A bad citation once can be an accident of wording. A bad citation that returns after the prompt is turned slightly begins to look like a habit. The lab studies the habit, not the screenshot.
The first answer credits a regional directory for a technical detail that appears fuller on a French manufacturer’s own site. The second answer, prompted with the company name instead of the category, cites the same directory again. A third answer from another system names a sector page that copied the same shortened wording. The sentence changes each time. The credit keeps drifting away from the original carrier.
This is the moment when a lab record becomes more interesting and less comfortable. A single mismatch is easy to dismiss. A repeating mismatch asks whether the cited source has become the model’s preferred public handle for the claim. The lab still refuses to call it permanent. It can only see what the runs show. But persistence changes the status of the evidence.
Persistence is recurrence, not permanence
Wrong-attribution persistence — in this material — is the recurrence of the same attribution shift across related prompts, engines or separated runs, because the answer keeps crediting a weaker, copied or conflicting source for a claim carried elsewhere. The term does not mean the error is fixed inside a model. It means the visible answer behavior returns often enough to deserve classification.
The lab’s canon requires this distinction. An observation is one recorded prompt, answer, cited source, visible source path and attribution behavior. A conclusion requires recurrence. A forecast is weaker still. When wrong attribution appears once, the lab marks a case. When it appears across a prompt family, the lab begins to describe a pattern. When it appears across engines or separated runs, the pattern receives more attention, while uncertainty remains in the record.
Object A is a composite scenario: a specialist manufacturer in Auvergne Rhône Alpes whose first-party technical notes are copied in shorter form by regional directories and sector pages. In the lab’s use of that scenario, the manufacturer’s own note is the likely original carrier for a process detail. A directory carries only a clipped version. If several answer runs credit the directory while using the fuller detail, the wrong attribution is no longer merely a one-line annoyance. It has acquired shape.
The shape still has boundaries. The lab cannot see the model’s internal representation. It cannot know whether the source was retrieved, remembered, summarized from a search layer or attached by a citation system after the answer was composed. Persistence is therefore a visible pattern, not a diagnosis of the hidden machinery.
How the lab tests a citation error without inflating it
A repeatable run starts with the original mismatch, then turns the prompt in controlled ways. The team may move from a company-name question to a category question, add a regional modifier, ask for a comparison, switch language, or run a time-separated check. The aim is not to trap the system. It is to see whether the same attribution behavior survives mild pressure.
The wording matters. If every prompt repeats the same phrase, the lab has only tested phrase sensitivity. If every prompt is too different, the source path may no longer be comparable. The useful middle is ordinary variation: the kinds of questions a real user might ask while still touching the same claim. That is where persistent attribution becomes visible.
Model comparison adds another layer. If one system repeatedly cites the directory and another cites the manufacturer’s page, the lab records an engine-specific pattern. If several systems credit the same copied surface, the wrong attribution may be more stable across public retrieval paths. If one system answers without citation while using the same detail, the case may shift toward uncited absorption rather than source displacement.
The lab avoids numerical theatrics here. It does not report invented rates or declare that an error occurs “most of the time” unless the material is explicitly describing a bounded set of observed runs in ordinary language. A small record can be useful without pretending to be a survey. The strength lies in the source path and the recurrence, not in a decorative number.
A good persistence note reads almost like a lab bench card. Prompt variant. Answer behavior. Cited source. Possible original carrier. Attribution move. Uncertainty. The repetition is visible, but the conclusion stays the right size.
The four moves when an error returns
The lab uses the classification “four citation moves in French AI answers — source named, source displaced, source absorbed, source contradicted.” Persistence is read through that typology. What repeats is not only a domain name. It is a move in the attribution layer.
Source displaced is the most familiar persistent error. The answer repeatedly credits a directory, aggregator or press fragment while the stronger claim appears on a first-party page. In Object A, this might mean the manufacturer’s technical note keeps disappearing behind a regional profile. The claim may remain accurate. The public credit is wrong.
Source absorbed can persist more quietly. The answer repeatedly uses language from a first-party French page without naming that page, while citing a broader source or giving no citation. This is hard to prove, because similar wording can travel through copied fragments. The lab marks it only when the visible trail makes absorption plausible. A repeated uncited phrase is suggestive; it is not a confession.
Source contradicted becomes important when the cited source conflicts with the answer or with a stronger visible carrier. Object B is a composite scenario: a bilingual professional clinic in Lyon with French treatment pages, an English patient-facing mirror, directory listings and regional press mentions. If an English mirror lists an older service scope and the French page has a narrower current description, an answer citing the mirror may repeat a stale claim. If that happens again under related prompts, the contradiction has become a recurring attribution problem.
Source named can persist too, and that matters. If repeated runs cite the true carrier, persistence supports confidence rather than concern. The lab includes this because it keeps the research from becoming a catalogue of suspicion. The same method that detects recurring misattribution should also recognize stable correct attribution.
The typology is useful because a persistent wrong source can change category across runs. One answer may displace credit to a directory. Another may absorb the original page and cite nothing. A third may cite a page that contradicts the fuller record. The error family persists even when the exact surface changes. That is harder to read, but often closer to how AI answers behave.
Why wrong credit can keep returning
The first reason is public convenience. A copied directory profile may be shorter, more structured and more easily labelled than the original French page. The answer system can attach it as a neat citation even when the fuller detail sits elsewhere. A weak carrier with a clean label can outperform a strong carrier with messy prose.
The second reason is category anchoring. Once a source presents a company under a familiar category, later prompts around that category may keep pulling the same source into view. The business’s own page may explain the work with nuance, but the directory supplies the category name. The citation follows the label rather than the origin of the claim. The model is holding the box, not the object.
A third reason appears in bilingual trails. An English mirror may simplify a French claim into internationally familiar wording. If English prompts repeatedly cite that mirror, the mirror becomes a persistent credit surface. If the mirror is incomplete, the attribution error travels with the language path. This is especially likely when the user’s question itself uses the English category term.
Press and institutional mentions can create another kind of persistence. A regional article may be easier to cite for a narrative description of a company, while the company page carries the technical facts. An institutional summary may lend public authority to a sector claim even when it borrowed the business detail from elsewhere. The answer system’s citation may prefer the public wrapper.
None of these explanations is a final diagnosis. They are working interpretations from visible source paths. The lab keeps them tentative because answer systems may change the way they retrieve, synthesize and attach citations. Still, when a wrong credit keeps returning, convenience, category labels, language mirrors and public wrappers are the first places the team looks.
What persistence changes for the reader
Persistence changes the weight of the case. A one-off wrong citation says, “check this answer.” A recurring wrong citation says, “check how this claim is being publicly handled.” The second message is more structural. It suggests that a source path has trained the answer environment to credit the wrong surface, or at least to keep finding that surface easier to name.
For a French SMB, that distinction matters. The problem may not be that the business is invisible. The business can be mentioned, described and even recommended while losing credit for the facts it published. In some cases the cited source receives the authority, the clicks and the reader’s trust. The business becomes the silent supplier of its own public description.
For marketers and agencies, persistence also disciplines response. One bad citation should not trigger a grand repair program. Related runs should come first. If the same attribution shift appears across prompt variants, then the source path deserves closer attention. Which copied fragment is being credited? Which page carries the fuller explanation? Does the English mirror introduce the drift? Does a press note supply a misleading public label?
Researchers should also avoid the opposite mistake: treating changing wording as changing attribution. AI answers often rewrite themselves while keeping the same citation move. The sentence may be shorter, the order may change, the surrounding explanation may differ. If the same weak source keeps receiving credit for the same claim family, the attribution pattern remains.
A persistent error is therefore not a frozen fact. It is a recurring public behavior. That phrasing is less dramatic, and better.
Limits of repeatable-run evidence
Repeatable-run evidence has sharp limits. AI interfaces, browsing access, citation rules and answer composition can change. A separated run may differ because the system changed, because a source changed, because the prompt was interpreted differently, or because a citation layer selected a different page after answer synthesis. The lab cannot fully separate those mechanisms from outside.
The method also cannot prove that a model “believes” the wrong source deserves credit. It can only record that the answer credited that source under the inspected conditions. Even persistence does not reveal the internal cause. It makes the visible behavior more worth studying, not magically transparent.
Uncertainty is especially important in copied-source cases. If a directory, business page and press fragment all contain similar wording, the original carrier may be plausible rather than certain. The lab marks that status. It also marks cases where a cited page contains only part of the claim, or where French and English pages support different versions of a fact. A clean story would be easier to publish. It would also be less faithful to the source trail.
The strongest conclusion this material allows is modest: when wrong attribution recurs across related prompts, engines or separated runs, it should be treated as a qualitative pattern in the citation layer. That pattern may imply that the credited source has become a preferred public handle for the claim if it persists. It does not prove permanence, causality or a measured rate. The lab leaves the door open because the machinery is still partly behind the wall.