Citation persistence: canonical definition

Citation persistence

Citation persistence designates the situation in which a deleted, retracted, corrected, or superseded source keeps influencing AI outputs not because it necessarily remains active as a direct origin, but because its content, framing, or formulations have already been redistributed through secondary surfaces that remain mobilizable.

In an interpreted web, a page may stop existing as a current page while continuing to exist as a reserve of quotations, a ranking template, a residual profile, or a reformulation already absorbed by the field. That indirect survival is what this term is meant to stabilize.

Definition

We call citation persistence the situation in which:

an origin source has lost its current status, availability, or primacy;
but citations, rankings, profiles, comparisons, summaries, excerpts, or other secondary artifacts continue to carry its framing;
and that secondary circulation is enough to maintain an effect on answer reconstruction.

Citation persistence therefore describes less the survival of a document than the survival of a statement that has already been redistributed.

Why it matters in AI systems

Deleting the origin is not enough. The field may already have replicated the signal.
Citation is not neutral. A third-party reprise may acquire more force than a more nuanced official source.
Local correction often fails. Correcting the origin without correcting the relay surfaces leaves part of the problem intact.

Citation persistence vs remanence vs archive

Interpretive remanence: the return of an old interpretation despite a visible correction of the canon.
Archive: preservation of an earlier state as trace, memory, or proof.
Citation persistence: maintenance of an effect through quotations, reprises, rankings, profiles, and secondary surfaces, even when the origin has disappeared or lost primacy.

An archive may feed citation persistence, but the two notions are not equivalent. An archive can remain passive. Citation persistence continues to act inside reconstruction.

What citation persistence is not

It is not sufficient proof that a model “keeps the page in memory.”
It is not necessarily a case of AI poisoning. No intentional corruption may be involved.
It is not mere popularity. The phenomenon implies that secondary reprises remain usable as authority inside the answer.

Practical indicators

A deleted page is no longer accessible, yet its framing still appears through third-party lists or profiles.
Outputs cite comparisons, directories, or rankings that continue to repeat an obsolete formulation.
The canon has been corrected, but the old wording keeps circulating through replicated excerpts.
The answer appears no longer to depend on the origin while remaining faithful to its earlier framing.

Why citation persistence is expensive to correct

Because it moves the center of gravity of the problem. The origin is no longer necessarily the dominant point of action. Once the framing has been redistributed, remediation often has to target:

the dominant third-party surfaces;
the most repeated formulations;
the environments where the old statement still functions as a shortcut;
the proof surfaces that distinguish quotation from distortion.

In other words, the problem stops being only endogenous. It becomes exogenous.

Minimum rule

Rule CP-1: when a source loses primacy but continues to influence outputs through secondary relays, the case must be treated as citation persistence. Remediation cannot be limited to the origin; it must map, qualify, and correct the reprise surfaces that prolong the framing.

Example

Case: a comparison page disappears, but directories, profiles, and rankings continue to repeat its verdict or category.

Diagnosis: citation persistence, potentially combined with surviving authority.

Expected correction: correction of the dominant relays, reinforcement of the canon, clarification of source hierarchy, and proof of fidelity against abusive reformulation.

Citation persistence