Notifio has 38 pages, and two of them are hubs: notifio.app/alerts lists the rental sites we support and notifio.app/guides lists the long-form pages. Both are grids of cards, which in Next.js means a grid of <Link> components, which means prefetching.

I have measured this before on a sister site, where loading a 13 KB page pulled 37 KB of other pages in 21 requests. That post worked out what the _rsc query token is keyed on. This one is about a worse case and a thing I did not notice then: the same destination gets fetched four times, at four different sizes.

Cold loads, no interaction at all

First, every URL in the sitemap, loaded once in a fresh context, three seconds of sitting still, nothing clicked and nothing scrolled:

total requests across 38 cold page loads 1,096 of those, _rsc prefetch requests 516 prefetches per page /guides 32 /alerts 30 /compare 20 /compare/*, /for/*, /guides/* (13 pages) 15 /alerts/* (15 pages) 11 /, /pricing, /download, /help, /about, /terms, /privacy 10 to 11 

Forty-seven requests to render notifio.app/guides, and 32 of them are for pages the reader has not asked for. The floor is ten, which is the nav and the footer: those links are on every page, they are in the viewport at load, so they are prefetched on every page.

Then I scrolled to the bottom

This is the measurement that changed my mind about it. Load the guides hub, scroll once to the bottom, wait five seconds:

prefetch responses 73 distinct URLs 73 (every one unique: no two share a query) distinct destination paths 23 distinct _rsc tokens 23 total bytes 365,776 

357 KB of other pages, to read a page that lists five guides. The reader clicked nothing. And the per destination breakdown is the part I had not seen before:

destination requests bytes of each /for/amsterdam 5 767, 2721, 365, 365, 23399 /alerts/rightmove 4 766, 365, 23631, 2676 /terms 4 655, 2580, 365, 23134 /about 4 655, 365, 2472, 9678 /alerts/kamernet 4 765, 2678, 365, 21107 /help 4 654, 2518, 365, 18198 /privacy 4 657, 2584, 365, 15292 /guides/rental-scams-netherlands 3 781, 2747, 2545 /pricing 2 12, 12 

Four shapes keep recurring. A short one of 650 to 790 bytes, one that is exactly 365 bytes every single time, one of about 2,500 to 2,750 bytes, and then the whole page payload at 9 KB to 25 KB. The App Router has more than one kind of prefetch, and the Vary header on these responses names all the inputs it keys on:

vary: rsc, next-router-state-tree, next-router-prefetch, next-router-segment-prefetch x-vercel-cache: HIT 

Four request headers participate in the cache key, so a destination asked for at four different granularities is four different cache entries and four round trips. The CDN says HIT on all of them, so none of this costs us anything on the server. It costs the reader, on their connection, before they have expressed any interest.

Our terms of service is in there because of the footer

where the guides hub's 23 prefetched destinations come from content cards 8 5 guides and 3 audience pages nav 4 /, /alerts, /pricing, /download footer 14 /terms, /privacy, /about, /help, /compare, /guides, /alerts and five individual /alerts/* pages (overlapping the nav on three of them) 

notifio.app/terms is the biggest source file on the site, and the footer links to it from every page. So scrolling to the bottom of a hub page downloads 23 KB of our terms of service. 15 KB of the privacy policy. 18 KB of the help page. A reader who lands on the guides hub, scrolls, and leaves, has downloaded the legal pages and never seen them.

I counted a while ago that 64 per cent of the internal links on this site are the footer. That was an SEO observation at the time. It turns out to also be a bandwidth one: a footer with 14 links is 14 prefetch candidates sitting in the viewport at the end of every page.

Why more than one request per destination

The hub cards are the other half. Each card on the guides hub wraps the whole tile in a link and then has a second link on the title, so one destination is two <Link> instances, and the browser does not get to dedupe them because the segment prefetch and the full prefetch are different requests with different _rsc tokens. /for/amsterdam managed five requests, two of which were the identical 365 byte shape.

Run to run this varies, which is worth saying plainly: repeat runs of the same scroll gave me 62 to 73 responses and 284 KB to 366 KB, depending on how long cards sat in the viewport. The shape of the result was stable every time.

What I am changing

Not prefetching itself. On a site whose whole pitch is being first to something, pages that open instantly when you click a card are worth paying for, and the HTML behind them is prerendered and cached at the edge so the server never feels it.

Two things are clearly wrong though:

  1. prefetch={false} on the footer. Nobody has ever clicked from a guide into our terms of service, and those 14 links are the reason a text page pulls four legal and help payloads on every scroll. This is one prop in one file.
  2. One link per card. Making the tile the only link, with the heading inside it, removes the duplicate instance for every card on both hubs.

My estimate is that the two together take the guides hub from 73 prefetches to somewhere near a dozen, one per destination a reader might plausibly click. I will measure it again after, because the only thing this exercise has taught me reliably is that the number of requests a Next.js page makes is not something you can read off the source.

If you want to see it on your own site, filter the network panel to _rsc and scroll to the bottom of your busiest index page. The row count is the interesting number, not the bytes.