What gets measured

I count visits to see what people read and where the site is slow or confusing. There are no ads, no profiles and no tracking across other sites, and nothing is sold.

For every visit, without asking, the site records:

  • which page was opened, how long it was on screen and how far down it was read;
  • the site you came from, and which share button you pressed, if any;
  • your approximate country and city, and the kind of device and browser;
  • which page you're on while you have the site open, for a live visitor count.

To tell visits apart, the server mixes the date, your IP address and your browser's name into a one-way hash. It changes every day, so it can't follow you from one day to the next, and the IP address itself is never stored. Nothing is saved in your browser for this.

To estimate your location and internet provider, the server sends your IP address to IPGeolocation. If that lookup fails or is incomplete, it can use ip-api as a fallback. The free fallback request uses an unencrypted connection. Map coordinates are rounded to about 11 km; IP locations are estimates, especially for mobile networks and VPNs. These results are cached under a one-way code for up to a day. Do Not Track and Global Privacy Control skip both lookups and retain only the country supplied by the network.

If you allow it

The first time you visit, the site asks whether it may remember your browser. If you say yes, it stores a random id in your browser so a visit next week counts as a return visit (on the analytics and the public visitor counter) and records anonymised sessions (every text field is masked) that I watch to fix rough edges. Recordings and analytics go to PostHog, hosted in the EU. If your browser sends Do Not Track or Global Privacy Control, the site treats that as a no and doesn't ask.

Other things the site keeps

  • On blog posts: how long you read, how far down you got, which headings you reached and which links in the article you clicked (the address only, without its query), counted once per day. With Do Not Track or Global Privacy Control, only the country and the kind of device go with it.
  • Liking a post stores a random id in your browser, so you can take the like back.
  • The terminal's leaderboard keeps the handle you pick, your score and your country, plus a random key in your browser that proves the handle is yours.
  • Cloudflare Turnstile checks that leaderboard scores come from real browsers.
  • Chatting with ophyx, my model, in the terminal: what you type goes to the model, which runs on Modal, and isn't stored or logged. To stop abuse, the chat counts messages against several codes at once and refuses when any of them is over its limit: a fingerprint of your browser, one of your device (screen size, time zone, processor count, operating system and which common fonts are installed, which every browser on it reports the same way), that device on your network (so switching browsers keeps your count), a random id kept in a cookie and in your browser's storage, and your IP address and its network (the first three of its four numbers). The site keys each of them with its own secret and keeps only those one-way codes. Turnstile checks the browser once per chat, a cookie keeps the chat open for 30 minutes, and an idle chat closes itself after 10. No fingerprint is impossible to dodge, so the model also has a fixed daily total for everyone.
  • When you start an ophyx chat or ask for an API key, and only then, the terminal also runs Fingerprint (fingerprint.com), a device-recognition service, to check that the browser isn't automated and to recognise it again in a private window. Fingerprint gets your browser's signals and IP address and returns an id; this site checks that id with Fingerprint's server and keeps only a one-way code of it. It is never used for analytics or ads. That code, and on a computer the code for your device on your network, link your browsers to one terminal profile, so your handle and scores follow you; a link nobody uses for 180 days is deleted. If an ad blocker stops Fingerprint, the chat still works; only a new API key needs it.
  • For the public usage numbers in the terminal (people, answers, tokens, speed, who used it most), each answer keeps its size and timings, the model, the time, and your terminal profile (or, without one, the one-way code for your device on your network). The name shown is your terminal handle. Never the prompt or the answer.
  • An API key for ophyx (when the API is open) is shown to you once. I keep only a one-way code of it, when it was made and last used, whether it was switched off, how many requests it made each day, and the codes above of the browser that made it. You can have one key at a time, and a key stops working 3 hours after it's made. Requests are counted the same way and their text isn't stored either.
  • Vercel, which hosts the site, counts page views and measures how fast pages load (Vercel Web Analytics and Speed Insights). It sets no cookies and keeps nothing that identifies you: it tells visits apart with a code that is thrown away after a day.
  • Short links on this domain (like the ones I share for promotions) count each click before redirecting: when it happened, the approximate location and internet provider looked up from your IP address, your device, browser and language, the page or app you came from, and any campaign tags in the link. Your IP address itself isn't kept, only its network part (the first three of its four numbers) and a one-way code used to count unique visitors. With Do Not Track or Global Privacy Control, only the country and device are recorded.

Analytics events are deleted after 90 days. Questions, or want something removed? Write to work@abhashchakraborty.tech.