Websites as references: access dates and archives

Citing a journal article is a solved problem: the article has a permanent identifier, the identifier retrieves a record, and the record will say the same thing in ten years. Citing a web page is not solved, and the reason is a single structural difference that is worth understanding before any of the formatting rules make sense.

A URL is a location, not an identity

A DOI is a name for a work. Whoever hosts it, whatever the publisher does with its website, the DOI continues to mean that article. A URL is a name for a place — and the thing at a place can be replaced, moved, updated or removed without any of that being visible in the citation.

doi:10.xxxx/jhs.2022.0447
  "this work, wherever it lives"
  → resolves to a record
  → the record is fixed
  → checkable by anyone, forever

https://example.org/policy/guidance
  "whatever is at this address today"
  → resolves to a page
  → the page may have changed since you read it
  → and nothing about the URL says whether it did

That last line is the whole problem. A dead URL announces itself — you click, you get an error, you know. A URL that still works but now shows a revised page is silent. Your citation resolves perfectly and supports a claim the page no longer makes.

The dangerous web reference is not the broken one. It is the one that still works and no longer says what you cited it for.

What an access date is actually claiming

This is why access dates exist, and why they feel pointless until you know what they are for. "Accessed 14 March 2025" is not bureaucratic decoration. It is a claim with a specific meaning: on this date, this address showed the content I am attributing to it.

It converts an unfalsifiable citation into a falsifiable one. Without it, a reader who finds different content has no way to tell whether you misread the page or the page changed. With it, the question is answerable.

The styles differ on when it is required, and the pattern behind the disagreement is consistent: the more stable the source, the less the access date matters. APA asks for a retrieval date only when the content is designed to change — a wiki, a live statistics dashboard, a page explicitly labelled as updated. MLA asks for it more often. Vancouver expects a "cited" date on electronic sources as a matter of course. If your handbook is silent, include it: it costs eight words and answers a question a reader may genuinely have.

The archive answer

An access date documents the problem. An archived snapshot solves it.

If a source matters to your argument and lives only on the open web, archive it and cite the archived copy. The snapshot has a permanent address, shows exactly the content you read, and will still do so when the original is gone or rewritten.

WEAKEST
  Ministry of Health. (2024). Adherence guidance.
  https://example.org/policy/guidance

  → if the page changes, your citation silently
    becomes wrong

BETTER
  …https://example.org/policy/guidance
  (accessed 14 March 2025)

  → a reader who finds something different knows
    when to look for the version you used

STRONGEST
  …archived at
  https://web.archive.org/web/20250314…/policy/guidance

  → the reader sees what you saw, permanently,
    regardless of what happened to the original

Archive the page at the moment you cite it, not at the end of the project. A snapshot taken in your final week captures whatever the page says then, which may already differ from what you read in year two — and the snapshot's date is a matter of public record, so it will not support a claim about an earlier version.

The mirror problem

A specific trap worth naming, because it looks like good practice and is not.

When a journal article is behind a paywall, it is tempting to cite the copy you could actually reach — an author's upload to an academic social network, a departmental page, a file on a course site. Those copies are frequently incomplete, sometimes a different version from the published one, and their metadata is often scrambled: author names in the wrong order, missing co-authors, the wrong year.

We refuse to attach links from mirror hosts as verified sources for exactly that reason. They duplicate papers, scramble author metadata and bypass the verification the publisher's record provides — so a link that appears to confirm a reference can quietly confirm the wrong version of it.

The right move is to cite the work by its DOI, which names the published article, and leave finding a readable copy to the reader — who has a library, and interlibrary loan, and possibly better access than you do. A citation is a claim about which work you used, not a hosting recommendation.

Which web sources are worth citing at all

A question that comes before the formatting, and gets skipped: should this be a reference?

Web sources divide roughly into three kinds, and they carry very different risk. The institutional and permanent — a government dataset, a standards body, a regulator's guidance — are genuinely citable, usually versioned, and frequently have an identifier if you look for one. The institutional but living — a statistics dashboard, an agency's current-guidance page — are citable with an access date and an archive, because the whole point of the page is that it changes. And the ephemeral — a blog post, a press release, a vendor's marketing page — may be perfectly appropriate as a source and are the ones most likely to vanish entirely.

The test worth applying is not "is this respectable" but what is it carrying? A web page supporting a background statement is low risk; if it disappears, a reader loses a footnote. A web page carrying a number your analysis depends on is load-bearing, and losing it means a claim in your thesis can no longer be traced to anything.

Archive by consequence, not by prestige. The page nobody would question is exactly the one nobody will check until it is gone.

Where a source is load-bearing and ephemeral, the strongest available move is to cite both: the original address for provenance, and an archived snapshot for permanence. It looks slightly cumbersome in a reference list and it is the difference between a claim that can be verified in five years and one that cannot.

What a check can and cannot tell you about a web source

Worth being precise, because the honest limit here is larger than for journal articles.

"Does this address still resolve?"
  → checkable. A dead link is detected.

"Does the page still say what I cited it for?"
  → NOT checkable by any automated means.
    It requires knowing what claim you attached
    to it, reading the current page, and judging
    whether the claim survives.

For a journal article the two questions collapse into one, because the record is fixed: if the DOI resolves to the right work, the content is the content. For a web page they come apart completely, and only the first is mechanical. We re-resolve every identifier in your list, which catches dead and misdirected links — and stops precisely at the point where reading is required.

So the practical division is: let a check tell you which of your web sources have broken, and put your own time into the handful whose content matters most to your argument. Those are the ones to re-read and, ideally, to have archived at the time.

Five rules

Prefer a DOI whenever one exists. Many things that look web-only have one — reports, standards, datasets, conference papers. Search the title before settling for a URL.

Archive anything load-bearing, at the time you cite it. If an argument depends on it, do not rely on the open web to preserve it for four years.

Include the access date unless your style forbids it. It is cheap and it makes your citation checkable.

Cite the publisher's version, not a mirror. Even when the mirror is the copy you read.

And re-check your web sources before submission, along with everything else that ages — the refresh pass we describe separately exists because a reference list is a snapshot of a moving literature, and web sources move fastest of all.

One category to take seriously

Most broken web references are ordinary decay: a reorganised site, a retired page, a ministry that rebranded. Nothing about that reflects on anyone, and the fix is to find the new location or the archived copy.

There is one exception worth keeping in mind, particularly if any part of your reading list came from a generative tool. A URL that has never existed — a plausible-looking address on a real domain that returns nothing and shows no trace in any archive — is a different finding from a page that moved. If a source cannot be shown ever to have existed, then the work is not merely unavailable; citing it as evidence is academic misconduct, whoever assembled that part of the list. It is worth ten minutes to distinguish the two before you decide which you have.