DOI

A DOI is a persistent identifier for a digital object: a string like 10.5194/gmd-15-1-2022 that resolves, through a registry, to wherever the object currently lives. The prefix identifies the registering organisation, the suffix is theirs to assign.

The distinction from a URL is the point, and it is organisational rather than technical. A URL names a location. A DOI names a thing, and separates the identifier from where it happens to be stored. When a journal changes platform or an institutional repository is restructured, URLs break and DOIs do not — provided the registrant updates the resolution target, which is the actual mechanism.

That proviso is worth being clear about. A DOI is a promise, not a property of the string. Persistence comes from an organisation committing to maintain the registration and from the registration agency requiring it, not from anything in the identifier itself. DOIs do rot when the registrant disappears. They are far more durable than URLs because the institutional commitment is real, not because the technology prevents failure.

Why they matter for data specifically

For articles, DOIs are administrative infrastructure. For data they change what is possible.

A dataset with a DOI is citable, which means the effort of producing and documenting it becomes visible in the same currency as a paper. This is a large part of why data publication has become viable at all — not because researchers were persuaded that sharing data is good, but because a citable object accrues credit.

It also makes reproducibility checkable. “We used ERA5” is ambiguous across versions and subsets; a DOI resolving to a specific versioned dataset is not. This is why FAIR treats persistent identifiers as foundational rather than as an implementation detail — the F in FAIR is largely about this.

Versioning is where care is needed. A DOI should identify a specific immutable version. If a dataset is corrected, the corrected version needs its own DOI, with the two linked. Repointing an existing DOI at revised data silently invalidates every analysis citing it, and nothing in the system flags this. Repositories like Zenodo handle it by minting a version DOI plus a concept DOI covering all versions, which is the pattern worth following.

See also: FAIR for the principles, datasets for what gets identified, provenance for the rest of the reproducibility chain, and metadata for the record behind the identifier.