymous
reviewers for their helpful feedback. The research leading
to these results has received funding from the EU’s H2020
Programme under grand agreements 830927 (project CON-
CORDIA) and 871370 (project PIMCITY). The paper reflects
only the authors’ views and the Commission is not responsible
fsor any use that may be made of the information it contains.
REF ERE NC E S
[1] F. Roesner, T. Kohno, and D. Wetherall, “Detecting and defending
against third-party tracking on the web,” in Proceedings of the 9th
USENIX Conference on Networked Systems Design and Implementation,
ser. NSDI’12. Berkeley, CA, USA: USENIX Association, 2012, pp.
12–12. [Online]. Available: http://dl.acm.org/citation.cfm?id=2228298.
2228315
[2] S. Englehardt and A. Narayanan, “Online tracking: A 1-million-site
measurement and analysis,” in Proceedings of the 2016 ACM SIGSAC
Conference on Computer and Communications Security, ser. CCS ’16.
New York, NY, USA: ACM, 2016, pp. 1388–1401. [Online]. Available:
http://doi.acm.org/10.1145/2976749.2978313
[3] P. Eckersley, “How unique is your web browser?” in Proceedings of
the 10th International Conference on Privacy Enhancing Technologies,
ser. PETS’10. Berlin, Heidelberg: Springer-Verlag, 2010, pp. 1–18.
[Online]. Available: http://dl.acm.org/citation.cfm?id=1881151.1881152
[4] G. Acar, M. Juarez, N. Nikiforakis, C. Diaz, S. G¨urses, F. Piessens,
and B. Preneel, “Fpdetective: Dusting the web for fingerprinters,”
in Proceedings of the 2013 ACM SIGSAC Conference on Computer
& Communications Security, ser. CCS ’13. New York, NY,
USA: ACM, 2013, pp. 1129–1140. [Online]. Available: http:
//doi.acm.org/10.1145/2508859.2516674
[5] N. Nikiforakis, A. Kapravelos, W. Joosen, C. Kruegel, F. Piessens,
and G. Vigna, “Cookieless monster: Exploring the ecosystem of
web-based device fingerprinting,” in Proceedings of the 2013 IEEE
Symposium on Security and Privacy, ser. SP ’13. Washington, DC,
USA: IEEE Computer Society, 2013, pp. 541–555. [Online]. Available:
http://dx.doi.org/10.1109/SP.2013.43
[6] L. Olejnik, T. Minh-Dung, and C. Castelluccia, “Selling off privacy
at auction,” in Network and Distributed System Security Symposium
(NDSS), 2014.
[7] M. A. Bashir and C. Wilson, “Diffusion of user tracking data in
the online advertising ecosystem,” Proceedings on Privacy Enhancing
Technologies, vol. 2018, no. 4, pp. 85–103, 2018.
[8] G. Acar, C. Eubank, S. Englehardt, M. Juarez, A. Narayanan, and
C. Diaz, “The web never forgets: Persistent tracking mechanisms
in the wild,” in Proceedings of the 2014 ACM SIGSAC Conference
on Computer and Communications Security, ser. CCS ’14. New
York, NY, USA: ACM, 2014, pp. 674–689. [Online]. Available:
http://doi.acm.org/10.1145/2660267.2660347
[9] P. Papadopoulos, N. Kourtellis, and E. P. Markatos, “The cost of digital
advertisement: Comparing user and advertiser views,” in Proceedings of
the 2018 World Wide Web Conference, ser. WWW ’18. International
World Wide Web Conferences Steering Committee, 2018, p. 1479–1489.
[10] P. Papadopoulos, N. Kourtellis, and E. Markatos, “Cookie synchroniza-
tion: Everything you always wanted to know but were afraid to ask,” in
The World Wide Web Conference. ACM, 2019, pp. 1432–1442.
[11] Council of the European Union , European Parliament, “Directive
2002/58/ec of the european parliament and of the council of 12 july
2002 concerning the processing of personal data and the protection of
privacy in the electronic communications sector (directive on privacy
and electronic communications).”
[12] ——, “Regulation (eu) 2016/679 of the european parliament and of
the council of 27 april 2016 on the protection of natural persons with
regard to the processing of personal data and on the free movement
of such data, and repealing directive 95/46/ec (general data protection
regulation),” 2016.
[13] “California consumer privacy act of 2018 [1798.100 - 1798.199],” 2018.
[14] C. Iordanou, G. Smaragdakis, I. Poese, and N. Laoutaris, “Tracing
cross border web tracking,” in Proceedings of the Internet Measurement
Conference 2018. ACM, 2018, pp. 329–342.
[15] M. Degeling, C. Utz, C. Lentzsch, H. Hosseini, F. Schaub, and T. Holz,
“We value your privacy... now take some cookies: Measuring the gdpr’s
impact on web privacy,” arXiv preprint arXiv:1808.05096, 2018.
[16] A. Dabrowski, G. Merzdovnik, J. Ullrich, G. Sendera, and E. Weippl,
“Measuring cookies and web privacy in a post-gdpr world,” in In-
ternational Conference on Passive and Active Network Measurement.
Springer, 2019, pp. 258–270.
[17] J. Sørensen and S. Kosta, “Before and after gdpr: The changes in third
party presence at public and private european websites,” in WWW ’19
Companion Proceedings of the The Web Conference 2019. United
States: Association for Computing Machinery, 1 2019.
[18] C. Utz, M. Degeling, S. Fahl, F. Schaub, and T. Holz, “(un) informed
consent: Studying gdpr consent notices in the field,” in Proceedings of
the 2019 ACM SIGSAC Conference on Computer and Communications
Security, 2019, pp. 973–990.
[19] P. Agarwal, S. Joglekar, P. Papadopoulos, N. Sastry, and N. Kourtellis,
“Stop tracking me bro! differential tracking of user demographics on
hyper-partisan websites,” in Proceedings of The Web Conference 2020,
ser. WWW ’20, 2020, p. 1479–1490.
[20] S. Zimmeck, J. S. Li, H. Kim, S. M. Bellovin, and T. Jebara, “A privacy
analysis of cross-device tracking,” in 26th USENIX Security Symposium,
ser. USENIX Security 17. USENIX Association, 2017, pp. 1391–1408.
[21] J. Brookman, P. Rouge, A. Alva, and C. Yeung, “Cross-device track-
ing: Measurement and disclosures,” Proceedings on Privacy Enhancing
Technologies, 2017.
[22] K. Solomos, P. Ilia, S. Ioannidis, and N. Kourtellis, “Talos: An automated
framework for cross-device tracking detection,” in 22nd International
Symposium on Research in Attacks, Intrusions and Defenses (RAID
2019), 2019.
[23] J. R. Mayer and J. C. Mitchell, “Third-party web tracking: Policy
and technology,” in Proceedings of the 2012 IEEE Symposium
on Security and Privacy, ser. SP ’12. Washington, DC, USA:
IEEE Computer Society, 2012, pp. 413–427. [Online]. Available:
http://dx.doi.org/10.1109/SP.2012.47
[24] M. Falahrastegar, H. Haddadi, S. Uhlig, and R. Mortier, “Tracking
personal identifiers across the web,” in International Conference on
Passive and Active Network Measurement. Springer, 2016, pp. 30–
41.
[25] Z. Yu, S. Macbeth, K. Modi, and J. M. Pujol, “Tracking the
trackers,” in Proceedings of the 25th International Conference on
World Wide Web, ser. WWW ’16. Republic and Canton of
Geneva, Switzerland: International World Wide Web Conferences
Steering Committee, 2016, pp. 121–132. [Online]. Available: https:
//doi.org/10.1145/2872427.2883028
[26] N. Nikiforakis, W. Joosen, and B. Livshits, “Privaricator: Deceiving
fingerprinters with little white lies,” in Proceedings of the 24th
International Conference on World Wide Web, ser. WWW ’15.
Republic and Canton of Geneva, Switzerland: International World Wide
Web Conferences Steering Committee, 2015, pp. 820–830. [Online].
Available: https://doi.org/10.1145/2736277.2741090
[27] A. Panchenko, F. Lanze, J. Pennekamp, T. Engel, A. Zinnen, M. Henze,
and K. Wehrle, “Website fingerprinting at internet scale.” in NDSS, 2016.
[28] A. Lerner, A. K. Simpson, T. Kohno, and F. Roesner, “Internet
jones and the raiders of the lost trackers: An archaeological
study of web tracking from 1996 to 2016,” in 25th USENIX
Security Symposium (USENIX Security 16). Austin, TX: USENIX
Association, 2016. [Online]. Available: https://www.usenix.org/
conference/usenixsecurity16/technical-sessions/presentation/lerner
[29] V. Kalavri, J. Blackburn, M. Varvello, and K. Papagiannaki, “Like a
pack of wolves: Community structure of web trackers,” in International
Conference on Passive and Active Network Measurement. Springer,
2016, pp. 42–54.
[30] T. Urban, D. Tatang, M. Degeling, T. Holz, and N. Pohlmann, “The
unwanted sharing economy: An analysis of cookie syncing and user
transparency under gdpr,” arXiv preprint arXiv:1811.08660, 2018.
[31] I. Sanchez-Rola, M. Dell’Amico, P. Kotzias, D. Balzarotti, L. Bilge, P.-
A. Vervier, and I. Santos, “Can i opt out yet? gdpr and the global illusion
of cookie control,” in Proceedings of the 2019 ACM Asia Conference
on Computer and Communications Security, 2019, pp. 340–351.
[32] Alexa, “The top 500 sites on the web,” https://www.alexa.com/topsites/
category/Top/, 2018.
[33] Disconnect, “Disconnect lets you visualize and block the invisible
websites that track your browsing history,” https://disconnect.me/, 2018.
[34] Q. Scheitle, O. Hohlfeld, J. Gamba, J. Jelten, T. Zimmermann, S. D.
Strowes, and N. Vallina-Rodriguez, “A long way to the top: significance,
structure, and stability of internet top lists,” in Proceedings of the
Internet Measurement Conference 2018. ACM, 2018, pp. 478–493.
[35] A. P. Felt, R. Barnes, A. King, C. Palmer, C. Bentzel, and P. Tabriz,
“Measuring https adoption on the web,” in USENIX Security, 2017.
[36] J. Sørensen and S. Kosta, “Before and after gdpr: The changes in third
party presence at public and private european websites,” in The World
Wide Web Conference. ACM, 2019, pp. 1590–1600.
[37] L. C. Freeman, “A set of measures of centrality based on betweenness,”
Sociometry, pp. 35–41, 1977.
[38] M. Girvan and M. E. Newman, “Community structure in social and
biological networks,” Proceedings of the national academy of sciences,
vol. 99, no. 12, pp. 7821–7826, 2002.
[39] P. Rombach, M. A. Porter, J. H. Fowler, and P. J. Mucha, “Core-
periphery structure in networks (revisited),” SIAM Review, vol. 59, no. 3,
pp. 619–646, 2017.
[40] M. Latapy, C. Magnien, and N. Del Vecchio, “Basic notions for the
analysis of large affiliation networks/bipartite graphs,” arXiv preprint
cond-mat/0611631, 2006.
[41] T. Urban, M. Degeling, T. Holz, and N. Pohlmann, “Beyond the front
page: Measuring third party dynamics in the field,” in Proceedings of
The Web Conference 2020, 2020, pp. 1275–1286.
[42] D. Zeber, S. Bird, C. Oliveira, W. Rudametkin, I. Segall, F. Wolls´en,
and M. Lopatka, “The representativeness of automated web crawls as a
surrogate for human browsing,” in Proceedings of The Web Conference
2020, 2020, pp. 167–178.
[43] J. M. Carrascosa, J. Mikians, R. Cuevas, V. Erramilli, and N. Laoutaris,
“I always feel like somebody’s watching me: measuring online be-
havioural advertising,” in Proceedings of the 11th ACM Conference on
Emerging Networking Experiments and Technologies (CONEXT), 2015.
