Skip to content
  • Home
  • News
  • Websites Tracking Without Consent

The 100 Most-Visited Websites Track You Before You Answer the Cookie Banner, Audit Finds

Published

Miklos Zoltan

By Miklos Zoltan

Founder - PrivacyAffairs

Laura Bede

Fact-Checked this

Leaving a cookie banner unanswered doesn’t stop the tracking, according to an audit of the 100 most-visited English-language websites published today.

Privacy lawyer and computer scientist Alexander Hanff loaded each site in a clean Chrome browser from Malta, inside the EU, and touched nothing. Across 265 pages, his tool recorded 518,770 events that affect your privacy: cookies written, device details read, fingerprinting scripts run and data sent to other companies. Every one of them happened while the consent banner was still waiting for an answer.

Hanff published the research on 17 November with the Dutch non-profit SOMI, which says it is looking at collective legal action. Hanff says he will file complaints with data protection authorities this week. When we published, The Register had reported it; otherwise it had only reached privacy circles on Mastodon.

One thing to know up front: WebSentinel, the audit tool, is Hanff’s own commercial product, and the research page ends with an offer to audit your website. He says he paid for the research himself.

How the Audit Was Done

The sample is the 100 most-visited English-language content websites on the Tranco list for September 2026: Facebook, YouTube, Wikipedia, Reddit, the New York Times, CNN, the BBC, the Guardian and so on. For each site Hanff opened up to three pages (the homepage, the privacy notice and, where there was one, the cookie notice) and left each open for five minutes. He ran the captures between 1 and 3 October 2026.

The browser started from a wiped profile for every site, and no banner was ever clicked. A consent banner appeared on 165 of the 265 pages. So the figures show what a first-time visitor’s device goes through before making any choice.

Hanff sealed each recording with hashes, signatures and timestamps so that regulators or courts can check nothing was changed afterwards. Anyone can open the per-site reports, an evidence viewer and screen replays on the research page.

What It Found

The audit in nine figures
What was recorded
The figure
Why it matters to you
Privacy events before any choice
518,770 across 265 pages; 2,695 on the median website
Every one happened while the cookie banner sat unanswered. Saying nothing is not consent under EU law.
Most serious events
14,675 rated tier 1 (interception, data sent out, high-entropy fingerprinting), on all 100 websites
No website in the sample came out clean.
Cookies
1,492 distinct cookies on 96 websites, 372 of them from third-party domains; the longest is set to expire in 2094
Rejecting later doesn’t remove what is already on your device.
Fingerprinting
75 websites used canvas, WebGL, WebGPU, WebRTC, audio, font, battery or sensor checks
A fingerprint follows you even after you clear cookies.
Disguised trackers (CNAME cloaking)
21 subdomains on 18 websites were aliases of another company’s servers
The browser treats them as the website itself and sends its cookies along. Many blockers miss them.
Where the data went
92.4% of 1,400 servers contacted sit on networks outside the EEA, almost all in the US; 1,034 belong to third parties
Transfers outside the EEA need a legal basis such as the EU–US Data Privacy Framework.
Code for the browser’s built-in AI
Delivered on 37 websites, on 36 of them through Google Publisher Tag
It checks for Chrome’s on-device AI model and times a session with it. The audit shows the code arrived, not that it ran.
Consent banners
115 of 165 banners used at least one dark pattern
They push you toward “Accept”. The EDPB’s cookie banner taskforce treats several of these designs as breaches.
Privacy notices
None of the 95 captured notices fully matched what the pages did; reading the median notice takes 28 minutes
The notice is not a reliable guide to what a site does to your device.
All figures come from one first visit per page in desktop Chrome from Malta, 1–3 October 2026, with no click on any banner. They are the researcher’s counts, not findings of a regulator or court.
Source: Alexander Hanff, “We Take Your Privacy Seriously”, WebSentinel aggregated report and summary, published 17 November 2026; checked 17 November 2026.

ESPN (57,194 events), Disney’s go.com (42,253) and the New York Times (30,158) logged the most. The fewest came from the Internet Archive (305) and the research archive arXiv (200), and even they recorded some tier 1 events.

Two findings stand out. The privacy and cookie notices, the very pages that explain how your data is used, produced more events (267,749) than the homepages did (251,021). And several websites say in their notices that cookies are set only with your consent while doing it anyway. NPR set 28 cookies, including advertising cookies from Criteo and Google’s DoubleClick, although its notice says “we only set these non-Essential cookies with your consent”. The Guardian, IMDb, Flickr and Unsplash are on the same list in the report.

Ad Code That Probes Chrome's Built-In AI

The least-known finding concerns Gemini Nano, the AI model that Chrome downloads to many computers to run features on the device. Websites can talk to it through new browser interfaces.

According to the report, Google Publisher Tag, Google’s advertising library, contains code that checks whether those interfaces exist, starts a session with the on-device model and records how long that takes. It reached 36 of the 100 websites through Google’s ad library and one more by another route, including the BBC, CNN, the New York Times, the Wall Street Journal, Stack Overflow and Nature.

Hanff’s reading is that this is a new way to gather device characteristics for ad measurement or fingerprinting, done without consent. His own report also sets the limit: the version of his tool used here didn’t log calls to these interfaces, so the audit proves the code was delivered to the browser, not that it ran.

If you’d rather websites couldn’t reach the model at all, Chrome lets you remove it. In Google’s help page, the switch is under Settings, AI Innovations, On-device AI. Turning it off deletes the model and stops new downloads, but it also turns off features that use it, including Chrome’s on-device scam warnings.

What You Can Do

Ignoring the banner, or rejecting cookies later, doesn’t stop what loads in the first seconds. What helps is cutting off the third-party code before it runs.

  • Use a browser that blocks trackers by default. Firefox’s Enhanced Tracking Protection in Strict mode and Brave’s Shields block many of the third-party scripts and cookies the audit counted. Neither stops everything, and some sites break in strict modes.
  • Add a content blocker. uBlock Origin in Firefox can also unmask disguised CNAME trackers like the ones the audit found on 18 websites. In Chrome, which only allows the cut-down uBlock Origin Lite, that check isn’t possible.
  • Expect fingerprinting to get through. 75 of the websites used advanced fingerprinting. Clearing cookies doesn’t reset a fingerprint; our guide to browser fingerprinting explains which browsers resist it best.
  • Ask what a site holds about you. In the EU and UK you can make an access request under Article 15 of the GDPR. Our guide to your data rights has a template letter, and SOMI offers a free tool, My Data Finder, for the same request.
  • Complain if you get nowhere. Our step-by-step complaint guide shows where to file in the EU, the UK and the US.

How Much Weight to Give It

The evidence is detailed and open to inspection, which is more than most tracking studies offer. Still, a few points matter when you read the headline numbers.

“Events” is a broad count. Tier 3, at 219,632 events, covers protocol mechanisms and passive reads that can be used for tracking, and every single read counts. Wikipedia shows why the totals alone mislead: it logged 2,554 events, more than Reddit or Facebook, yet only 13 were rated high or critical, against 768 on the BBC. The 14,675 tier 1 events are the core of the case.

The legal conclusions are Hanff’s. He holds that almost none of this was strictly necessary to show a web page, so all of it needed consent. Regulators may agree on much of it, but the EU’s cookie rule does allow some storage without consent, and websites will argue that security tools such as bot checks belong there. The report itself notes that the largest single source of fingerprinting was a bot challenge.

It’s also one visit per page, from one EU country, in desktop Chrome. Websites can serve different code to other countries, browsers and phones.

What it does show is that the gap between what a cookie banner promises and what happens behind it is wide on the web’s biggest sites. Regulators have fined it before: our GDPR fines tracker lists 42 fines that mention cookies, worth about €750 million together. How the rules themselves may change under the EU’s Digital Omnibus plans is covered in our guide to cookie consent.

What Happens Next

Hanff says he is sending the reports and evidence to every EU data protection authority through the European Data Protection Board, the European Commission, Members of the European Parliament, the US Federal Trade Commission, California’s privacy agency and US state attorneys general. He plans formal complaints about the audited websites in the coming week, and SOMI says it is exploring collective legal action.

We’re contacting Google about the code in its ad library that calls Chrome’s on-device AI, and will add its answer here.

Our Mission

We believe security online security matters and its our mission to make it a safer place.