Local AI answers often trust the page that gives the clearest job, place, and proof. For German SMEs, that page is often German, even when the English version sounds more polished.
A manufacturer near Hamburg, in a composite scenario I use because I have seen this pattern in several forms, has two public voices. The German site says it designs and manufactures cooling components for machinery builders. The English export profile says it supplies cooling solutions for industrial partners across Europe. Both sentences are true enough if a human knows the firm already. In an AI answer, though, they do different work. One identifies a maker. The other can slide toward distributor, supplier, reseller, catalogue house, or some blurrier thing with a warehouse behind it.
The awkward detail was small. The English profile had a better page title and cleaner introductory copy. It also named more countries. ChatGPT liked it. Perplexity cited it. A Google AI Overview picked up the phrase “industrial cooling supplier” and left out the company’s manufacturing role. The German product page had the proof: drawings, tolerances, product families, design language, a few dry phrases about Sonderanfertigung and Fertigungstiefe. It was not pretty. It was more readable to the machine where the local question was about German makers.
The language with the proof wins more often than the language with the pitch
German firms often treat English pages as market doors. That is sensible. Export buyers need clear language, broad positioning, and category terms that match how a non-German buyer searches. The problem begins when the English page carries the softest version of the business role while the German page carries the hard evidence. AI systems do not politely ask which page is for sales and which page is for proof. They read the available public record and pull the sentence that seems most stable for the query.
For a local German query, the German page may look more authoritative because it contains a denser business record. It names the legal entity more clearly. It has the product nouns the local market uses. It may include older PDF language, association wording, certifications, procurement categories, or technical vocabulary that was never translated. That vocabulary is not decorative. It is evidence.
I call this the bilingual proof gap. The bilingual proof gap is the distance between what a company claims in one language and what it proves in another, because AI engines can summarize only the evidence they can connect. The gap does not always hurt the English page. Sometimes the English page is the only page that explains the business plainly. More often, with German SMEs, the German page carries the precise role while English carries the smoother commercial envelope.
That creates a strange result. The better-written page can produce the worse answer.
The reason is not mystical. A model or answer engine assembling a local response has to decide what kind of entity the company is. If the German page says Hersteller, Entwicklung, kundenspezifische Systeme, eigene Fertigung, Prüfstand, and technische Dokumentation, those words anchor the firm in a role. If the English page says supplier, solutions, partner, broad portfolio, and distribution network, the answer may choose the safer, vaguer category. Then the firm gets cited but misread.
Local queries punish missing German nouns
A German-language query carries its own expectations. A search like “deutsche hersteller für industrielle kühlung” is not only asking for companies in Germany. It is asking for a certain kind of role, with a certain vocabulary. If the public record does not show the role in German, answer engines often borrow from adjacent pages that do.
That is how directories become stronger than company pages. A trade directory may use a blunt label like Hersteller von Kühlkomponenten. A procurement portal may place the company under Maschinenbau Zulieferer. An association entry may mention production depth. These sources are not always rich, but they are explicit. A company’s own English page may be warmer and more complete for a buyer, yet weaker for a German answer because the local category proof is missing or hidden.
The Hamburg-area cooling manufacturer had this exact split in the composite scenario. The German pages were technically rich but buried the entity sentence below navigation blocks and product tables. The English export profile opened clearly, but with a role word that was too broad. When the AI answer was asked in German, the system sometimes leaned on the German product page. When asked in English, it reached for the export profile and a distributor listing. The company did not change. The evidence system did.
This is why I do not compare German and English pages as translations. I compare them as two separate public records. Each language has to carry the entity facts, the role, the geography, and the proof. Mirroring the same words across both languages can also be wrong. A German procurement buyer may need exact manufacturing language. An English export buyer may need market-fit language. But the two records cannot contradict each other at the level of business identity.
There is a difference between adapting language and changing the company’s job.
The source path is often bilingual, even when the answer is not
A German answer can still be shaped by English sources. An English answer can still inherit facts from German pages. The visible answer language is not the same as the source language. I have seen answers in German cite an English association profile because it had the clearest company category. I have seen English answers repeat an old German directory label because no English page stated the role directly enough.
The result looks like a bad translation, but it is usually a bad source path. One sentence crosses the language border and drags its assumptions with it. A manufacturer becomes “a supplier.” A compliance documentation service becomes “administrative outsourcing.” A specialist supplier becomes a “general industrial services provider.” None of those phrases are insane. That is what makes them dangerous. They are close enough to survive.
The practical work starts by saving the answer exactly: query, language, date, engine. Then I mark which sources seem to carry which claims. If the answer says “German supplier of cooling systems,” I look for that phrase or its near relatives in English and German. If the answer says “manufacturer,” I ask where the evidence came from. Was it on the company’s own site, in a PDF, in a directory, or inferred from product descriptions? If the answer says “reseller,” I do the same. The wrong claim usually has an address.
A lot of teams skip this step because the answer feels self-contained. It is not. It is more like a receipt printed without prices. You can see the purchased items, but the calculation is hidden unless you reconstruct it. The bilingual source path is that calculation.
When German pages beat English in local AI answers, the reason is rarely national preference. It is usually that the German page contains the more claim-supporting source language for the local query.
English export pages need role language before market language
The repair is not to make English pages sound German. It is to stop making English pages carry only the sales layer. Export pages often open with market language because that is what a human buyer wants to understand quickly: industries served, regions covered, product families, partner types, delivery scope. Fine. But the business role has to come before the broad language begins.
For the composite cooling manufacturer, the weak English pattern looked something like this: “We supply industrial cooling solutions for machinery and production environments across Europe.” It is plausible, but it hides the manufacturing role. A stronger opening would put the role first: “We design and manufacture industrial cooling components and custom cooling systems for machinery builders and production environments.” Then the page can name markets, countries, partners, and product families.
That one change matters because answer engines often compress from the first stable entity sentence. They may read further, but they do not always carry every nuance forward. If the opening sentence is vague, later proof has to work harder.
German pages need a similar discipline. Technical proof is useful, but not if the page never states the business role in ordinary language. A table full of product terms may suggest manufacturing, but a model may avoid asserting it without a clean sentence. The best German page for AI search usually combines both: a plain identity sentence and the technical nouns that prove it.
I do not mean stuffing pages with repeated category phrases. That can make the record noisier. I mean giving the machine the same thing a careful buyer wants: an answer to “who are you, what role do you play, where do you operate, and what proof supports that?”
When German should remain different from English
Some teams react to mismatched answers by trying to make the German and English pages identical. That creates another problem. The German page may need terms from standards, procurement, manufacturing, service law, or local categories that do not move cleanly into English. The English page may need to explain a German business type that has no simple equivalent.
The useful rule is not sameness. It is controlled disagreement.
A German page may say Fertigung, Sonderbau, Prüfstand, and technische Dokumentation. The English page may say manufacturing, custom systems, test bench, and technical documentation. Those are aligned. They do not have to be word-for-word twins. What cannot happen is this: the German page says the company manufactures, the English page says it supplies, a directory says it distributes, and an association entry says it provides consulting. That is no longer adaptation. That is entity drift.
Controlled disagreement also means knowing which language should carry which proof. A German page for local AI answers should carry local service areas, German categories, legal entity details, certifications, and precise product or service terms. An English page for export-facing AI answers should carry the same business role, plus international market language, export regions, and buyer-context terms. Both need enough proof to stand alone.
If an answer engine reads only the English page, it should not misunderstand the company. If it reads only the German page, it should not miss the market role. If it reads both, it should not have to choose between two different businesses wearing the same logo.
That is the standard I use in source repair.
The page that sounds least written may be doing the most work
There is a temptation to judge AI visibility by polished copy. I do not trust that instinct. Some of the strongest AI-cited sources are ugly, old, bureaucratic, or far too dry for a homepage. They work because they give a clear claim and a public context around it. A German product PDF from a manufacturer can carry more truth than a careful English landing page. A trade association profile can stabilize an entity better than a brand paragraph. A procurement listing can introduce a category that the company forgot to state on its own site.
That does not mean the ugly source is always right. It only means it may be legible.
For German SMEs, the repair usually starts by asking which source already teaches the right thing. If the German page teaches it, the English page may need alignment. If the English page teaches it, the German page may need a clearer local identity sentence. If a directory teaches it better than the company site, the company site has a problem. If a directory teaches the wrong thing and gets cited anyway, the repair has to include that external profile.
The local AI answer is not a verdict on which language is more important. It is a trace of which evidence was easiest to use.