{"id":"defff6b3-0038-48b2-963d-ae5b50308ec8","arxiv_id":"1908.02261","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Analyzing 30k sensitive domains, this paper shows they carry many third-party trackers, mostly the same mainstream companies that dominate the web, plus niche trackers that specialize by category.","lead":"The authors built a tool to recognize sensitive websites (health, religion, politics, sexual orientation, porn) and then measured who tracks visitors to those sites. They found thousands of such sites are tracked by familiar advertising and analytics companies, and some niche trackers specialize in specific sensitive categories.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"All third-party domains are counted as trackers, so the headline medians (10/7/6) are inflated by CDN and font hosts that appear prominently in the paper's own top-20 lists; the qualitative result likely survives.","rationale":"I agree with the Reader's weakest assumption and do not see a different concern that is more likely to undermine the central claim. The classifier evaluation is reasonably careful (held-out split, feature sensitivity, manual validation on unlabeled sites), and the CSync statistics align numerically with a prior independent study, which is good supporting evidence. The paper is also upfront in Section 5.4 that visible CSync activity is a lower bound. The unresolved issue is that the main tracking counts treat every third-party domain as a tracker. This is not fatal for the qualitative claim: the top-20 coverage plots independently demonstrate that major ad-tracking domains such as google-analytics.com and facebook.com are present on a large fraction of sensitive sites. But it is a real load-bearing flaw for the precise medians that the abstract and findings highlight, because the paper's own tables prominently include non-tracking CDN and font hosts. Whether the sensitive-vs-TopK gap shrinks depends on the distribution of such non-trackers across categories, which is exactly what the proposed blocklist-based recomputation would reveal. Since the Reader already flagged this assumption and assigned CONDITIONAL, I do not move the verdict.","tokens_in":27037,"tokens_out":4968,"duration_ms":55218,"concrete_test":"Recompute the per-category median number of unique tracker eTLD+1 domains using a tracker blocklist (e.g., Disconnect's tracking-protection list or EasyList+EasyPrivacy) instead of counting all third-party domains. Concretely, re-crawl the same Curlie subcategories and label every observed third-party host against the blocklist, then compare the resulting medians to Table 5. If the Health, Political Beliefs, and Sexual Orientation medians drop substantially (e.g., from 10/7/6 to near 0-3), the headline counts are materially inflated. Separately, manually audit a random sample of 50 third-party hosts that appear on sensitive sites but are absent from the blocklist to test the claim that most unfiltered third-party domains are trackers.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing weakness is the unrestricted definition of a tracker. Section 4.1 states that the authors choose to include all third party domains in their dataset without any filtering. The headline counts in Table 5 (median 10 for Health, 7 for Political Beliefs, 6 for Sexual Orientation) are therefore counts of any third-party domain contacted during rendering, not counts of entities that actually track users. The paper's own top-20 lists confirm the problem: fonts.googleapis.com, fonts.gstatic.com, ajax.googleapis.com, cdnjs.cloudflare.com, and code.jquery.com appear in multiple categories. These are CDN, font, and library hosts, not trackers. The Section 4.1 assertion that most of the third party domains encountered are indeed trackers is never quantified and appears inconsistent with the displayed rankings. Because the abstract and Section 1.3 present the per-category medians as the central quantitative finding, the paper overstates the number of entities tracking visitors to sensitive sites. The qualitative conclusion that mainstream advertising trackers appear on sensitive sites does survive, since the same figures show Google and Facebook properties with high coverage.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates which third parties track visitors to websites in GDPR-sensitive categories (Health, Ethnicity, Religion, Sexual Orientation, Political Beliefs, plus Porn). It first builds a classifier that labels sensitive domains using manually filtered Curlie categories as training data, reaching 85.66% accuracy with content and metadata features. It then uses a headless-browser crawler that renders roughly 30k sensitive domains and records all third-party requests, including recursive inclusions. Headline findings are that sensitive sites are tracked less intensely than popular TopK sites (median unique third-party domains: 10 for Health, 7 for Political Beliefs, 6 for Sexual Orientation, versus 17 for TopK), that the same mainstream trackers dominate, that some niche trackers specialize by category, and that mainstream and niche trackers exchange data via cookie synchronization. The paper concludes that sensitive domains are tracked to an alarmingly high degree, with direct inclusion by site owners implying tracker awareness.","tokens_in":27270,"tokens_out":2930,"duration_ms":32582,"significance":"If the results hold, the paper contributes a useful large-scale measurement of third-party presence on sensitive domains, a lightweight methodology for building sensitive-domain classifiers from Curlie, and one of the first looks at cookie synchronization between mainstream and niche trackers on such sites. The study's strengths include a fresh crawl rather than reuse of existing measurements, explicit documentation of the inclusion-chain construction, and an effort to verify niche trackers by manual inspection of their self-descriptions. The core qualitative finding -- that mainstream advertising trackers such as Google Analytics, DoubleClick, and Facebook appear on sensitive sites with high coverage -- is credible and important. However, the headline numerical claims about the number of trackers are weakened by the operationally unrestricted definition of a tracker, which counts CDN, font, and library hosts as trackers. The paper deserves publication after the quantitative claims are made robust to that concern.","major_comments":[{"comment":"The central quantitative claim is the median number of trackers (10 for Health, 7 for Political Beliefs, 6 for Sexual Orientation, presented in the abstract and Section 1.3). Section 4.1 defines trackers as all third-party domains observed during rendering, choosing to include all third-party domains without filtering, and asserts without quantification that most are indeed trackers. The paper's own top-20 lists show that fonts.googleapis.com, fonts.gstatic.com, ajax.googleapis.com, cdnjs.cloudflare.com, and code.jquery.com appear in multiple categories. These are CDN, font, and library hosts, not tracking services in the usual sense, so counting them as trackers inflates the headline medians. The qualitative conclusion that mainstream tracking services are present on sensitive sites would likely survive re-analysis, but the specific numbers labeled 'number of third party trackers' are not supported as stated. Please re-run the analysis using a tracker filter list (e.g., EasyList/EasyPrivacy or an equivalent) or, at minimum, provide a sensitivity analysis that excludes CDN/font/library hosts, and report both the filtered and unfiltered results.","section":"Section 4.4, Table 6"},{"comment":"The niche-tracker analysis is presented as evidence that some trackers specialize in sensitive categories. The method excludes domains whose coverage on other categories exceeds a threshold q, but q is set to different values per category (1.0% to 5.5%) without a stated justification or sensitivity analysis over q. Several listed domains are plausible niche trackers (e.g., medtargetsystem.com, nationbuilder.com), and the manual verification is a good step. However, as written the table is suggestive rather than demonstrative; please add robustness checks across a range of q values or include a statistical test comparing within-category coverage against the baseline, so the claim of category-specialized trackers is not an artifact of the chosen thresholds.","section":"Section 4.4, Table 6"}],"minor_comments":[{"comment":"The word 'described' is misspelled as 'descriped' in the sentence about filter lists and methodologies.","section":"Section 4.1"},{"comment":"The figure caption and text refer to 'a.discuscdn.com' and 'disquiscdn', while the text elsewhere uses 'disqus.com'; the names should be made consistent and correct.","section":"Figure 5"},{"comment":"The last sentence of the Contributions paragraph is garbled: 'for the benefit of other researchers and' is an incomplete fragment that should be rewritten or removed.","section":"Section 1.2"},{"comment":"The phrase 'Each raw of the confusion matrix' should be 'Each row of the confusion matrix'.","section":"Section 3.2.1"},{"comment":"The sentence 'In total, we have 62 keyword' should be '62 keywords', and the keyword list is described only by examples; consider providing the complete list or a pointer to it.","section":"Section 5.2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful core of this paper survives reading: it gives a large-scale look at third-party presence on GDPR-sensitive domains, and the qualitative finding—that mainstream advertising trackers like Google and Facebook appear on health, politics, and sexual-orientation sites—is solid and worth knowing. The scale is the main novelty: roughly 30k sensitive domains across six categories, using a Curlie-based classifier that is evaluated on held-out data and reaches about 86% accuracy with TF-IDF. The niche-tracker analysis and the cookie-sync comparison between niche and mainstream trackers are also new, and the CSync results align with prior work.\n\nThe soft spot is real and load-bearing for the quantitative claims. Section 4.1 says the authors deliberately include all third-party domains without filtering, and the paper's own top-20 lists show fonts.googleapis.com, cdnjs.cloudflare.com, and code.jquery.com. Those are not trackers. So the headline medians (10 for Health, 7 for Political Beliefs, 6 for Sexual Orientation) are counts of any third-party contact, not counts of entities that actually track. The assertion that most third-party domains are trackers is never quantified and looks inconsistent with the displayed rankings. The qualitative conclusion that sensitive sites are tracked by mainstream advertising companies survives, because the same figures show Google Analytics above 40% coverage in every category.\n\nTwo smaller issues: the per-category niche-tracker thresholds q are chosen without sensitivity analysis, and the preprint ships no code or data (the text promises release upon publication). Both are fixable. The related work section honestly lists prior studies that touched sensitive-domain tracking, so the novelty is scale and specific findings rather than the basic question.\n\nThis paper deserves a serious referee. The fix is not to reject but to require a proper tracker definition—report both raw third-party counts and filtered tracker counts—and to justify the niche-tracker threshold choice. I would bring it to a reading group to discuss methodology, and I would cite it with a caveat on the numbers. Send it to peer review with a request for revision, not a desk reject.","headline":"The paper has real evidence that sensitive domains are tracked by mainstream ad-tech, but its headline tracker counts are inflated by counting every third-party domain, including font and CDN hosts, as a tracker.","tokens_in":27783,"tokens_out":1634,"would_cite":true,"duration_ms":19056,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Sensitive websites are tracked by mainstream ad companies, with median 10 trackers on health sites.","keywords":["web tracking","sensitive domains","GDPR","third-party trackers","cookie synchronization","Naive Bayes classification","privacy measurement"],"falsifier":"Re-run the same crawl pipeline but classify each third-party domain against a tracker list (e.g., EasyList/EasyPrivacy or a manually audited tracker domain list) and compare the median counts; if the filtered medians drop to near zero for sensitive categories while remaining high for TopK, the paper's central quantitative claim would be refuted.","tokens_in":26815,"feed_emoji":"🕵️","tokens_out":1599,"duration_ms":20235,"temperature":0.7,"pith_summary":"This paper asks who tracks websites that most people would rather not be seen visiting, such as health, political, and sexual-orientation sites, and answers with a large-scale measurement. The authors build a classifier that automatically finds sensitive domains using a human-edited web directory, then render about 30,000 of those pages in a real browser and record every third-party request. They find that sensitive sites are tracked less intensely than popular mainstream sites, but still heavily: median 10 third-party domains on health sites, 7 on political-belief sites, and 6 on sexual-orientation sites, versus 17 on top-1000 popular sites. The same well-known advertising trackers dominate, some smaller niche trackers specialize in particular sensitive categories, and mainstream and niche trackers often exchange data through cookie synchronization. The paper matters because it supplies concrete, reproducible evidence about a privacy harm that data-protection law treats as especially serious, and because the authors argue the trackers are usually there intentionally, included directly by site owners who join ad-revenue programs.","feed_headline":"Health sites carry a median 10 third-party trackers","feed_subtitle":"A crawl of 30,000 sensitive domains shows mainstream ad trackers dominate, and niche trackers sync cookies with them.","key_machinery":"The argument is carried by three linked instruments: (1) a sensitive-domain classifier trained on hand-filtered subcategories of Curlie.org (the human-edited Open Directory successor), using TF-IDF features over webpage text plus meta tags, which reaches about 85.66% accuracy and can detect sensitive domains in unlabeled lists like Alexa TopK; (2) a Puppeteer-based crawler that renders pages, scrolls to trigger lazy-loaded requests, and waits after window.onload to capture recursive third-party inclusion chains up to nine levels deep; and (3) an inclusion-chain construction that maps each request to its initiator and hop level, letting the authors distinguish direct inclusions (0 hops, where the tracker certainly knows the first-party domain) from indirect ones, and a keyword-based cookie-synchronization detector that recognizes ID-passing URL arguments between third-party domains.","core_discovery":"The central claim is that sensitive domains are tracked by third parties, though less intensely than mainstream domains. Using five GDPR-listed sensitive categories (Health, Ethnicity, Religion, Sexual Orientation, Political Beliefs) plus Porn, the authors report median third-party counts of 10 for Health, 7 for Political Beliefs, 6 for Sexual Orientation, 5 for Religion, 5 for Ethnicity, and 7 for Porn, versus 17 for a TopK baseline. The top trackers on sensitive sites are the same mainstream advertising and analytics services that dominate the wider web, with google-analytics.com present on at least 40% of sensitive sites in every category. Beyond those, the authors identify niche trackers that are absent from non-sensitive sites but present on specific sensitive categories, several of which advertise their ability to target health, political, or adult audiences. They also detect cookie synchronization between mainstream and niche trackers, finding that 26.87% of health websites host at least one synchronized pair, and that in Health and Political Beliefs roughly 12.8% and 11.7% of synchronization pairs involve niche trackers. The authors conclude that in most cases the tracker is deliberately included by the sensitive site owner, so the mainstream tracker is aware of its presence on the sensitive domain.","pith_inferences":["The headline tracker counts deliberately count every third-party domain, including font CDNs, JavaScript libraries, and other infrastructure; a refined tracker-only list would likely lower the medians, though the qualitative finding that mainstream advertising trackers are present would probably survive.","The paper's own logic suggests a testable regulatory lever: sensitive-domain owners could be offered advertising programs that refuse admission to sensitive categories, and the authors' classifier could mechanically audit whether such exclusion is honored.","The cookie-synchronization result implies a concrete data-flow concern that the paper does not fully quantify: even if a niche tracker holds no personal identifiers, synchronization with a PII-holding mainstream tracker can attach a pseudonymous identity to a history of sensitive visits.","The methodology could be extended to measure tracking after GDPR enforcement actions, comparing tracker presence on sensitive domains before and after regulatory changes to estimate deterrence effects."],"forward_implications":["If the paper is right, a visit to a cancer forum, a gay dating site, or a niche political blog routinely leaks to mainstream advertising companies that the visitor was there, and the leakage is not accidental: the site owner usually included the tracker deliberately.","The tracker census implies that data-protection enforcement cannot rely on the naive hope that sensitive sites are too small or too marginal to attract trackers; even low-coverage niche trackers appear on up to 14% of sites within a sensitive category.","Because cookie synchronization links niche trackers (which may have observed users on sensitive domains) to mainstream trackers (which often hold personal identifiers), the paper implies that the mainstream tracker can enrich its profile with sensitive-visit information without itself appearing on the sensitive site.","The classifier itself is a reusable instrument: the authors argue it can be retargeted to arbitrary sensitive or non-sensitive categories in roughly five minutes of manual effort, enabling proactive monitoring and compliance audits at scale."],"supporting_citations":[{"why":"Curlie.org supplies the human-edited taxonomy of labeled websites from which sensitive training sets are drawn.","marker":"[23]"},{"why":"The GDPR text defines the sensitive categories (health, sexual orientation, political beliefs, etc.) that the study operationalizes.","marker":"[28]"},{"why":"The 1-million-site tracking measurement provides the baseline comparison for how intensely mainstream domains are tracked.","marker":"[72]"},{"why":"A global mobile tracking study supplies prior methodology for measuring third-party trackers at scale.","marker":"[63]"},{"why":"The cookie synchronization study supplies the detection keywords and the 1.47% baseline percentage that the paper compares its own CSync rates against.","marker":"[60]"},{"why":"The study of information flows between ad exchanges supplies the concept of cookie synchronization and the mechanism for passing IDs via request URLs.","marker":"[11]"},{"why":"A mobile third-party tracking study supports the recursive multi-hop inclusion detection methodology.","marker":"[13]"}],"fun_headline_variants":["Sensitive domains are tracked, though less than mainstream","Cookie sync ties niche trackers to mainstream on health sites","Health sites host 10 trackers, politics 7, sexual orientation 6","Niche trackers target sensitive categories, often with cookie sync","Trackers on sensitive domains: fewer, but niche players appear"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that every third-party domain observed during page rendering is a tracker, because it intentionally filters nothing; if many of those domains are font hosts, CDNs, or code libraries that do not track users, the reported median tracker counts overstate tracking intensity.","fun_headline_variants_meta":{"raw":{"variants":["Sensitive domains are tracked, though less than mainstream","Cookie sync ties niche trackers to mainstream on health sites","Health sites host 10 trackers, politics 7, sexual orientation 6","Niche trackers target sensitive categories, often with cookie sync","Trackers on sensitive domains: fewer, but niche players appear"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000924,"raw_usage":{"total_tokens":3972,"prompt_tokens":968,"completion_tokens":3004,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":2917}},"tokens_in":584,"tokens_out":3004,"duration_ms":24609,"temperature":1.0,"reasoning_tokens":2917,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:49:19.830176+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same crawl pipeline but classify each third-party domain against a tracker list (e.g., EasyList/EasyPrivacy or a manually audited tracker domain list) and compare the median counts; if the filtered medians drop to near zero for sensitive categories while remaining high for TopK, the paper's central quantitative claim would be refuted.","supporting_citations":[{"cited_title":"Curlie - the largest human-edited directory of the Web","cited_arxiv_id":null,"evidence_quote":"Curlie.org supplies the human-edited taxonomy of labeled websites from which sensitive training sets are drawn."},{"cited_title":"Data protection in the EU, /T_he General Data Protection Regulation (GDPR); Regulation (EU) 2016/679","cited_arxiv_id":null,"evidence_quote":"The GDPR text defines the sensitive categories (health, sexual orientation, political beliefs, etc.) that the study operationalizes."},{"cited_title":"In ACM CCS (2016)","cited_arxiv_id":null,"evidence_quote":"The 1-million-site tracking measurement provides the baseline comparison for how intensely mainstream domains are tracked."},{"cited_title":"In NDSS (2018)","cited_arxiv_id":null,"evidence_quote":"A global mobile tracking study supplies prior methodology for measuring third-party trackers at scale."},{"cited_title":"P.Cookie Synchronization: Everything You Always Wanted to Know But Were Afraid to Ask","cited_arxiv_id":null,"evidence_quote":"The cookie synchronization study supplies the detection keywords and the 1.47% baseline percentage that the paper compares its own CSync rates against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The study of information flows between ad exchanges supplies the concept of cookie synchronization and the mechanism for passing IDs via request URLs."},{"cited_title":"V., Z/h.sc/a.sc/o.sc, J., L/i.sc/b.sc/e.sc/r.sc/t.sc, T., /a.sc/n.sc/d.sc S/h.sc/a.sc/d.sc/b.sc/o.sc/l.sc/t.sc, N","cited_arxiv_id":null,"evidence_quote":"A mobile third-party tracking study supports the recursive multi-hop inclusion detection methodology."}],"review_version":1}