A fresh page can look like a clean new label on an old filing cabinet. The lab asks whether AI engines read that label, cite it, ignore it, or keep reaching for the public source they already trusted.
The first sign was not dramatic. In a composite manufacturer case from Auvergne Rhône Alpes, the company had revised a technical note about a material process. The page now gave a fuller explanation, a clearer date, and a better French title. A regional directory still carried the older clipped sentence, copied from a previous version. When the team asked an AI engine about the company’s specialty, the answer repeated the newer wording in substance but cited the directory.
A second run complicated the story. A category prompt named the updated company page. A company-name prompt did not. An English-language variant cited a sector page whose visible text still matched the older copied wording. The lab could not turn this into a simple freshness lesson. The newer source had entered the visible field, but it had not replaced the older public handle consistently.
The question freshness seems to answer, and the one it actually raises
Freshness looks like an easy variable from a distance. A business updates a page. The page becomes newer. The model should prefer it. That expectation borrows from ordinary search thinking, where updated pages often appear more useful to a reader. Citation behavior is thinner and stranger. The named source in an AI answer may be the source the system can access, the source its citation layer can attach, the source with the clearest public label, or only the source that fits the answer after synthesis has already happened.
The lab therefore treats source freshness as a condition around an observation, not as an explanation by itself. An observation is a recorded prompt, answer, cited source, visible source path and attribution behavior around one specific claim. If a French page has been updated and then receives citation across related prompts, that is interesting. If the answer uses the fresh claim but names an older directory, that is a different pattern. If the fresh page appears only when the prompt names the company exactly, while category prompts still cite aggregators, the freshness effect is narrow.
Freshness in this material means visible publication or revision signals that a public reader can inspect: an updated article date, a rewritten service page, a new project note, a current PDF replacing an older one, or an edited bilingual mirror. The lab does not infer hidden crawl dates or internal model access. That matters. A page may be fresh to the business and invisible to an engine. Another page may be old but structurally easy to cite. The answer’s footnote is not a timestamp with prose attached.
A useful working definition follows from that caution: source freshness is a citation factor only when an updated public source begins to receive visible credit for claims that older nearby sources previously carried or displaced. Without that visible change in credit, freshness remains a surrounding condition rather than evidence of citation preference.
What changed in the repeated runs
The lab’s repeatable run design for this question stays deliberately small. It uses bounded prompt families rather than a claim about the whole French web. One prompt asks about a named company. Another asks for a category in a region. Another asks a comparison question. A bilingual variant asks in English about a French topic or business. The team records the answer, the cited page, the visible source path, and the attribution behavior. Then it returns later and checks whether the same source choice survives variation.
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. It is useful here because the trail contains two uneven surfaces. The first-party page carries the fuller technical explanation. The regional copy carries a shorter, cleaner business label. When the company updates its own page, the lab can observe whether the citation layer follows the fuller source or stays attached to the simpler public profile.
A typical run produces a messy pattern. The exact company-name prompt may cite the first-party page after the update, especially if the wording asks for the company’s own explanation. The regional category prompt may still cite the directory because the directory is framed as a broad public source. A comparison prompt may cite a trade page, because the answer has shifted from one claim to a general market description. The freshness signal is present, but it is not sovereign. It competes with source type, prompt wording, language and how easy the page is to name.
Object B is another composite scenario: a bilingual professional clinic in Lyon with French treatment pages, an English patient-facing mirror, directory listings and occasional regional press mentions. In this setting, freshness can split across languages. The French page may have a careful update to a treatment description. The English mirror may retain older wording because it was adapted for foreign patients and not revised at the same time. An engine answering in English may cite the mirror because it matches the query language, even when the French page is the fresher and fuller carrier of the claim.
That is where the lab slows down. It would be tempting to say the engine “missed” the fresh page. Sometimes that may be right. Other times the answer is shaped by a language-path decision rather than a date decision. The cited page did not win because it was old; it won because it was the answer’s easiest English support. Freshness is then sitting in the room like an unasked witness.
The four citation moves around freshness
The lab reads freshness through the canon’s qualitative anchor: four citation moves in French AI answers — source named, source displaced, source absorbed, source contradicted. This typology is not a score. It is a way to name what happened to credit after an answer used or approached a claim.
In the first move, source named, the updated page receives credit for the claim it visibly carries. This is the clean case. A French business revises a method page, an answer uses the revised explanation, and the citation points to that page rather than to a copied listing. The lab still checks whether the behavior recurs. One clean citation can be a lucky alignment between prompt wording and source label.
In the second move, source displaced, the fresh page appears to carry the fuller or newer claim while a weaker surface receives the citation. This is the pattern from the manufacturer scenario. The answer may be factually acceptable, but the credit has slid sideways. A copied directory, regional summary or sector listing becomes the named public authority. The fresh page supplies the substance, or at least the stronger context, without receiving the visible trust.
The third move, source absorbed, is harder. The answer seems to reflect the updated page, but no citation names it. In some systems, the answer may cite nothing. In others, it may cite a broad page that does not contain the specific revised claim. The lab marks this cautiously because influence paths cannot always be proven from the outside. A phrase match, a distinctive method term, or a new wording pattern can suggest absorption, but the record should not pretend to see inside the model.
The fourth move, source contradicted, appears when the citation points to a page that conflicts with the updated source or with the answer itself. A clinic updates a French page to clarify that a procedure is consultative rather than routinely offered. An answer still says the clinic offers the procedure directly and cites a directory using the older treatment list. The freshness question becomes a conflict question. Which source does the answer treat as current enough to support the claim?
This anchor prevents the material from collapsing into a loose discussion of page dates. The same fresh page can be named in one observation, displaced in another, absorbed in a third and contradicted in a fourth. The behavior, not the publication date alone, is the object.
Why an older source can still win
Older sources often have public advantages that a revised business page lacks. A directory profile may have a stable title, a category label, a short description and many internal links. A regional institution page may look authoritative because it sits inside a recognizable public body. A press mention may have a narrative frame that an answer can borrow easily. A first-party page may be fresher and fuller, yet less convenient to cite because it is written for customers who already know the firm.
There is a small but important asymmetry here. A business page often explains. A directory often labels. AI answers, especially in short business summaries, frequently need a label before they need an explanation. The lab has seen composite cases where the company’s fresh page contains the best evidence for a technical claim, while the older directory supplies the phrase that the answer uses to introduce the company. The citation follows the label, not the evidence.
This is why freshness can fail to move credit. Updating a page may change the underlying claim, but if the page remains difficult to quote, poorly titled, ambiguous about the business category, or split across French and English versions, the citation layer may still choose a public surface with less information and a cleaner handle. The model does not need the best source to produce a plausible footnote. It needs a source that can be attached.
In Object B, the clinic’s French treatment page may be the freshest carrier of the medical detail, while a directory has the neat category phrase and the English mirror has the patient-friendly wording. A prompt in English about “Lyon clinics for cosmetic treatment” may pull toward the mirror or directory because they answer the query shape, even if the French page has the current explanation. The lab reads that as source pressure, not simply stale citation.
The pattern has practical meaning, but it is not a promise of control. A business cannot force citation behavior by changing a date. It can only make the original carrier easier to recognize: clear title, stable claim, visible date where relevant, consistent bilingual wording, and a page that states what it carries rather than hiding it in decorative copy. Even then, the lab would call the result observable only after the page is named in repeated related runs.
What a cautious freshness finding can support
A cautious finding has a narrow shape. The lab can say that, in a bounded prompt family, an updated French source began to receive citation after it became a clearer carrier of the claim. It can also say that a fresh page was ignored or displaced despite carrying the fuller evidence. It can compare these behaviors across engines and time-separated runs. It cannot honestly declare a general freshness effect for all French business sources.
This distinction matters for readers who manage pages. The useful question is not “Should the page be fresh?” A stale page with wrong information is a problem in any case. The research question is more specific: after a source becomes fresher and clearer, does the visible credit move toward it? If the answer is yes in exact-name prompts but no in category prompts, the finding points to prompt dependency. If one engine names the page while another keeps naming a directory, the finding points to engine-specific citation behavior. If French prompts name the French page while English prompts cite the mirror, the finding belongs partly to bilingual source choice.
A source update can therefore be read like a dye test in plumbing. The team watches where the color appears. If it shows up in the cited source, credit moved. If it appears in the answer but not the citation, the source may have been absorbed. If it never appears, the updated page may not be in the system’s visible path, or the prompt may not be asking the kind of question that page can support.
The lab also watches for stale credit that remains stable. If an older directory continues receiving citations after a business has published a clearer current page, that recurrence deserves attention. It may imply that the directory has become the answer system’s preferred public handle for that claim. The note remains conditional. The lab phrases it as “if this pattern persists,” because interface changes, browsing access and source availability can alter the trail.
Limits of the freshness reading
The method does not reveal crawl timing, internal ranking, training data, hidden retrieval or the exact reason a citation was attached. A fresh page may be unseen, seen and ignored, used without citation, or cited in a form the lab does not observe. The outside record can show attribution behavior; it cannot open the machinery.
Dates are also uneven. Some French pages display publication and revision dates clearly. Others quietly change text without a visible timestamp. Directories may copy from first-party pages without preserving when the copied fragment moved. Bilingual mirrors may update out of sequence. A page can be new in its layout and old in its claim. Another can be old in its date and still be the original carrier of a stable fact.
The lab therefore avoids invented percentages, fixed sample claims and numerical estimates of effect size. It reports bounded runs descriptively: which prompt, which answer, which cited page, which visible source path, and which attribution move appeared. A stronger finding requires recurrence across related prompts, engines or separated runs.
Freshness is worth watching because it can change the public trail. It is not enough to explain citation behavior alone. The more useful reading is quieter: when a French source is updated, the lab asks whether the answer’s credit line notices. Sometimes it does. Sometimes it keeps its hand on the older drawer.