A French business can appear once in an AI answer and still remain fragile. The question is whether that first selected source becomes a groove the system follows again, or only a footprint left by one run.
The first appearance feels larger than it is. In a composite scenario built around a bakery equipment supplier near Tours, an AI search system finally selects the supplier’s own French product page instead of the older directory record. The answer names the business, uses the right town and lists the owned site as a source. For a reader watching the trace, it looks like a small door opening.
Then the same query is run again under comparable conditions. The owned page does not disappear completely, but the directory returns to the top of the source trail. In another run, a review page takes the visible slot. The answer still sounds sensible. The retrieval evidence has shifted under its feet.
A citation is an event, not a contract
Repeat retrieval — this is the reappearance of the same business evidence under comparable query frames because the retrieval path remains stable enough to select it again. The definition is deliberately narrow. It does not mean the answer repeats a sentence. It does not mean the company is mentioned somewhere in the output. It means the same page, listing or source trail is visibly retrieved again when the lab keeps the query frame as steady as the interface allows.
Indexe Clair starts from a plain suspicion: one citation can mislead people into thinking a business has crossed a permanent threshold. That is tempting because citation looks official. A source box, a linked page, a neat answer paragraph — together they create the feeling that the system has now “learned” the business. The lab avoids that word. From the outside, the team can observe selection, not internal learning.
A single source selection shows that retrieval was possible in that run; it does not show that retrieval has become dependable.
The distinction matters for French SMBs because their evidence often sits in several public places at once. An owned site, directory-style listings, review profiles, municipal mentions, regional press and sector directories can all point toward the same business with small differences. If one source wins once, it may have won because of the query wording, interface state, available live retrieval, ranking order or a temporary preference in the source trail. The next run may rearrange the shelf.
This is not failure in the dramatic sense. Search systems vary. AI search systems vary with an extra layer of answer synthesis on top. The lab’s point is quieter: a citation is the smallest unit of visible success, not proof of durable visibility.
What repeated runs actually compare
The lab keeps the comparison close to the conditions that can be observed. The query wording is held stable. The language is recorded. The location frame is kept: “near Tours” is not casually swapped for “Centre-Val de Loire,” and a French query is not silently replaced by an English one. The system and interface conditions are noted as far as the interface allows. Then the team watches which business, page and source trail appear.
The output can vary in several ways. Sometimes the answer wording changes while the source remains the same. That is a synthesis change, less central to this material. Sometimes the source list changes while the business remains the same. That is more interesting, because the entity may be stable but the selected evidence is not. Sometimes a competing business appears. That is a stronger retrieval shift. Sometimes the owned site appears in one run and only a directory appears in the next, which raises the question this work-item is built around.
Indexe Clair does not count repetition as a polished paragraph returning. The group is looking for whether the same evidence trail survives.
In the composite Lyon peri-urban repair service scenario, repeat retrieval can be especially thin. One run for a French service query may select the independent repairer’s service page. A second run may pull a larger chain from Lyon because the geographic intent collapses toward the city. A third may mention the independent but cite a review profile. The business has not become invisible in a simple sense. It has become unstable across retrieval paths.
That instability is the object of study. It tells the reader that the first selected source did not necessarily build a lasting route. It may have been closer to a ferry crossing than a bridge: real, useful, visible in the moment, but not guaranteed the next time someone arrives at the bank.
The four gates of repeat selection
The lab applies its anchor classification here as a way to name the stages where repeat retrieval can break. A French business must pass four retrieval gates — discovered page, indexed entity, ranked evidence, selected source. In a repeat run, the question is not only whether it passed once. The question is which gates stay passable across comparable prompts.
At the discovered page gate, the owned site or listing must remain reachable to the system. A deep service page can appear once because the query closely matches its wording, then vanish when the system takes a broader route through directories. At the indexed entity gate, the business has to be recognized as the same entity, not split across a legal name, trade name, old address or bilingual variant. At the ranked evidence gate, the page must compete with other public traces. At the selected source gate, the interface must actually show or use it.
One citation can occur with weakness at any earlier gate. A page may be discovered by a narrow query but not strongly attached to the entity. A business may be indexed as an entity, while its owned site ranks below review pages. A source may be selected once because other traces were not retrieved in that run. The four gates prevent the lab from calling all repeat failures by the same name.
This is where the team’s work differs from a simple visibility check. A marketer might ask, “Did the AI cite us?” The lab asks, “Which gate did the cited source pass, and did it pass again?” That second question is less satisfying. It is also more useful.
A repeated selected source suggests a stronger retrieval path, while a repeated mention without the same source suggests only partial stability.
The lab also watches for source substitution. If a system keeps naming the business but alternates among the owned site, a directory and a review page, the entity may be stable while evidence selection remains unsettled. For some business needs, that may be enough. For source-trail research, it is a different pattern from stable source selection.
Why one successful selection may not hold
There are several ordinary reasons a selected source fails to repeat. The first is query sensitivity. A French phrase with the town and category may retrieve the owned site. A slightly broader phrase may retrieve a directory because the directory is stronger for category matching. The lab does not treat this as noise to be averaged away. The change shows how narrow the route may be.
The second reason is competing structure. Directories and review platforms often package business identity into consistent fields. Owned sites, especially small French SMB sites, may spread identity across pages: logo in an image, phone number in a footer, location in a contact page, services in paragraphs and seasonal details in a banner. A system can select the site once and still prefer the structured source later.
The third reason is language routing. Noémie Arcas’s part of the lab often sits near this problem. A French query, English query and mixed-language query can retrieve different trails for the same business. If the first citation happened under a French frame, it does not guarantee the same source under an English frame. In some cases, the English frame reaches a bilingual directory before it reaches the French owned page.
The fourth reason is hidden system variation. Interfaces may blend live retrieval with stored knowledge. They may show only some sources. Ranking may shift without a visible explanation. The lab can record the trace; it cannot open the machine and inspect every decision.
That uncertainty is not an excuse to stop measuring carefully. It is a reason to treat one citation gently. A single retrieval event is a useful observation. It becomes evidence of repeatability only after comparable runs show the same pattern holding.
Reading weak and strong repeatability
Indexe Clair uses plain language for the patterns rather than a numerical score. Weak repeatability appears when the business is named again but the selected source changes or disappears. Medium repeatability appears when the same source returns in some comparable runs but loses under nearby query frames. Stronger repeatability appears when the same owned page or source trail is selected across stable language, location and category wording. These are qualitative readings, not a scale with hidden numbers.
The weaker pattern still deserves respect. A French SMB that is repeatedly named but sourced through mixed trails is not absent. It may have entity visibility without source control. The stronger pattern is different: the owned page becomes part of the recurring evidence path. That is closer to what many business owners imagine when they say they want to be “found.”
A small imperfection often helps classify the pattern. Suppose the system cites the owned site three times but once repeats an old founding year from a directory. The source selection looks stable, while one detail may still leak from another trail. The lab would avoid calling that clean success. It would say the owned site is recurring as visible evidence, with a remaining source conflict in synthesis or auxiliary retrieval.
This is why the answer itself cannot carry the whole interpretation. Fluent prose can conceal source instability. A source list can show instability even when the paragraph reads smoothly. Indexe Clair keeps those layers apart because the business question is not only “what did the AI say?” but “what did it reach for when it said it?”
Limits of repeat-run evidence
Repeat runs do not produce a permanent guarantee. They produce a better description of a retrieval pattern under recorded conditions. The lab cannot prove that a source will appear for every user, every location, every personalization state or every interface version. It also cannot assume that a source hidden from the visible list was absent internally. Some systems expose source trails more plainly than others.
The method has another limit: comparable does not mean identical. AI search is not a locked laboratory instrument. The same wording may still meet a changed index, altered ranking or different live retrieval state. Indexe Clair treats that variation as part of the field, but it does not pretend to control every variable. The result is a disciplined reading, not a mechanical proof.
The team is also careful with forecasts. If current interface patterns persist, one-time citation will probably remain a fragile signal for French SMBs with scattered public evidence. But that statement is an interpretation of observed behavior, not a rule. The safer finding is narrower and more durable: a citation becomes meaningful evidence of visibility only when the same source trail returns under comparable query frames.
The last note is almost uncomfortable because it removes the small celebration too quickly. Being selected once is still something. It means the page, listing or entity reached the surface. But the lab does not let that moment harden into a claim it cannot support. A footprint is evidence that someone crossed the mud. It is not yet a road.