notifio.app shows the app working instead of a screenshot. Four animated mockups run down the homepage: a browser with a listing appearing in it, a notification sliding in, a reply being typed. I wrote about the 997 lines behind them a while ago, and one line in that post has been bothering me since: "timers that run whether you scrolled there or not".
That was an observation from reading the code. This is the measurement.
Sixty seconds of nothing
Load a page, wait for it to settle, then do absolutely nothing for a minute while reading Chrome's own counters at both ends:
const cdp = await ctx.newCDPSession(page); await cdp.send('Performance.enable'); const before = await cdp.send('Performance.getMetrics'); await page.waitForTimeout(60000); const after = await cdp.send('Performance.getMetrics'); // TaskDuration, ScriptDuration, RecalcStyleDuration, LayoutCount, RecalcStyleCount Plus a MutationObserver on document.body with every flag turned on, counting what changes.
| Page | Task | Script | Style recalc | Layouts | Style recalcs | DOM mutations |
|---|---|---|---|---|---|---|
| / | 8,756 ms | 1,828 ms | 1,132 ms | 956 | 6,531 | 11,468 |
| /terms | 1,329 ms | 208 ms | 0 ms | 0 | 0 | 0 |
| /help | 1,382 ms | 218 ms | 0 ms | 0 | 0 | 0 |
The two control pages are the important column. Not "a few": zero layouts, zero style recalculations, zero DOM mutations in sixty seconds. A server-rendered page with nothing animated on it is genuinely idle, and that is the baseline the homepage is being measured against.
About the homepage's 8,756 ms. Roughly 1.35 seconds of it is my own instrument, because counting frames with requestAnimationFrame costs 7,200 callbacks a minute and the control pages show what that costs. So the honest figure for the page's own work is about 7.4 seconds of main thread per minute, or 12 per cent of one core, held for as long as the tab is open.
I ran the homepage five times and the shape is stable even when the machine is not:
| Run | Task | Style recalcs | Mutations |
|---|---|---|---|
| 1 | 8,100 ms | 6,523 | 11,466 |
| 2 | 10,556 ms | 6,531 | 11,288 |
| 3 | 8,678 ms | 6,522 | 11,481 |
| 4 | 9,353 ms | 6,529 | 11,284 |
| 5 | 8,756 ms | 6,531 | 11,468 |
Task duration moves with whatever else the laptop is doing. The work itself does not: 6,525 style recalculations and about 11,400 mutations a minute, every time.
Which demo, and what kind of change
The same observer, bucketed by the nearest section and the mutation type, over 30 seconds:
2611 MAIN > SECTION > Watches your search in the background [attributes:style] 1311 MAIN > SECTION > Replies to new listings for youBeta [attributes:style] 805 MAIN > SECTION > Get notified before anyone else responds [attributes:style] 338 MAIN > SECTION > Replies to new listings for youBeta [characterData] 45 MAIN > SECTION > Replies to new listings for youBeta [childList] 27 MAIN > SECTION > Never miss a rental listing, ever again. [childList] 24 MAIN > SECTION > Get notified before anyone else responds [childList] 14 MAIN > SECTION > Never miss a rental listing, ever again. [attributes:style] 13 MAIN > SECTION > Watches your search in the background [childList] 5,202 mutations in half a minute, and 4,741 of them, 91 per cent, are writes to one inline style attribute. The headings double as component names, because each section holds exactly one demo:
| Section | Component | Mutations in 30s | Share |
|---|---|---|---|
| Watches your search in the background | BrowserDemo | 2,624 | 50.4% |
| Replies to new listings for you | AutoReplyDemo | 1,694 | 32.6% |
| Get notified before anyone else responds | NotificationDemo | 829 | 15.9% |
| Never miss a rental listing, ever again. | HeroDemo | 55 | 1.1% |
The hero, which is the one every visitor sees, is responsible for one per cent of the work. The expensive one is the second section, which most readers scroll past in a second or two.
Why that one costs the most is visible in its source. Each listing is a motion.div with layout and a spring:
<motion.div key={l.price} layout initial={l.isNew ? { opacity: 0, y: -10 } : false} animate={{ opacity: 1, y: 0 }} exit={{ opacity: 0 }} transition={{ type: "spring", stiffness: 340, damping: 30 }} A spring has no fixed duration, it runs until it settles, and layout means the element's box is measured and re-measured while that happens. That is where the 945 layouts a minute come from. The driver is an async loop that never ends:
await delay(1400); setShowNew(true); setShowToast(true); await delay(3000); setShowToast(false); await delay(3500); setShowNew(false); await delay(600); if (!cancelled.current) run(); 8.5 seconds per cycle, three demos running their own cycles out of phase with each other, and the result is a page that is never quiet. Counting mutations second by second:
158 188 56 37 175 315 547 40 141 163 37 221 96 432 126 463 88 25 53 129 128 147 576 482 50 44 223 2 35 12 Zero of the 30 seconds are idle. The quietest second still had 2 mutations, the busiest had 576.
Two things that change nothing
Scrolling the demos off screen. I scrolled to 4,836 pixels of a 5,736 pixel document, so all four mockups were far above the viewport, and measured another minute:
| State | Style recalcs | Mutations |
|---|---|---|
| Demos on screen | 6,531 | 11,468 |
| Demos scrolled out of view | 6,529 | 11,284 |
Identical. There is no IntersectionObserver anywhere in those components, so being invisible costs exactly as much as being watched. The earlier post said this would be true from reading the code. It is nice, in a bleak way, to have the number.
Forcing reduced motion. Launching a context with reducedMotion: 'reduce', which makes matchMedia('(prefers-reduced-motion: reduce)').matches return true in the page:
| State | Style recalcs | Mutations |
|---|---|---|
| No preference | 6,531 | 11,468 |
prefers-reduced-motion: reduce | 6,522 | 11,481 |
Also identical, which confirms the accessibility gap I found by reading the code with a measurement instead of a grep. A reader who has asked their operating system to stop animations gets every animation, at full cost.
What this is actually worth
For a desktop visitor on mains power, 12 per cent of one core on the page that explains the product is defensible. The demos are the product pitch. A static screenshot of an app that watches for new listings cannot show the thing it does.
For a phone, it is not defensible, and the three numbers that make that case are the ones above: the work continues when the demos are off screen, it continues when the reader has asked for less motion, and it never pauses between cycles. The fix is two small pieces of work that the measurement has now sized for me:
- an
IntersectionObserveraround each demo's loop, which removes the cost for every reader who has scrolled past, and since the second section is 50 per cent of the total, that is most of it - a
prefers-reduced-motioncheck that holds each demo on its final frame, which is both the accessible behaviour and free
What I will not do is replace framer-motion. 945 layouts and 6,525 style recalculations a minute is what three concurrent spring animations cost, and the spring is the reason the mockups look like software rather than like a GIF. The problem was never the library. It was that nothing ever told it to stop.
You can check the whole thing yourself in a console on notifio.app: open DevTools, Performance Monitor, and watch the CPU line while you do nothing. Then do the same on /terms and watch it sit at zero.

