User Agent Parser

Almost every token in a user agent string is there because some site once sniffed for it, and the three biggest consequences are the reason most parsers get the same three things wrong.

The first is the engine. Every Chromium browser sends "AppleWebKit/537.36 (KHTML, like Gecko)" and "Safari/537.36" and has done since 2013, while its engine is Blink. A parser that reads the engine off the AppleWebKit token calls Chrome, Edge, Opera, Brave, Vivaldi and Samsung Internet WebKit, which is exactly the mistake the token was put there to cause. Here the engine is derived from the browser instead, and the reason is printed next to the answer.

The second is iOS. App Store rules require every browser on iOS to use the system engine, so Chrome is CriOS, Firefox is FxiOS, Edge is EdgiOS and Opera is OPT. None of them contains a Chrome token at all, and all of them contain "like Mac OS X", which is what makes a naive macOS test match first and report every iPhone in the world as a Mac. The order of tests here puts iOS and iPadOS ahead of macOS for that reason, and each of the four browsers is named rather than falling through to Safari or to Unknown.

The third is that several fields are frozen and carry a constant whatever the machine is. macOS has reported 10_15_7 in the user agent string for every version since Big Sur, so the number says nothing. Windows 11 reports "Windows NT 10.0" exactly as Windows 10 does, because Microsoft did not raise the NT version, so a tool that tells you which one you are on is guessing. Chrome's reduced user agent sends "Android 10" and the device model "K" on every Android device there is, and zeroes the minor, build and patch of its own version. Each of those is labelled frozen rather than reported as a reading, and the page says what the constant is and why.

Where the string cannot answer, the browser itself sometimes can. navigator.userAgentData is the browser answering the same questions about itself in structured form, and on a Chromium build it carries the real platform version, the device model and the CPU architecture. That is the only reference a user agent parser can be checked against that is not another parser, so when the box holds your own browser's string the two are shown side by side with the disagreements marked. On the machine these tests run on, the string says macOS 10.15.7 and the browser says macOS 26.6.2, and the string cannot give the architecture at all while the browser reports ARM64. Safari and Firefox declined to implement the API, and the page says so rather than leaving the panel blank.

Non-browser clients are the other half. A keyword list built from bot, crawl, spider, slurp, search, fetch and archive misses facebookexternalhit, curl, python-requests, Go-http-client and okhttp, which between them account for a great deal of what actually hits a server. Named crawlers are matched on the token each one publishes, including the AI crawlers that appeared recently, and anything left that writes the product/version form RFC 9110 section 10.1.5 describes without opening with Mozilla is reported as a library rather than as a desktop browser.

Finally, the whole string is broken up by the RFC 9110 grammar rather than on whitespace, so "Windows NT 10.0" stays one item instead of three, and each piece is shown with what it means. Above all of it sits the fact that decides how much any of this is worth: a user agent string is whatever the client chose to send. Nothing measures it, every browser lets an extension change it, and a crawler that wants to look like a person simply writes a browser string.