Nobody asks a marketing page for bytes 0 to 99. Browsers send Range for media and for resumed downloads, and the 111 pages of PubTrivia contain neither. So I spent an afternoon asking anyway, because a header nobody exercises is a header nobody has checked, and the answers turned out to be more interesting than the question.

The headline: every unsatisfiable range on our static files answers 416 Range Not Satisfiable and then sends the entire 953,385-byte file along with the refusal.

The matrix

$ curl -s -D - -o /dev/null -H 'Range: bytes=0-99' https://pub-trivia.app/faq 
Target Range asked Status accept-ranges Bytes returned
/faq (prerendered page) none 200 absent 78,655
/faq bytes=0-99 206 absent 100
/faq bytes=-100 206 absent 100
/guides/not-a-real-guide (404) bytes=0-99 206 bytes 100
/robots.txt bytes=0-9 206 bytes 10
/sitemap.xml bytes=0-49 206 bytes 50
/manifest.webmanifest bytes=0-9 206 bytes 10
/logo.svg bytes=0-99 206 bytes 100
/opengraph-image bytes=0-99 206 bytes 100
/_next/image?...&w=1200&q=75 bytes=0-99 206 absent 100
/demo-2.webp bytes=0-0 206 bytes 1
/demo-2.webp bytes=953380- 206 bytes 5
/demo-2.webp bytes=953380-999999 206 bytes 5
/demo-2.webp bytes=100-199,300-399 200 bytes 953,385
/demo-2.webp bytes=abc 200 bytes 953,385
/demo-2.webp bytes=99999999- 416 bytes 953,385
/demo-2.webp bytes=100-50 416 bytes 953,385
/demo-2.webp bytes=-0 416 bytes 953,385
/ (per-request page) bytes=0-99 200 absent 170,342
/pricing (per-request page) bytes=0-99 200 absent 102,002

The 416 that sends everything

Three different ways to write an unsatisfiable range, three identical outcomes:

$ curl -s -o body.bin -w 'code=%{http_code} got=%{size_download}\n' \ -H 'Range: bytes=99999999-' https://pub-trivia.app/demo-2.webp code=416 got=953385 $ head -c 8 body.bin | xxd 00000000: ffd8 ffe1 0080 4578 ......Ex 

The status line is correct, the content-range: bytes */953385 is correct, and then the body is the whole representation, starting at its first byte. content-length says 953385 too, so this is not a stray transfer: the response genuinely carries the file it just said it could not give you a piece of.

The point of 416 is to tell a client the range is impossible so it can stop. Here, a client that asks for a byte past the end pays for the entire object, and a client that retries with a corrected range pays for it a second time. Resumable downloads commonly probe with a range beyond what they have; on this origin that probe is a full download.

It arrives from cache as well, x-vercel-cache: HIT, so there is no origin round trip involved in producing it.

For comparison, the cases the same origin gets exactly right: a malformed unit (bytes=abc) is ignored and answered 200 with the whole file, which is what RFC 9110 asks for; If-Range with the current ETag returns 206; If-Range with a stale ETag returns 200 and the whole file, which is the correct way to tell a resuming client that its copy is out of date.

HTML serves ranges it does not advertise

A prerendered page answers 206 and names no accept-ranges at all:

$ curl -s -D - -o /dev/null -H 'Range: bytes=-100' https://pub-trivia.app/faq | head -3 HTTP/2 206 content-range: bytes 78555-78654/78655 content-length: 100 

So a client that checks before asking concludes ranges are unavailable, and a client that just asks gets served. The static file handler advertises it; the page handler does it silently. The 404 page does both, because it is a static file too.

The two sets of responses have different validators as well:

ETag Last-Modified accept-ranges
/faq and the other 108 prerendered pages "bu8obk0nja1oot" absent absent
/demo-1.webp and the other static files "5f5a2f9089..." present bytes

That static ETag is worth a second look. It is the md5 of the file in our repository:

$ md5 -q public/demo-1.webp 5f5a2f908982fddbe254f5984271e00e $ curl -sI https://pub-trivia.app/demo-1.webp | grep etag etag: "5f5a2f908982fddbe254f5984271e00e" 

Which is a good property: a ranged read of a file is safe to reassemble, because the validator is a hash of content rather than a label tied to a deploy.

Last-Modified is not nearly as good. Two requests two seconds apart:

age: 8 last-modified: Fri, 20 Nov 2026 10:58:28 GMT age: 10 last-modified: Fri, 20 Nov 2026 10:58:28 GMT 

Subtract the age and that timestamp is the moment this edge filled its copy, not the moment the file changed. The same file on the same origin reported 18 November an hour earlier, from an object that had been sitting there for two days. So If-Modified-Since on our static assets is comparing against cache furniture, and the ETag is the only validator here that means anything.

Two pages refuse the whole idea

The homepage and /pricing are the only per-request pages on the site, because they resolve the visitor's display currency on the server. They answer bytes=0-99 with 200 and the entire document: 170,342 and 102,002 bytes respectively, no content-length in the header at all because the response is chunked, no ETag, no range support. Streamed responses and byte ranges do not mix, and there is nothing to fix there.

A range turns compression off

This is the one that would matter if anybody did use ranges on HTML:

$ curl -s -o /dev/null -H 'Accept-Encoding: br' -w '%{size_download}\n' \ https://pub-trivia.app/faq 12218 $ curl -s -D - -o /dev/null -H 'Accept-Encoding: br' -H 'Range: bytes=0-99' \ https://pub-trivia.app/faq | grep -iE 'content-(range|encoding|length)' content-range: bytes 0-99/78655 content-length: 100 

No content-encoding. The 100 bytes are raw. Ranges are expressed against the identity representation here, so any partial read is uncompressed, and asking for the first 12,218 bytes of this page would cost more than taking the whole page compressed. The crossover for /faq is about 15 per cent of the document.

Multiple ranges in one request get the same treatment as nonsense: bytes=100-199,300-399 returns 200 and the full file. No multipart/byteranges anywhere on this origin.

The one range that earns its keep

$ curl -s -r 0-15 https://pub-trivia.app/demo-1.webp | xxd 00000000: ffd8 ffe1 00b4 4578 6966 0000 4949 2a00 ......Exif..II*. $ curl -sI https://pub-trivia.app/demo-1.webp | grep -i content-type content-type: image/webp 

ff d8 ff e1 is a JPEG with an APP1 segment. The CDN says image/webp because the filename says .webp, and the filename is wrong: all three of our landing page photographs are baseline JPEGs with the wrong extension, which I wrote about separately when I weighed them. Sixteen bytes is enough to prove it, out of 1,769,688.

Push it to 701 bytes and the whole metadata block arrives:

$ curl -s -r 0-700 https://pub-trivia.app/demo-1.webp | strings \ | grep -oE 'Exif|xpacket begin|dc:creator|dc:rights' | sort | uniq -c 1 Exif 1 xpacket begin 2 dc:creator 2 dc:rights 

0.04 per cent of the file tells you its camera metadata, its XMP packet and the name of the photographer who holds the rights, and the site has no attributions page. That is a genuine use for a range request on a static site: reading a file's headers without reading the file. It is also the strongest argument I have found for stripping metadata at build time, since anybody can do the above in one line.

What I cannot claim from this

Every measurement came from one network through one edge location, fra1, within about an hour. Vercel's edge is not something I configure: there is no vercel.json in this repository, so none of these behaviours are ours to set, and a different POP or a later platform release could answer differently. If you repeat this against your own Vercel project and get a clean 416 with an empty body, I would believe you over my own table.

The 416 body is reported by curl's own size_download, with a content-length to match and JPEG magic at offset zero, so it is not a counting artefact. What I have not established is whether it is a deliberate choice somewhere in the static handler or a bug, and I am not going to guess.

And the honest framing on all of it: no real visitor to a pub quiz website sends a Range header. This is a census of a code path that exists, works mostly correctly, is never exercised by the product, and is one careless retry away from being expensive for the person on the other end.

If you want to see the things these bytes add up to, the five free tools are at pub-trivia.app/tools, the comparison pages at pub-trivia.app/compare, and the quiz night product itself is free to start with no card at pub-trivia.app.