Service Worker: a heavy request count that mostly stays first-party

Third-heaviest in requests, tenth of thirteen in companies

The median site running a service worker made 51 requests during the measured load. Against the thirteen technologies broken out in this index, that ranks 3rd on request volume, where rank 1 is the heaviest.

The same median page reached only 5 distinct companies, which ranks 10th of 13 on that measure, and 1 classified tracker, which ranks 9th of 13. So the traffic is near the top of the set while the number of separate organisations behind it sits near the bottom. That gap is what a service worker is built to produce: it intercepts and replays requests, and a cache that answers from the site's own origin generates request volume without adding anyone new to the page.

95.5% of these sites made contact before a banner could be answered

273 of the 286 sites reached a third party inside the opening window — the stretch that ends where a visitor could first have read a banner and answered it. That share ranks 4th of the 13 technologies here, with rank 1 the highest. The median first contact landed at 593 ms; the earliest single site in the group reached out at 80 ms.

The median number of companies first contacted after that window had closed is 0. The index measures that per site and takes the median of the result; it is not the difference between two other medians. For the median service-worker site, the third-party contact list is complete before the visitor has done anything.

The banner mix leans hard on the advertising framework

Consent software was detected on 117 of the 286 sites. The tool seen alongside most often was an IAB TCF framework banner, on 76 sites. OneTrust followed on 34, then CookieYes and Cookiebot on 2 each, Usercentrics on 1 and Iubenda on 1. One page can run two of these at once, so those are sightings, not slices of the 117.

TCF is the advertising industry's own consent framework, which encodes a visitor's choices into a string that bidders read. Its lead here is worth noting precisely because the pages themselves are not ad-heavy by company count: five companies at the median is a light page. What that suggests is two different populations sharing one technology — app-like sites that installed a worker for offline behaviour, and large commercial publishers that installed one and also sell inventory.

Two populations, visible in the named sites

reddit.com reached 5 companies, all 5 inside the window, first contact at 279 ms with no banner detected. twitch.tv reached 5 with 3 inside the window at 721 ms behind a TCF banner. amazon.co.uk contacted 4 companies, all 4 inside the window, at 172 ms, and amazon.de the same 4 with all 4 inside, at 272 ms — both behind TCF banners. soundcloud.com is the busiest of the group at 15 companies, 11 of them inside the window, first contact at 228 ms behind OneTrust. t-mobile.com reached 26 companies but only 6 inside the window, at 951 ms.

The opposite pattern shows up clearly in a cluster of Russian-language properties. vk.com contacted 6 companies with none of them inside the window and first contact at 3,653 ms. userapi.com: 6 companies, none early, at 3,153 ms. vkontakte.ru: 6 companies, none early, at 2,854 ms. vk.ru breaks with its siblings — 6 companies with 5 inside the window at 1,488 ms. A first contact after three seconds usually means the page rendered its own content from cache before it went looking for anyone else, which is exactly what a service worker makes possible.

The lightest entries are ampproject.org with 1 company at 776 ms and amazonalexa.com with 3, all inside the window, at 914 ms.

Why request count is the wrong number to worry about

Fifty-one requests sounds like a lot, and on this group it mostly is not a privacy figure. A cached image fetched through a worker is a request; so is a font and a background sync. The measure that maps onto exposure is how many separate organisations a page reached, and on that measure a service-worker page is quieter than most technologies in this index.

What the corpus records is timing and destination, not payload.