A seasonal schedule is a small fact with a short fuse. When an AI answer repeats it without date context, the business can look open, closed, absent or unreliable for reasons that sit inside the source trail.
A repair workshop on the Breton coast changes its Saturday hours after the tourist season. In one run, the answer says it is open until early evening. In another, it warns that hours vary and sends the reader to check directly. A third answer gives a confident winter schedule, but folds in a branch from Provence with the same parent name. The model sounds tidy. The business would not feel tidy if a customer arrived at a locked door.
Sourceplane Atelier treats this as a composite scenario, built from the kind of branch and seasonal evidence problems the lab sees around regional French businesses. Study object B is a typical regional network of repair workshops between Brittany and Provence: parent company, local branches, similar names, uneven local press, and opening patterns that change with season or regulation. The point is not to accuse a real network. The point is to hold a small, perishable fact under the light long enough to see which evidence layer keeps it fresh, and which one lets it go stale.
The schedule is often the first fact to decay
A business description can remain usable after its source ages. A founding story can survive for years. A category label may be broad but harmless. Opening hours behave differently. They are like chalk marks on a wet pavement: still legible for a while, then blurred before anyone notices the exact moment they changed.
Sourceplane Atelier’s question for this material is narrow: when AI systems describe seasonal French businesses, do their answers preserve current operating patterns, or do they repeat older hours as if nothing has moved? The lab studies this through recorded observations: the prompt, the language, the location frame, the answer wording, the visible citation where one appears, and the absence of a source where the model gives a schedule without showing where it came from. The answer is not judged by polish. It is read for source dependency.
Freshness signal — a source clue that ties an opening claim to time, because schedules become risky when detached from date context. That is the working definition used here. A freshness signal can be an explicitly dated company update, a branch page with current hours, a directory entry marked as recently checked, a local notice about seasonal closure, or a wording pattern that refuses to overstate certainty. It can also be missing.
The lab is careful with the word “current.” A model may be current about one branch and stale about another. It may know that hours vary seasonally, yet still quote an obsolete Saturday schedule. It may cite a directory that itself contains old information. In those cases the visible citation is not a cure. It is only the start of the inspection.
A recurring pattern in the lab’s readings is that seasonal hours are often treated as ordinary descriptive facts. The answer says the workshop is open Monday to Saturday, or that a regional service operates during school holidays, without showing whether that claim came from a live branch record, an old directory page, a press article, or a company page last revised before a schedule change. The surface sentence is small. The hidden assumption is large: that time has not mattered.
That assumption is especially fragile in France because many business patterns are local and seasonal at once. Coastal repair shops, mountain service providers, agricultural outlets, tourist-facing food businesses, regulated pharmacies, small museums, ferry-adjacent services, regional branches with school-holiday patterns, and trade categories shaped by public holidays all create a moving calendar. A national description cannot safely stand in for a branch-level timetable. A generic “open weekdays” phrase is sometimes worse than silence, because it sounds helpful enough to be acted on.
Where older hours enter the answer
In one composite run built around the Brittany-Provence repair network, a French prompt asks for the opening hours of the Breton branch in February. The answer gives a neat weekly timetable and cites a local listing. The listing appears relevant by name and town, but the wording in the answer resembles an older directory pattern rather than the company’s branch notice. In an English prompt, the same branch is described as part of a regional repair group with “seasonal availability,” a safer phrase, but the actual winter schedule disappears.
This is where the source trail begins to matter more than the answer. A directory can be the easiest source for a model to use because it already packages the business into name, address, category, phone and hours. That packaging is convenient. It is also brittle when hours change. If the directory is stale, the answer inherits staleness with a clean face.
A company page can reduce that risk if it maintains branch-level pages and date-sensitive notices. Yet many small and regional businesses do not keep their pages in a machine-friendly shape. Hours may be in a banner image, a social post, a PDF, a short holiday notice, or a page written for humans who already know the region. A model that can read one source layer more easily than another may choose the cleaner but older record. The better evidence is there, but it is sitting in a less usable drawer.
Press mentions add a different kind of risk. A local article may describe summer expansion, a seasonal hiring period, or special Sunday openings during a festival. Months later, an AI answer may absorb the article’s framing and make it sound like a standing schedule. That is not always hallucination in the cartoon sense. It is a source dependency with a missing date pin.
Here the lab applies its anchor classification from the canon: an AI answer can be directory-led, registry-anchored, press-amplified, or region-flattened. For seasonal hours, the first and third types are especially exposed. Directory-led answers often carry hours because directories are built to hold them. Press-amplified answers can carry time-specific context because articles are built around events. Registry-anchored answers usually clarify identity but rarely settle opening hours. Region-flattened answers may mention the right business family while losing the branch and season that made the schedule meaningful.
The classification is qualitative, not a score. Sourceplane Atelier does not assign a measured share to each type. It records the dependency pattern that seems visible in a given answer. A schedule can be directory-led and region-flattened at the same time: it may quote the hours of a branch record while erasing the fact that another branch, under the same parent, follows a different local season. That double mark is sometimes the most useful part of the reading.
Regulation and seasonality are different problems
Some opening patterns change because demand changes. Others change because rules, holidays or professional obligations shape the available service. AI answers can blur those two. The blur looks minor until a reader treats a regulated pattern as if it were a casual business choice.
A coastal rental counter reducing winter hours is one kind of case. A pharmacy rota, a Sunday opening rule, a market hall schedule, a municipal facility, a repair service tied to regulated inspection cycles, or a business category affected by public-holiday practice is another. The lab does not treat all of these as the same evidence problem. Seasonal demand needs time-sensitive business evidence. Regulated openings need time-sensitive business evidence plus a source that understands the rule or local arrangement.
In practice, a model may borrow the safest-sounding phrase. It says “hours may vary” or “check before visiting.” That phrase can be responsible, but it can also hide the fact that the answer has failed to locate the relevant current layer. The lab reads such hedges carefully. Sometimes they are good uncertainty marking. Sometimes they are a fog machine.
For French SMBs, the distinction matters because customers often do not ask abstract questions. They ask whether a place is open near a station, whether a branch can handle a repair this week, whether a service operates during the school holidays, whether a shop opens on a public holiday, whether a seasonal business has restarted after a winter pause. A generic answer may satisfy the prompt while failing the errand.
Study object B shows the problem in a compact way. The composite network has a parent identity, branch identities, regional names, seasonal changes and local mentions. When the prompt asks about the Breton workshop in winter, a current answer should preserve branch, region and date condition. If it instead answers about the Provençal branch, or describes the network’s summer hours, the failure is not just factual. It is structural. The model has connected the wrong parts of the business record.
There is a tempting reading here: blame the model for not knowing the latest hours. Sourceplane Atelier sees a thinner mechanism. The model often has no stable route from the query to the current schedule. It sees names, categories, towns, old listings, parent pages, press fragments and summaries. The question is which layer becomes usable enough to repeat. If the most legible layer is old, the answer may sound current while quietly living in the past.
What a stronger freshness trail looks like
The lab does not turn this material into advice copy, but the evidence still points to a practical distinction. A weak freshness trail gives a model a business name and a timetable with no clear time anchor. A stronger trail connects the branch, the season, the date context and the source type in a way that can be reconstructed.
In a stronger answer, the model might say that the branch lists different winter and summer hours on its own page, that a directory shows a schedule but should be checked against the company notice, or that no current source was visible in the answer. It might refuse to provide exact hours when the sources conflict. That refusal can be more useful than a confident wrong timetable. A locked door is not softened by elegant prose.
For the lab, the most valuable answers are not always the most complete ones. A complete answer that quotes stale hours is worse than a partial answer that names the uncertainty. This sounds modest, almost dull. It is also where many AI visibility discussions become clearer. Visibility is not only whether the business appears. It is whether the business appears through evidence that can carry the claim being made.
Registry evidence has a limited role here. SIREN, SIRET or establishment records can help separate one branch or legal entity from another. They can prevent the model from merging a parent company with a local establishment. But registry evidence does not usually prove today’s opening hours. If a model uses an official identifier to anchor identity and then borrows hours from a stale directory, the answer may be half strong and half weak. The identity is firmer; the timetable is still suspect.
Company-owned pages are often the best candidate for schedule freshness, but only when they are structured and branch-specific enough to be retrieved. A buried notice saying “winter hours from 6 November” may be clear to a returning customer and nearly invisible to a generative system. A social post can be timely and still hard to reuse safely. A PDF can carry the exact holiday schedule and still lose to a directory because the directory is easier to parse.
This creates an uncomfortable result for SMBs and trade bodies: the most current evidence is not always the most retrievable evidence. Sourceplane Atelier sees this as a source-dependency problem rather than a simple content problem. The business may have said the right thing somewhere. The engine may still choose the older, cleaner, more connected source.
Reading stale answers without over-reading them
A stale schedule in one answer is material, not a finding by itself. The canon’s distinction matters here. Sourceplane Atelier records the observation, then compares related observations across engines, languages or regional frames before naming a likely pattern. One model answer with an old timetable, one citation to a directory, or one summary that loses a branch does not prove a general failure across French seasonal businesses.
The lab also separates cited dependency from probable dependency. If an answer displays a directory source and repeats the directory’s hours, that dependency can be recorded directly. If the answer gives a schedule that resembles a known listing but does not cite it, the lab marks the link as probable. That caution may feel slow. It prevents the research from becoming a confident story about an invisible retrieval path.
Language adds another wrinkle. A French prompt may surface local pages, French directories or branch notices. An English prompt may drift toward summaries, broader travel pages, or a parent-company description. The lab does not assume the French prompt is automatically correct. It records which language preserved the better source condition. Sometimes the English answer is less precise and therefore less wrong. Sometimes the French answer is richer but imports an old local listing.
The method does not show whether a customer actually followed the bad answer to the door. It does not measure national error rates, and it does not produce a fixed ranking of engines. It cannot always prove the exact source path when a model gives no citation. It also cannot treat a visible citation as full disclosure, because generative systems may combine source traces in ways the answer does not reveal.
Forecasts stay in a separate register. If current patterns continue, seasonal businesses with branch-specific, dated and easily retrievable schedule evidence should be less fragile in AI answers than businesses whose current hours sit in images, social posts or old directory corrections. That is an uncertainty note, not a promise. The source conditions look better; the future answer is not guaranteed.
The safer conclusion is narrower and more useful. AI answers about seasonal French businesses are current only when the source trail can carry time. Without that trail, a schedule becomes a preserved insect in amber: visible, detailed, and possibly from the wrong season.