I told readers to date a Vercel deploy by reading two headers, and both of them moved without a deploy

On 31 October, the author published a post about the three caches in front of PubTrivia's 72 public pages. They ended the section with a tip to read the age header to determine when a Vercel site last shipped. The age header reported about 1,052,000 seconds or 12.2 days, matching the last deploy. However, when the author re-ran the same sweep, nothing had shipped since 18 October.

The age header was supposed to be a deployment log, but it now pointed at a Sunday evening three weeks after the build. The trick was incorrect, and the investigation revealed something more useful. The deploy was the old one, as the git log ended at 18 October. Every asset URL on every page still carried the same deployment ID.

The cache object had moved, with 70 pages showing HIT and 62 distinct ages from 79,333 seconds to 226,655 seconds. The date header represented the moment the origin generated the page, while age showed how long ago that was. The date plus age was simply the current time, not when the page was built.

The ETag, which was printed in the previous post, was already dead. If the build had not changed, a prerendered page should be the same bytes it was last week, and the validator should still match. However, the ETag had changed while the content remained the same. The chain was that the edge evicted an object after a few days, Vercel re-rendered the page, the trace ID was fresh, the hash was new, and returning readers got a 82 KB response instead of a 304. The content had not changed since September; only the receipt had.