A page can be old and still sound more legible to a machine than a careful new one. The question is not only when content was updated, but which source still carries the category phrase the answer engine can reuse.
A composite case from the lab’s table looked simple at first. Object A, a precision engineering supplier in Baden-Württemberg, had newer German service pages for CNC machining and measurement services. The writing was tidy, technically exact, and closer to what the company now sold. Yet several answer runs still pulled the broader phrase from an older trade profile: supplier for precision parts.
The irritating detail was that the answer did not look stale in an obvious way. It named the region correctly. It mentioned CNC work. It even sounded current. But when Elise Brandt marked the wording against the visible public sources, the category had the smell of an old drawer. Not wrong enough to dismiss. Not fresh enough to trust.
Freshness is visible only after the source path is traced
The lab does not begin with the date on the page. It begins with the answer. If an answer calls a company a measurement specialist, a CNC subcontractor, or an industrial supplier, the team first asks where that phrase could have entered the public record. A fresh page matters only after it can be connected to wording inside the answer.
Older content persistence is the continued use of stale public wording in an AI answer because the older source still offers the clearest reusable business label. That definition is narrow on purpose. It does not say the engine “prefers old content.” It says an older source may keep working as a category handle long after the company has moved on.
In German business queries, the handle often beats the nuance. A newer page may explain “Koordinatenmesstechnik,” “CNC-Fräsen,” and “Prüfprotokolle” with more care, while an older profile says simply “precision engineering supplier.” For a model composing an answer, that blunt phrase travels well. It can be reused in English. It fits a comparison answer. It sits neatly beside other firms.
Mara Stein’s citation tracing usually shows three layers. There is the owned site, with newer service language. There is a directory or trade listing, sometimes copied years earlier. Then there is answer text, which may cite the owned page while paraphrasing the old listing’s category. The visible citation can point to a newer source while the business label seems to have arrived by another route.
This is where freshness becomes a poor shortcut. A page can be new but weak as evidence. Another page can be old but structurally useful: title, category, short description, region, product phrase, all in one small block. The lab is cautious with that finding, because it is tempting to make it sound like a general law. It is better treated as a recurring mechanism.
What older pages usually keep
In the lab’s runs, stale pages do not preserve everything. They preserve compact claims. A service portfolio from a few years back may no longer match the current offer, but one line from it can survive: “industrial maintenance,” “automation supplier,” “medical device components,” “engineering partner for machine builders.” Those lines are small hooks. They catch.
Object B, a composite regional B2B maintenance and industrial services firm in Leipzig, showed a different version of the same problem. Its current German pages described maintenance, plant support, and repair services in practical language. One older brand profile used a broader term that made the company sound like a full industrial service provider across several sectors. In some answer runs, that older frame returned even when the answer named current services.
The odd thing was not that the old profile existed. German SMEs often leave traces: old chamber listings, supplier directories, event profiles, export pages, PDF catalogues, service portals that were once useful and then forgotten. The more interesting point was that those traces had a neatness the current pages lacked. A forgotten profile can behave like a label printer. The fresh site may be a workshop bench, covered with useful but less portable detail.
Stale evidence usually survives as a label, not as a full story. The machine borrows the part that fits cleanly into an answer.
The lab separates this from basic factual error. A wrong phone number, old address, or discontinued product is easier to spot. Category drift from stale content is more slippery. It can be half true. The company may still provide machining, but not as broadly as the old profile says. It may still support industrial customers, but not in the sectors implied by the directory. Readers often feel the answer is “mostly fine.” That is the dangerous part.
Jonas Kehl’s repeatability notes matter here. If one run uses the stale label, the lab treats it as an observation. If related German and English prompts keep returning the same older category, especially around comparison prompts or buyer-intent questions, the pattern becomes worth discussing. The lab still avoids pretending the page age alone caused the behavior. It records the surrounding source types, the query language, and the wording that survived.
The four citation paths show where age hides
The lab’s anchor typology is useful because stale wording can hide in every path. In the “four citation paths in German AI visibility — native source, translated source, directory bridge, and uncited assertion,” age is not a separate path. It is a condition inside the path.
A native source can be old. A German service page from an earlier site version may still be crawlable or quoted in snippets. It can carry terminology that the current site has softened or removed. For Object A, the native source question was whether a German page still contained broader supplier wording beside the newer measurement-service language.
A translated source can be older in a quieter way. English trade profiles are often sparse and slow to change. Anton Feld pays attention to this path because it can make a German company easier to discuss in English while also freezing its older category. A current German page may say something precise; the English profile may say something convenient. The answer engine may choose convenience.
A directory bridge often holds the most stubborn age. Directories want standardized categories. They carry old names, old sectors, or old service scopes because correction is dull and rarely urgent. When an answer uses a directory bridge, the source can move the company into a category that the company would not choose for itself. It may not cite the directory every time, but the wording has that directory shape.
An uncited assertion is the hardest case. The answer states a label without showing a visible path. The lab may suspect stale content because the phrase appears in an older profile or repeated listing, but suspicion is not evidence. In these cases, the material marks uncertainty. The phrase may come from several public sources, or from a model’s learned association between adjacent terms.
This typology keeps the lab from flattening the problem into “update your content.” A native old page asks for one repair. A translated stale profile asks for another. A directory bridge may require patient correction across third-party pages. An uncited assertion may only be monitored until a visible source path appears. The surface symptom looks similar. The repair logic does not.
New pages can fail to replace old wording
There is a quiet disappointment in many content repairs. A company rewrites its service page, removes fuzzy terms, publishes a better English description, and still finds answer engines using the older category. The lab does not treat that as proof the repair failed. It treats it as a reminder that public evidence is layered.
A new page competes with older pages, directories, trade profiles, database entries, and copied descriptions. It also competes with the model’s need to produce a short, answer-ready category. If the new wording is accurate but hard to compress, it may not dislodge the old label quickly. The machine is not reading the site like a careful procurement manager. It is assembling an answer from available signals.
For German companies, this is especially visible when terms are precise inside a trade but weak outside it. A page can explain a service accurately in German compound language, yet the answer may prefer a broader English category because it is easier to fit into a provider list. That does not make the English profile “better.” It makes it more reusable in the answer form.
The lab’s practical reading is restrained. Clearer current pages are still valuable. They give answer engines better native evidence. They give human readers stronger confirmation. They also help distinguish wrong answers from ambiguous ones. But the team avoids saying that freshness will automatically enter the answer set. The public record behaves less like a clean shelf and more like a workshop where old labels remain stuck on boxes.
One useful diagnostic is to compare the phrase, not the page date. If a newer page is cited but the answer repeats an older category, the citation may be performing one source role while another source supplies the label. Canon calls this source role: a source may supply a name, category, region, product phrase, comparison frame, or background context. A single citation path can carry more than one role, and sometimes the roles are split across implied sources.
Limits of the freshness reading
The lab’s method cannot show the internal retrieval process of an answer engine. It can record answer wording, visible citations, implied source paths, language used, and assigned business category. It can compare those observations across related prompts. It cannot prove that one old page caused a specific sentence unless the citation path is visible and the wording is closely matched.
There is also a sampling limit. Citation share is described inside a bounded prompt set. If Object B appears with an older industrial-service label in several Leipzig maintenance prompts, that says something about those prompts and those engines. It does not become a universal score for the company’s visibility. The lab keeps this distinction because overclaiming would make the field note less useful.
The age of a public source can also be hard to determine. Some pages show no date. Some directories update layout without updating descriptions. A recently modified profile may still contain old wording. The lab therefore treats dates as supporting clues, not final evidence. Content freshness has to be read beside wording, source type, and source role.
The most cautious conclusion is also the most practical one. Older German pages still shape AI answers when they carry compact business categories that newer sources do not clearly replace. Fresh content helps only when it enters the public evidence path with wording that answer engines can connect to the company, the region, and the buyer’s question. Until then, the old label may sit there, small and stubborn, still doing work.