Post
Every January, Every Footer Is Wrong
At some point on the first of January, every static site built before midnight starts telling visitors it was made last year.
1© 2007-2025 example.com
The fix takes ten seconds. Finding out you need it takes ten months, because nobody reads their own footer.
This is not an oversight in the generator. It is a consequence of something the generator does deliberately, and worth understanding before reaching for the obvious workaround.
The guarantee that causes the problem
SSG never calls the clock while rendering. Given the same sources, the same configuration and the same version, it produces the same bytes.
That guarantee earns its keep in ordinary places:
1a rebuild after an edit shows only what the edit changed
2a deploy uploads the files that differ, not all of them
3a CI check can compare output against a recorded baseline
4two machines building the same commit agree
The moment any template can ask what time it is, all four weaken. A diff after a one-word edit now includes every page carrying a timestamp. The baseline check fails every day for a reason that is not a regression.
So the generator declines to know. And a generator that declines to know what year it is cannot write the year into your footer.
Why the obvious fix is the wrong shape
The instinct is a template function:
© 2007-{{ now.Year }}
An agent editing over MCP reached for exactly this and found it did not exist. Reasonable instinct — most template languages have it — and it is worth being precise about why it is not the right answer here.
A function reads the clock when it is called. A build renders pages one after another, and a large site takes minutes. Nothing prevents this:
123:59:59.8 posts/first/index.html rendered → © 2026
200:00:00.3 posts/last/index.html rendered → © 2027
One build, two answers. Rare, and unpleasant precisely because it is rare: the site is internally inconsistent for a year, and nobody diffs a footer.
The same reasoning rules out a build that quietly re-reads the clock per page, per template, or per partial. The problem is not where the clock is read. It is that it is read more than once.
One value, read once
flowchart LR
A["build starts"] --> B["clock read once"]
B --> C[".BuildTime"]
C --> D["front page"]
C --> E["posts"]
C --> F["archives"]
C --> G["taxonomy pages"]
The clock is read when the generator is constructed, before any page renders,
and that one value reaches every template as .BuildTime:
© 2007-{{ .BuildTime.Year }} {{ .Domain }}
It is a time.Time, so the whole of Go's date formatting is available:
{{ .BuildTime.Format "2 January 2006" }}
Last built {{ .BuildTime.Format "2006-01-02 15:04 MST" }}
Two pages of one build cannot disagree, because there is nothing for them to disagree about. The value is decided before rendering begins.
Keeping reproducibility for whoever needs it
Reading the clock once is still reading the clock. A build at 09:00 and a build at 09:05 differ, which breaks the guarantee for the people who were relying on it: distribution packagers, CI comparing against a baseline, anyone verifying that a published artefact matches its source.
Those people already have a convention. SOURCE_DATE_EPOCH is the
reproducible-builds standard: a Unix timestamp in the environment that a build
must use instead of the current time.
1# Pinned: the same input produces the same bytes, forever
2SOURCE_DATE_EPOCH=1700000000 ssg content simple example.com
3
4# Unpinned: the real clock, which is what a scheduled rebuild wants
5ssg content simple example.com
Pin it and the build is a pure function of its input again. Leave it alone and the footer is correct. Nobody has to choose between the two globally — it is decided per build, by whoever is running it.
The value is interpreted as UTC deliberately. A pinned epoch exists so that two machines agree, and rendering it in local time would make the printed date depend on where the build ran, which is most of the way back to the problem.
A malformed value is ignored, not fatal
If SOURCE_DATE_EPOCH holds something that is not a number, SSG uses the clock
and carries on.
That is a deliberate choice and the opposite of what strictness usually suggests. The variable is frequently set by a surrounding toolchain the site owner does not control — a base image, a CI template, a build wrapper someone else maintains. Turning a third party's typo into a failed deploy makes the site owner responsible for a variable they never set and may not be able to change.
The build stays internally consistent either way, which is the property that actually matters.
The bug that only a real build could find
Worth recording, because it is the sort of thing unit tests are structurally unable to see.
SSG hands templates their data through four different paths. Pages and
archives get a map. The front page and taxonomy indexes get anonymous
structs, built in place with exactly the fields those views need.
Adding a value to a map is additive; a template that does not read it never notices. Adding a value to a struct means naming a field. Miss one, and Go templates do not fall back to empty — they fail:
1executing "index.html" at <.BuildTime.Year>:
2can't evaluate field BuildTime in type struct { Site *models.SiteData; ... }
Not a blank year. A hard error, on the most linked document the site has.
The first implementation added the field to both maps. Every unit test passed — they asserted on the data the maps carried, and the maps were correct. It was building an actual site with an actual footer that produced the error above.
The regression test is now a full build whose every template reads
.BuildTime, asserting that each rendered document agrees. It fails without
the fix, which is the only property that makes a regression test worth keeping.
Where this actually pays off
A footer is the visible case, but the useful one is a site that rebuilds without a person present.
1a nightly scheduled build
2a webhook from a CMS
3a republish trigger after content changes upstream
Any of those crosses midnight on the 31st of December on its own, and the footer is right the next morning. Nobody edits anything. Nobody remembers.
That is the whole feature: not that the generator learned to read a clock, but that it reads one exactly once, and can be told not to.