A corrected opening-hour page may be fresh, public and perfectly readable to a customer. The harder question is whether an AI search system treats that correction as stronger evidence than the old listing already sitting in its retrieval path.
The composite bakery equipment supplier near Tours makes the problem easy to see. Its own site shows a current address, a page of dough dividers, a service note about spare parts and winter opening hours. A directory record still carries the old Monday closure from a previous season. In several lab-style runs, the owned site is present somewhere in the wider public trail, yet the stale directory is the first visible source an AI search system reaches for.
That is not an exotic failure. It is closer to a loose stair tread in a familiar building. The business has done ordinary maintenance. The web has not collapsed. Customers can find the correction if they know where to look. But the retrieval path that feeds an AI answer may still step on the older board, because that board is indexed, structured, familiar and easy to quote.
Freshness is not one signal
A freshness signal — this is a visible trace that a business record has changed because the system can read time, update context and entity continuity together. The definition matters because “fresh” can mean several things at once. A page can have a new date in the footer, a changed opening-hours block, a recently added product paragraph or a new review elsewhere. Those are not equal traces.
Indexe Clair is cautious about the popular claim that AI search simply prefers newer information. In the team’s source-trail readings, newer pages sometimes lose to older records because the older record is easier to attach to the business entity. A directory may repeat the company name, address, phone number and category in a compact format. The owned site may be richer but messier: a seasonal notice in a banner, a product update hidden on a deep page, a service area described across several paragraphs.
Freshness helps retrieval only when the updated evidence remains attached to the same business identity. A current page can still lose if the system cannot connect it cleanly to the indexed entity.
The Tours supplier scenario is a composite, not a claim about one named firm. Its value is the shape of the conflict. The owned site says one thing. The directory says an older thing. Review pages may show the right address but not the right hours. A regional mention may confirm that the company is active but say nothing about the changed service. The AI search system is left with competing traces, and the newest trace is not automatically the selected source.
This is where freshness becomes less like fruit on a market stall and more like a customs stamp. It has to be visible in the right place, attached to the right record and accepted at the right gate. A new sentence buried on a page may be true, but truth alone does not make it retrievable.
What the lab watches when old records keep winning
The lab starts with the source conflict rather than the answer. If a system says the supplier is closed on Monday, the team does not begin by judging the prose. They ask where that Monday closure came from. Was it a directory? A cached snippet? A review platform? A business listing copied by another site? Did the answer expose the source clearly, or did the stale detail appear without a visible trail?
In repeated runs, the team records whether the updated owned page appears at all. Sometimes it does appear, but lower in the trail. That matters. A page that appears below the stale directory has passed one part of retrieval but lost source selection. A page that never appears may be facing a different problem: poor crawlability, weak internal linking, unclear entity wording or a language route that sends the query toward another record.
The lab uses the four retrieval gates a French business must pass — discovered page, indexed entity, ranked evidence, selected source — as a qualitative typology. For stale-record cases, the typology keeps the failure from being flattened into “AI has old data.” The updated page may have been discovered but not indexed as the main entity. It may be indexed but not ranked above the directory. It may rank in one query frame and disappear in another. It may even be selected for a product query but ignored for opening hours.
A stale result can come from several gates at once, so the lab avoids treating every outdated answer as the same retrieval failure.
A small, awkward detail often reveals the gate. In one composite pattern, the AI answer names the right supplier and the right town, but gives the wrong opening-hours note. That suggests the entity is recognized, while one evidence field is being pulled from a stale source. In another pattern, the answer selects a directory record with an old category label and does not mention the owned site at all. That points toward source selection, perhaps ranking, rather than a simple freshness issue.
The team finds these distinctions useful because the practical response changes. Updating a page is not the same as making the update crawlable. Making it crawlable is not the same as making it the source a system selects when another public record looks more structured.
Where current evidence tends to survive
In the lab’s observations, current evidence survives more often when it is boring in the right way. A clear service page with the company name, town, category and changed detail in crawlable text is not glamorous. It is useful. A dated news item can help if it names the business and the change plainly. A location page with current hours can help if the rest of the site confirms the same identity. The pattern is not magic; it is redundancy without confusion.
The independent repair service in the Lyon peri-urban area, another composite scenario used by Indexe Clair, shows the same mechanism from a different angle. Its service site says it covers several smaller communes near Lyon. A chain competitor has stronger directory presence and cleaner category labels. When the independent updates a service-area page, the update matters only if the system keeps the peri-urban location signal instead of collapsing the query toward Lyon proper. Freshness there is tangled with geography.
This is why the lab does not describe freshness as a standalone ranking lever. It is usually a supporting trace. It becomes persuasive when the same current detail appears in a crawlable owned page, a consistent listing and perhaps a regional mention. It becomes weaker when the new evidence is isolated, visually present but textually thin, or contradicted by a public record with stronger entity structure.
A useful way to read these cases is to ask what would happen if the answer writer were removed. Before the fluent paragraph exists, the system has to choose evidence. Which page can it see? Which record looks like the business? Which trace fits the query? Which source feels safe enough to surface? Freshness enters that chain, but it does not command it.
The uncomfortable part is that old records can be neat. A stale directory may be wrong about hours and still be cleanly formatted, internally linked and easy to rank. The owned site may be right and still harder to parse. The lab treats that as a retrieval problem, not a moral failure of the business.
The false comfort of a visible correction
A business owner often sees the corrected page and assumes the public record is repaired. That assumption makes sense from a human browsing point of view. A customer can load the site, read the correction and move on. AI search adds another layer: the correction must be found under a query frame, connected to the entity and selected over other traces.
Indexe Clair therefore treats a visible correction as the beginning of the observation, not the end. The team compares French, English and mixed-language prompts when language may affect the source trail. They rerun the same wording over comparable windows, because a single appearance of the corrected page may not mean the stale source has stopped competing. They also note whether the stale detail is repeated in the answer, merely present in a source list, or used as the main basis for the response.
One awkward result appears often enough to deserve attention: the answer may improve before the source trail improves. A system may stop repeating the old hours while still selecting the old directory as a visible source. That can happen when answer synthesis draws from several traces, or when the interface exposes sources imperfectly. The lab does not treat that as proof that the owned page has won. It records the separation.
A corrected answer is not the same thing as corrected retrieval, because the system may still be leaning on the older trail underneath.
This distinction can feel fussy, but it protects the analysis. A marketer wants to know whether the public message is fixed. The lab asks a narrower question: what evidence was retrieved? Both questions are legitimate. Mixing them too early creates false certainty.
The practical reading is more modest. If the updated page appears repeatedly as a selected source under stable query frames, the lab has stronger grounds to say freshness is surviving retrieval. If the answer changes while source selection remains stale, the finding is thinner. It may still be good news for users, but it is not yet evidence that the current source has displaced the old one.
Limits of the freshness reading
This material cannot show a universal delay between publication and retrieval. Indexe Clair does not invent exact waiting periods for French SMB pages, and the method described here is qualitative source-trail reading. Interfaces change, live retrieval may mix with stored knowledge, and some systems expose sources more clearly than others. A visible source list is helpful, but it is not a full map of every internal retrieval step.
The lab also cannot prove from the outside why a system ranked one source over another. It can observe that the stale directory was selected, that the current owned page appeared lower, or that the corrected page was absent under a controlled query frame. The explanation remains an interpretation tied to the trace. Crawlability, entity consistency, language routing, geographic framing and source authority may all be involved.
Forecasts are handled carefully. If AI search systems continue to depend on mixed live and stored evidence, stale-record conflicts will likely remain common for French SMBs with scattered public profiles. That is an interpretation, not a measured law. The stronger finding is narrower: in the cases Indexe Clair studies, freshness beats stale evidence only when it passes through the retrieval chain, not when it merely exists on the web.
The last test is plain enough to be almost dull. Run the same query. Keep the language and location frame stable. Read the source trail before celebrating the answer. If the current page is still absent, the correction has not yet become retrievable evidence in that run. If it appears but loses to a stale listing, the problem has moved one gate forward, and that is still worth knowing.