Grey Zone · Episode 3Trackers

The Grey Zone Index

How many advertising and audience-measurement companies a website contacts before you touch the cookie banner. Not after you accept. Not after you refuse. Before you are asked.

Captured 10 August 2026 · from Finland · first ~10 seconds of load, before the cookie banner was touched

20
Sites tested
13
Sent data before consent
7
Sent nothing

One company, 6 of the 20 sites

Google appears on more of this sample than any other company — on 6 sites, before consent. No other company appears on more than 2. That is the finding, more than any single site's number.

The twenty

Sorted worst first. “Ignored” is the count of requests thrown out as the site doing its job — its own CDN, fonts, the consent banner, payments, anti-bot. Those are not held against anyone. Every site name links to its raw capture file.

#SiteRequestsCompanies
1www.aljazeera.com
Google, Freestar, Amplitude, Chartbeat
244
2www.foxnews.com
Braze, True Anthem, Amazon, Knotch, Adobe, Pix
196
Fox's own identity-graph hosts (prod.idgraph.dt.fox, loki-ingest.fox.com) are first-party and therefore not counted here.
3www.bbc.com
Optimizely, Dotmetrics, Chartbeat, Google
144
4www.actvli.com
Google
111
Our own site, left in the sample. 11 requests to Google ad and analytics endpoints before consent.
5www.zalando.fi
Google
81
6www.amnesty.org
Zoho, Google
62
The 6 counted here are Zoho and Google. Not counted, and arguably the more interesting one: a Stripe beacon sent a full browser fingerprint before consent — operating system, language, installed plugins, screen geometry and a hash. Stripe is ignored under the fairness rule as a payment provider, which is exactly where that rule is most generous.
7yle.fi
AudienceProject, Kilkaya
52
8um.dk· Danish Ministry of Foreign Affairs
Silktide
41
9www.amazon.de
Amazon
31
All three counted requests go to amazon-adsystem.com, which Amazon owns. By company this is arguably a zero; the domain rule is applied identically to all twenty sites.
10www.nordea.fi· Bank
Tealium
31
11www.anthropic.com
HubSpot
21
12www.indiehackers.com
Google
21
13www.sitra.fi· Finnish innovation fund
Matomo
21
14github.com00
15www.ebay.com00
Zero only because PerimeterX anti-scraping (perfdrive.com) is treated as functional. Same call as LinkedIn.
16www.facebook.com00
17www.hus.fi· Helsinki University Hospital00
18www.linkedin.com00
Zero only because PerimeterX anti-scraping (protechts.net) is treated as functional. It is genuinely security and genuinely fingerprinting: two POSTs totalling roughly 21 KB of encoded device data went out before the banner was touched. The capture file is published so this call can be argued with.
19www.nhs.uk· NHS00
Evidenced by screenshot, not HAR. The export ran with the third-party filter active and there was nothing left to export — the Network panel read 0 / 13 requests, all of them first-party.
20www.temu.com00

Who the companies are

How many of the 20 sites each company was reached from, before consent.

The raw capture files

Every site name in the table above downloads that site's HAR file — the complete record of what the browser did, exported from Chrome DevTools. Drag one back into the Network tab to inspect it, or run the counting script over it and check our arithmetic. They reproduce every figure on this page exactly; that was verified after scrubbing rather than assumed.

What was changed before publishing them

Response bodies were removed — they are other people's copyrighted JavaScript and 94% of the file size, and nothing in the analysis reads them. API keys sent to the browser were replaced with a marker, keeping the header name, because that a site ships an API key before you consent is part of the finding while a working key is someone else's problem to have abused.

URLs, query strings, request headers and timings are untouched. Redacting those would leave you taking the counts on trust, which is the opposite of the point. Full notes on what was changed and why.

How this was measured

The rule that makes it valid

Measurement stopped the instant the cookie banner was touched. No clicking, no scrolling, no “continue to site”. Everything counted here happened before any consent was given or refused. Each site was loaded in a fresh Chrome profile — no cookies, no storage, no memory of a previous visit.

What counts, and what does not

A host is ignored if it is the site's own infrastructure, or if it exists to deliver the thing you asked for — assets, fonts, the consent banner itself, payments, login, anti-bot. A host is counted if data goes out to it for advertising, audience measurement or profiling, regardless of who owns it.

Calls were made from full URLs, not hostnames. www.google.com serving a CAPTCHA and www.google.com serving an ad-conversion ping are the same hostname and opposite verdicts.

Domains and companies are counted separately

The domain count is mechanical and reproducible. The company count is a judgement — Google alone ships six domains, so a site running nothing but Google scores 6 on one measure and 1 on the other. Both are published side by side so the grouping does not have to be taken on trust.

What this data does not yet show

  • No cookie or storage counts. Every measurement here is network requests going out. What gets written onto the device before consent is a separate and arguably sharper question, and it has no data behind it yet. It is coming.
  • Roughly ten seconds per site. A live news page never finishes loading — Al Jazeera made 113 requests at ten seconds and 1,092 at ninety. Every figure here is therefore conservative.
  • Ad auctions vary between runs. Counts on ad-heavy sites genuinely differ load to load. Treat single-digit differences between sites as noise.
  • One sample, one location, one date. Captured from Finland on 10 August 2026. Results elsewhere, especially outside the EU, are likely to be worse rather than better.

Our own site is in the sample

www.actvli.com is row 4. Eleven requests to Google advertising and analytics endpoints before consent. Removing it and hoping nobody checked was the other option. It is in the table for the same reason every other number in it should be believed.

Check your own, in about thirty seconds

  1. Open the site in a private window. Press F12 for developer tools, then the Network tab.
  2. Reload the page and do not touch the banner.
  3. Type -domain:*theirdomain.com into the filter box. What is left is everyone else your browser just talked to.

Check the places you assume are fine. Not because they will be worse — because otherwise you are guessing.

Two payloads are worth opening yourself: amnesty.org's Stripe beacon, which sent a browser fingerprint before consent, and linkedin.com's, which sent roughly 21 KB of encoded device data to an anti-scraping vendor while scoring zero on this page. That second one is the weakest call in the whole dataset, and the file is there so you can argue with it.

Want a specific site checked in a future episode? Send it to us. For scam and fraud risk on a domain rather than tracking, use the RiskScope scanner or read the blog.