A French business can update the public web and still meet an older version of itself inside AI search. The question is not whether the answer sounds current, but which trail the system actually reached.
In a composite scenario, a bakery equipment supplier outside Tours changes its opening hours after moving the collection desk to a side entrance. The owned website is updated first. The product pages still load, the contact page now shows the new hours, and a short local mention from a trade association repeats the corrected schedule. A controlled French query asks for the supplier’s opening times and nearby pickup details. One AI search system returns the new hours. Another gives the old hours from a directory listing, with a neat sentence that sounds fresher than the evidence beneath it.
That is the kind of small mismatch Indexe Clair keeps returning to. In the lab’s composite Tours supplier scenario, the problem is not dramatic hallucination. The business is recognized. The place is approximately right. The category is understood. Yet the source trail bends backward. A stale listing is easier for the system to select than the corrected owned page, or the answer blends two records into a tidy paragraph. The update exists in public. The retrievable version has not fully caught up.
The local answer can look live while the trail is older
AI search interfaces often give a live feeling. They write in the present tense, mention current-seeming details, and sometimes expose sources beside the answer. For a business owner reading quickly, this can make the answer feel newly checked. Indexe Clair is careful with that impression. A fluent present-tense answer is not evidence that the system just crawled the current business page.
Live retrieval — in Indexe Clair’s working use — is a visible retrieval event that reflects recently available public evidence, because the selected source trail points to a current page, listing or location signal. The definition matters because “live” cannot be inferred from tone. It has to be read in the trail: which page appeared, what date or detail it carried, whether the owned source or a refreshed third-party record was selected, and whether the same query repeats the behavior.
A cached-looking answer can be harder to identify than a wrong answer. If a system names a closed restaurant, the mismatch is loud. If it names an active supplier but gives the previous opening hours, the issue hides inside a mostly useful response. In several lab-style runs around French SMB categories, the stale part is often small: one old phone number, an older category label, a previous address line, a directory title that still uses the business name before a minor rebrand. The answer feels current because most of it is plausible.
This is why the lab treats the page, listing, business name, geographic signal or source trail as the observation unit. A sentence alone is too smooth. It may be synthesized from current retrieval, older stored knowledge, a search-result snippet, a directory cache, or a blend that the interface does not fully expose. The visible retrieval event is rougher and more useful. It leaves a handle.
What changed public evidence does to the run
In the composite Tours supplier case, the team uses a simple contrast. First, they describe the old trail: a directory entry with outdated hours, an owned website with older but consistent product pages, and a few local mentions. Then they describe the changed evidence: the contact page is corrected, the local mention updates, and the directory remains stale. The query frame is held still as much as the interface allows: same business category, same location framing, same language, same intent.
The revealing result is rarely a clean switch. One system may select the updated owned page for a direct brand query, then return the stale directory for a category query like “fournisseur matériel boulangerie près de Tours.” Another may cite the owned site but preserve the old hours in the answer. A third may show no sources and give a sentence that cannot be traced with confidence. These are not three levels of accuracy on a neat ladder. They are three different retrieval situations.
The lab reads them as competing evidence trails. The owned page may have passed discovery and indexing, while the directory still wins source selection. Or the business entity may be indexed through the directory, while the page update has not entered ranking for the relevant query. Sometimes the current page appears only when the query includes the exact business name. Remove the name, keep the category and town, and the system falls back to a more established listing.
A page can be current, crawlable and still lose the moment of source selection to an older record with stronger retrieval gravity.
That phrase, retrieval gravity, is not a metric. It is the lab’s shorthand for the pull some sources seem to have inside observable answer trails. A directory page may be structured, linked, old enough to be trusted, and easy to parse. The owned page may be more accurate but thinner, deeper in the site, or less clearly tied to a category phrase. In a local query, the system may prefer the record that looks like a business object over the page that looks like a small commercial website.
The four gates help separate delay from preference
The lab’s anchor classification is useful here because “live or cached” is too blunt. Indexe Clair reads the four retrieval gates a French business must pass — discovered page, indexed entity, ranked evidence, selected source. Each gate describes a different failure shape.
A discovered page means a crawler or retrieval layer can find the page. That alone does not prove the system treats the business as a stable entity. An indexed entity means the business can be recognized as a retrievable object, perhaps through an owned site, a directory, a review profile, or some mixture of public traces. Ranked evidence means the relevant page or listing rises for the query frame being tested. Selected source is the final visible choice: the source the system uses or exposes in the answer.
The stale-hours problem can occur at any gate after discovery. A corrected contact page may be discoverable but not yet ranked for the local category query. A business may be indexed through a directory rather than through its own site, so the stale directory becomes the system’s entity spine. A current page may rank in a general web search surface but fail to become the selected source inside the AI search answer. The outcome looks like “cached knowledge,” yet the mechanism may be source preference rather than delay.
This distinction changes how the lab reads evidence. If the updated page never appears in any query frame, even a direct brand query, the question leans toward discovery or indexing. If the page appears for exact-name prompts but disappears for category prompts, the issue is ranking under broader intent. If the page appears as a source while the answer still uses old details, the synthesis layer may be blending evidence, or the source snippet may not contain the relevant updated detail. The laboratory does not collapse these into one label too early.
In the Lyon peri-urban repair service composite, this becomes visible through geography. A small repair firm outside Lyon updates its service area page to clarify several communes. Direct French queries retrieve the owned service page. A “near Lyon” frame retrieves a larger chain, then a review profile, and only later the independent firm. If one of the third-party profiles has stale coverage details, the answer can look cached. Yet the source trail suggests a different story: the system preserved the larger-city retrieval path and selected sources that already had stronger geographic signals.
How the lab tests without pretending to see inside the system
Indexe Clair cannot inspect the internal index of ChatGPT Search, Perplexity, Copilot or Google AI Overviews. The lab works from visible traces. That makes the method modest, but also disciplined. The team records the query frame, language, location wording, system conditions as exposed by the interface, source trails when available, and any conflict between selected sources and current public evidence.
The work depends on repetition, not on a single striking screenshot. A query is rerun with the same wording, then with controlled variants. Exact business name, category plus town, category plus “près de,” French wording, English wording, mixed-language wording. The point is to see whether the updated evidence appears only under one frame or whether it becomes stable across the retrieval path. Variation is expected. The interesting part is whether the source selected changes.
For the “live or cached” question, the team pays particular attention to details that have a before-and-after form: hours, address, page title, service area, product availability phrasing, business category, or a newly added local mention. These are easier to trace than vague brand reputation. A changed sentence on an “à propos” page is often too soft; a changed opening-hour block or corrected location line leaves a sharper mark.
The method also avoids turning every stale answer into proof of caching. A system might retrieve a current page but use a snippet that was generated before the page changed. It might retrieve a stale directory because that directory ranks more clearly for the category query. It might have no live source exposure, making the answer impossible to classify beyond visible behavior. The lab’s finding can therefore be narrow: “this run selected the stale directory after the owned page had been updated,” rather than “the system is cached.”
That restraint may feel unsatisfying. It is also the difference between research and irritation.
What this means for French SMB evidence
For a French SMB, the practical lesson is that updating the owned website is necessary but does not close the retrieval loop. The updated page still has to be found, associated with the entity, ranked for the query frame, and selected over other public traces. A correction that appears on one page may remain weak if the rest of the source trail keeps telling the older story.
The lab’s observations suggest that freshness signals work best when they are not isolated. A changed opening-hours page may become more retrievable when internal links point to it clearly, the location language is unambiguous, the business name and category are consistent, and stale third-party records do not keep presenting a conflicting version. This is an interpretation, not a rule. The observed mechanism is that AI search systems often use several public traces, and a single corrected trace can be outvoted or bypassed.
This is especially sharp in French local search because evidence is scattered across owned sites, directories, review pages, municipal mentions, sector listings and regional references. The system may choose the source that is easiest to parse, not the source closest to the business. A small firm can become present on the public web and still remain unevenly retrievable. The old record sits there like a folded road sign in a garage: no longer useful to the driver, still legible to the machine.
The lab is also cautious about advice that promises speed. Submitting pages, improving crawlable text, clarifying structured business details and correcting directory conflicts may help visibility in observable ways. But “help” is not the same as “guarantee.” The only responsible claim is to test whether the retrieval event changes after the evidence changes.
Limits of the live-versus-cached reading
This material cannot show what an AI search system stores internally, how often it refreshes a given source, or which ranking signals it applies behind the interface. Some systems expose sources clearly. Others give a finished answer with little trace detail. Personalization may be partly hidden. Live retrieval can mix with retained knowledge, and a visible source may not be the only source that shaped a sentence.
The method also depends on public evidence that can be compared. If a business has no clear before-and-after change, the live-versus-cached question becomes muddy. A vague claim about “better visibility” cannot be tested the way an updated address line can. For this reason, Indexe Clair treats the strongest observations as those where a visible retrieval event points to a specific current or stale source.
The answer to the title question is therefore deliberately uneven. French local AI search can behave live for one query frame and cached-looking for another. It can retrieve a current owned page and still select an older listing. It can find the business as an entity while missing the updated evidence that matters to a customer. The lab’s conclusion is less tidy than a platform comparison chart, but closer to the source trail: the real question is not simply whether the system is live or cached. It is which version of the business became retrievable when the system looked.