If you imagine a typical ecommerce application makes a purchase with a HTTP request like “POST /store/checkout HTTP/1.1” we can look for requests similar to this to understand the activity.
I basically wouldn't jump to "Cloudflare have a bunch of sensitive data and will use it in bad ways", I suspect they have less data than we might assume from the article, and in general their security/privacy stance is great.
I'm not sure I understand this sentence. What about analyzing HTTPS traffic is analogous to analyzing "semi-public" data? By design and convention, HTTPS traffic is considered private data. Portions are by technical necessity still public, such as the hostname and IP addresses. In order to produce these findings, private data (which is certainly not public or semi-public) such as full URLs and user-agents needed to be analyzed.
Do we have any sense how often a mobile device is on wifi? For my brother it's 0%. For me it's 90%.
Most of it seems to be domain and user-agent based. I'd suspect you could even guess the device type based on IP with reasonable accuracy. It wouldn't be entirely correct, but with the kind of scale Cloudflare deal with it doesn't have to be to be useful.
The domain (which is public) gives you the info about whether users are browsing or checking out (if a browser hits api.stripe.com, worldpay, PayPal, some checkout API domain for Amazon, etc, then you can infer).