The repository behind PubTrivia has not changed since the commit that is deployed. So I built it on my laptop, fetched all 111 pages of the live site, and compared them file by file. The question was simple: how much of what the CDN serves is a function of our source, and how much is a function of the machine that built it?

Answer: every word of content is identical, and the difference is 404,095 bytes of deployment id, two meta tags, and the names of 32 JavaScript files.

Method

$ curl -s https://pub-trivia.app/sitemap.xml | grep -o '<loc>[^<]*' | sed 's|<loc>||' > urls.txt $ wc -l urls.txt 111 

Fetch each URL anonymously, then pair it with the HTML the build left in .next/server/app. 109 of the 111 pair up. The two that do not are the homepage and /pricing, which are the only per-request pages on this site, so the build writes no HTML for them at all.

Then two comparisons: raw bytes, and visible text with every <script> block and every tag stripped and whitespace collapsed.

The raw numbers

Live Local build Difference
109 pages, total bytes 8,520,294 8,116,199 +404,095
/faq 78,655 74,953 +3,702
/features 82,565 78,866 +3,699
/legal/privacy 43,291 41,047 +2,244
/tools/round-planner 73,797 69,497 +4,300

Where the 404,095 bytes go: 9,562 occurrences of ?dpl=dpl_4xbGu8hwmfCw7SjeoEiF32fkeCcC across the 109 pages, between 48 and 103 per page with a median of 88. Vercel stamps the deployment id onto every asset URL in the document and again inside the flight payload, and adds data-dpl-id to the <html> element and one Sentry trace meta tag.

You can see one of them without leaving this page:

$ curl -s https://pub-trivia.app/faq | grep -o 'dpl_[A-Za-z0-9]*' | sort -u dpl_4xbGu8hwmfCw7SjeoEiF32fkeCcC $ curl -s https://pub-trivia.app/faq | grep -c '?dpl=' 88 

The content

With scripts and tags removed, 109 of 109 pages are character-for-character identical to the ones my laptop produced. Not one heading, paragraph, price, link text or table cell differs between the build on this machine and the deploy that has been serving for days.

That is the result I was actually after, and it is the boring one, which is the correct outcome for a static marketing site whose content lives in TypeScript files.

What still differs, and why

After normalising the deployment id away, /faq still has 73 differing regions. All of them are names:

live: <link rel="stylesheet" href="/_next/static/chunks/00_d8kbt9dn9n.css" disk: <link rel="stylesheet" href="/_next/static/chunks/0pqey50yjq7_l.css" live: 2:I[880492,["CHUNK","CHUNK"],"ThemeProvider"] disk: 2:I[17738,["CHUNK","CHUNK"],"ThemeProvider"] 

Both CSS chunk names differ. The numeric module ids inside the flight payload differ. And of the 33 distinct JavaScript chunks the live pages reference, 32 do not exist in my build at all.

The cause is in the first four lines of my own build output, which I had been ignoring for months:

⚠ Warning: Next.js inferred your workspace root, but it may not be correct. We detected multiple lockfiles and selected the directory of /Users/daniel/package-lock.json as the root directory. 

There is a 5,247-byte package-lock.json sitting in my home folder from some unrelated experiment, and because it is higher up the tree than our pnpm-workspace.yaml, Next picks my entire home directory as the project root. Module identity in Turbopack is derived from the module's path relative to that root, so every id in my build is computed from a path three directories longer than it should be.

The proof is in the server chunk filenames, which spell the path out:

$ ls .next/server/chunks | grep -c Documents_programs_pub-trivia 16 $ ls .next/server/chunks | head -1 Documents_programs_pub-trivia_0-uqo0e._.js 

On Vercel the root is its own checkout directory, so the same source produces different names. Same bytes of content, different labels on the boxes.

The one chunk that matched

One of the 33 live chunk names does exist in my build, and it is identical:

$ curl -s https://pub-trivia.app/_next/static/chunks/0cz1d0mv5g_q7.js -o live.js $ md5 -q live.js .next/static/chunks/0cz1d0mv5g_q7.js 846118c33b2c0e922d7b3a7676f81f6f 846118c33b2c0e922d7b3a7676f81f6f 

112,594 bytes, byte-identical across two machines. It is the noModule legacy bundle, the core-js polyfill chunk that ships on every page for browsers that do not support ES modules and is therefore never fetched by any browser that visits us. The single piece of our deployment I can reproduce exactly on my laptop is the piece nobody downloads.

Two local builds of the same commit

If the environment is what makes the difference, then the same environment twice should make none. I built it again, six minutes later, and compared the two outputs file by file.

Count
Files in both builds 2,710
Byte-identical 1,651
Differing 1,059

Every one of the 1,039 server chunks, all 80 client chunks, every manifest and every trace is identical. The 1,059 that differ are 119 .html, 939 .rsc and one .body, and in the pages the difference is one string:

$ cat .next/BUILD_ID 6Mh6nPjo_wz-TeOOS6oxY 

faq.html is 74,951 bytes in both builds and differs in exactly four places, all of them inside that 21-character id, which appears once per page as "b":"<id>" in the flight payload.

The one .body that differs is sitemap.xml.body, and it is 12,899 bytes in both. The difference is three timestamps:

build 1: <lastmod>2026-11-20T10:36:00.022Z</lastmod> build 2: <lastmod>2026-11-20T10:42:37.427Z</lastmod> 

Three URLs on this site take their lastmod from the build clock instead of from the content registry, because they are the three legal documents and they are the only pages with no registry node. Our own app/sitemap.ts says so in a comment, and you can see them in the live sitemap as the only three entries that are not midnight UTC:

$ curl -s https://pub-trivia.app/sitemap.xml | grep -o '<lastmod>[^<]*' \ | sed 's|<lastmod>||' | sort | uniq -c | sort -rn 44 2026-09-14T00:00:00.000Z 35 2026-11-16T00:00:00.000Z 14 2026-10-15T00:00:00.000Z 9 2026-11-04T00:00:00.000Z 4 2026-11-15T00:00:00.000Z 3 2026-11-16T18:52:56.539Z 1 2026-10-18T00:00:00.000Z 1 2026-10-03T00:00:00.000Z 

That lone 18:52:56.539Z row with three pages in it is the deploy time of the running build. The other 108 are dates somebody wrote down on purpose.

What this does not prove

The "visible text" comparison strips script blocks, which is where the flight payload lives, so it is a claim about rendered content and not about the whole response. The byte comparison covers that gap only as far as the normalisation goes, and my normalisation is where I nearly published a false finding: deleting data-dpl-id="..." left <html lang="en"> with two spaces, which showed up as the first difference on all 109 pages and looked for a few minutes like Vercel rewriting our markup. It was my regex.

I have also not tested the fix. Deleting that stray lockfile, or setting turbopack.root in the config, should make my module ids match a build rooted at the repository, but it would not make them match Vercel's, whose root is a different absolute path again. Reproducible chunk names across machines would need the ids to be derived from repository-relative paths, which is a Turbopack question and not a configuration I can set.

And "the repository has not changed" is a claim about git status and git log, not about the deploy. The deployment id above is the only honest identifier of what is actually running.

The useful conclusion

For a content site, a deploy is a pure function of the commit plus two pieces of environment: a random build id, and the clock at the moment the sitemap was generated. Everything a reader sees is in the repository. Everything else is a label.

If you would like to read the identical words off the CDN instead of out of my .next, they are at pub-trivia.app/faq and pub-trivia.app/guides/how-to-host-a-pub-quiz. The free tier needs no card: pub-trivia.app.