Privacy
This site measures how fast your device draws frames and shares the numbers, anonymously, so framebudget's scores and thresholds can be calibrated on real devices instead of one desktop. Other sites that use framebudget and turned its sharing on send the same anonymous reports here too. Nothing in them identifies you, and you can say no.
Your choice in this browser
What is collected
At most one report per browser and site each week, sent to /api/report on this site, from this site and from other sites that turned sharing on. It holds only these numbers and names:
| Field | What it is |
|---|---|
| Calibration version | Which set of reference numbers the scores were computed with. |
| Scores | The device score (100 is the reference phone), and the cold and warm benchmark scores, rounded to whole numbers. |
| Kernel rates | How much work each benchmark kernel did per millisecond, to 3 significant digits: float math, typed arrays, allocation and canvas paths, plus canvas rasterization from framebudget versions that measure it. |
| Clock resolution | The precision of the page's timer, in milliseconds. |
| Hardware hints | The number of logical cores and the rounded memory size the browser reports, the Compute Pressure state when available, and whether reduced motion is on. |
| Decision | The tier, the effects that ran, and the effects framebudget stepped down because frames were late. From framebudget versions that send them: the effects the page used, and the effect a page view turned on just below its threshold to calibrate it, if any. |
| Frame rates | The frames per second of the page and of each effect that reported frames, whole numbers. From framebudget versions that send it: the share of those frames that came late, to 2 decimals. |
The server adds a few coarse fields from the request and then discards the request headers:
| Field | What it is |
|---|---|
| Browser engine | The engine family (Blink, Gecko or WebKit) and its major version, for example Blink 140. |
| Operating system | The family only: Android, iOS, Windows, macOS, Linux, ChromeOS or other. |
| Form factor | Mobile or desktop. |
| Country | The two-letter country code Cloudflare assigns to the request. |
| Day | The date the report arrived, in UTC, without the time of day. |
Reports from other sites
framebudget is a library that other sites use. It never shares anything by default: a site's developer has to turn sharing on. When they do, the browsers of that site's visitors send the reports described above to https://framebudget.dev/api/report, and that destination cannot be changed or removed. A site may also send the same report to an address of its own; what happens there is up to that site and its privacy policy.
- What is in a report: exactly the fields listed above, nothing more. The server derives the same coarse fields from the request (engine and major version, operating system family, mobile or desktop, country, the day in UTC) and stores neither the IP address, the user agent string, nor the
Originheader, so not even which site sent a report. - What is not: no cookies, no identifiers, no page addresses, referrers or time of day, as listed below.
- How often: only a sample of page views reports, 10% unless the site chose otherwise, and each browser reports at most once every 7 days per site (sooner only when the calibration changes). Those sites also fetch
/api/calibrationfrom here at most once a day, without cookies. - Never when Global Privacy Control or Save-Data is on, on any site, or while a device is simulated or a tier is forced.
- Saying no: Global Privacy Control or Save-Data stops it on every site. The Don't share choice on this page covers framebudget.dev only, since it lives in this site's storage; another site turns sharing off with its own setting (
share: null), so ask that site or use Global Privacy Control.
These reports are kept and used exactly like this site's own: same database, same 400 days, only to calibrate framebudget, and only aggregated numbers are ever published.
What is not collected
- No IP address. Your browser sends it with every request, as it does to any site, but it is not stored.
- No user agent string, only the engine, its major version and the operating system family derived from it.
- No cookies, no identifiers, no fingerprint, nothing that links one visit to another.
- No page addresses, referrers or search terms, no time of day, and not which site a report came from.
- No analytics or advertising scripts. On this site, the report goes to this site only.
When it is sent
- At most once every 7 days per browser (sooner only when the calibration changes), as one
navigator.sendBeaconrequest when the page is hidden or closed, after it has loaded. The date of the last report stays in your browser's local storage, next to the effects framebudget learned to skip. - Never when Global Privacy Control or Save-Data is on. Both are checked again right before sending.
- Never while a device is simulated or a tier is forced, since those numbers would not describe your device.
- Never after you choose Don't share, here or on the notice at the bottom of the home page. The choice is kept in this browser's
localStoragefor framebudget.dev, and covers this site only: other sites follow their own setting, and Global Privacy Control or Save-Data.
The site also fetches /api/calibration at most once a day, without cookies, to use updated reference numbers on your next visit. It follows the same rules.
Storage and retention
Reports are stored in a Cloudflare D1 database, behind the Cloudflare Worker that serves this site. A daily job deletes every report older than 400 days. The reports are used only to calibrate framebudget: reference rates, thresholds, effect weights, the load capacity and the score scale. Only aggregated numbers, such as medians and percentiles, are ever published.
On your device
framebudget keeps one localStorage entry per site with the cached score and the effects that stuttered, so the next visit starts with the right effects. The site also remembers your sound and sharing choices. None of this leaves your device.
The lab
The lab is a separate, opt-in test. Nothing in it runs until you press Run the test and share the results on that page. It then runs the site's effects one at a time on your device for about a minute and sends one record to /api/lab/runs on this site:
- Device: the framebudget, calibration and lab test versions, the device score and the cold and warm benchmark scores (one decimal), the kernel rates (3 significant digits), the clock resolution, the number of logical cores and the rounded memory size, the measured refresh rate, the pixel ratio (one decimal), the window width rounded to 100 pixels, and whether reduced motion and Save-Data are on.
- Steps: for each step, its name, the effects that ran, the number of frames, the duration, the median, 95th percentile and longest frame time, and how many frames were later than one and a half refresh intervals. Also the main-thread work per frame (the page's JavaScript plus style, layout and paint): its mean, median and 95th percentile in milliseconds to 2 decimals, and the number of frames with a work sample. Compositor and GPU work, such as a backdrop blur, is not in it; it shows up as late frames when it makes the device drop them. Each step is added to the same record as soon as it ends.
- From the request: the browser engine and its major version, the operating system family, mobile or desktop, the country code Cloudflare assigns, and the day in UTC, like the reports above. The IP address and the user agent string are not stored.
Cloudflare Turnstile, a check by Cloudflare that a browser and not a script pressed the button, runs when you press the button, and only on the lab page. Its script comes from challenges.cloudflare.com and is subject to Cloudflare's privacy policy; this site only receives its pass or fail answer.
Each test gets a random id and a random write key from the server. Both live only in the page's memory and in the record, never in your browser's storage, and the key stops working when the test ends or after 15 minutes. It is erased from the record when the test ends, or by the next daily cleanup. When the test is done, the browser keeps framebudget-lab in localStorage with the calibration version, the version of the test and the day, and no id, so the lab does not offer the test again for the same calibration and test.
Lab records are kept in the same database and deleted after 400 days. The server accepts a limited number of new tests per day, and stops accepting more until the next day once it is reached.
Contact
Questions, or a request about this data: privacy@framebudget.dev. Since reports carry no identifiers, there is no way to find the reports of one visitor, so none can be shown or deleted individually.