A B2B website can be clear to a buyer and still unclear to an answer engine. The buyer fills gaps from context; the machine borrows the nearest public sentence.
A service company in North Rhine-Westphalia has a neat website. It says it supports mid-sized industrial firms with compliance documentation and supplier onboarding. The German pages sound careful. The English page sounds broader, probably because somebody wanted it to travel better. Two directory profiles use different categories. One old listing calls the company a consulting office. Another makes it look like a software platform. A machine reads the pile and tries to be helpful.
In a composite scenario assembled from recurring B2B observations, the company has 42 people, real clients, and a practical niche. It helps industrial firms get supplier documents, compliance files, and onboarding evidence into usable order. In AI answers, though, it changes shape. One answer calls it a local consultant. Another places it beside software vendors. A third makes it sound like general administrative outsourcing. In one run, the company name was right but the service area was shifted toward HR paperwork, a small wrong detail that showed how loose the role had become. German B2B sites often do not lose AI visibility because they lack content. They lose role accuracy because the content gives answer engines several competing ways to describe the same business.
The first mistake is describing the activity without naming the role
Many B2B pages begin with verbs: support, accompany, improve, manage, coordinate, advise. The verbs are not wrong. They may match sales conversations. But a machine trying to classify a company needs nouns too. It needs to know whether the business is a consultancy, software provider, documentation service, manufacturer, agency, broker, integrator, auditor, or managed-service operator.
A buyer can infer this from tone, pricing, case studies, and a call with sales. An answer engine is less patient. If the site says “we support companies in supplier onboarding” and a directory says “software for supplier management,” the machine may choose the cleaner category. If another profile says “consulting for compliance,” the answer may drift again. The company becomes whatever public source uses the firmest noun.
This is what I call role drift. Role drift is the AI-search error where a company’s activity remains roughly visible, but its business role changes because public sources use competing category nouns. It is common in German B2B because many firms sit between services, systems, documentation, audits, and software-enabled work.
In the North Rhine-Westphalia scenario, the company did not need louder claims. It needed a stable role sentence. Something like: “The company is a B2B compliance documentation service for industrial supplier onboarding, combining managed document work with process advice.” That sentence is not beautiful. It is useful. It gives the machine a role, client type, function, and boundary.
Without that kind of sentence, every surrounding source gets a vote.
The second mistake is letting directories write the cleanest category
German B2B companies often treat directory profiles as minor housekeeping. Somebody filled them out years ago. A category was selected from a dropdown. A short description was copied from an old brochure. The listing stayed online. Later the website changed, the service changed, the English page changed, and the directory kept its little fossil.
AI search can like fossils. They are short. They are structured. They state a category plainly. A human may skip the directory and read the company website. A machine assembling an answer may find the directory easier to use than a nuanced service page.
In the composite service-company case, one directory profile used the equivalent of “business consulting.” Another leaned toward “supplier management software.” Neither was absurd. Both were incomplete. Together they created a category fork. When the model saw “supplier onboarding,” “compliance,” “platform,” “consulting,” and “documentation,” it had to settle the issue somehow. The answer that called the company a software platform was not random. It followed the strongest external noun.
This is why directory cleanup belongs inside AI SEO, even though it feels like old search hygiene. The point is not cosmetic consistency. The point is claim control. A directory profile may become the source that carries the category. If that source is wrong, the answer can be wrong with confidence.
The repair is usually dull and specific. Find the profiles that appear in answers, appear in search results, or repeat in snippets. Check the company name, role noun, service boundary, geography, and language. Do not turn every profile into a slogan. Make it boringly correct. Boringly correct is a good state for source evidence.
The third mistake is making German and English pages solve different problems
German pages and English pages often have different jobs. The German site carries precise service descriptions, legal vocabulary, procurement language, and domestic credibility. The English page carries export posture, market friendliness, and shorter explanations. That division makes sense for humans. It becomes risky when answer engines treat each language as a separate evidence system.
In the service-company scenario, the German pages made the documentation work clearer. The English page tried to sound accessible and used broader terms such as “supplier management support.” That phrase opened the door to software-platform readings. It did not say enough about the managed service component. It did not rule out outsourcing either. The machine had to combine signals from both languages, and it welded them at the wrong seam.
German-English mismatch does not mean the two versions must be identical. Mirrored translation can be just as bad. Each language should carry the same entity facts: what the company is, who it serves, what role it plays, what it does not replace, and which proof supports the claim. The examples can differ. The market language can differ. The core role cannot keep changing clothes.
A small roughness I often find: the German page uses one category in the headline, another in the footer, and the English page uses a third in the metadata. No human notices because each page works alone. The answer engine reads across the set. Then the inconsistency becomes a category machine.
A bilingual role sentence is sometimes the simplest repair. Write the German version first, because the technical and legal terms may be more exact. Then write the English version as a role-preserving explanation, not a sales paraphrase. If the English phrase becomes broader, add the boundary back.
The fourth mistake is hiding the proof behind service language
B2B service pages often describe what the company helps with, then leave proof scattered elsewhere. The case studies may be vague. The process page may be separate. The certificates may sit in a PDF. Team experience may appear in biographies. The answer engine sees the service claim but not the evidence that makes it credible.
For AI search, the proof should sit close to the role. If the company handles supplier onboarding documentation for industrial firms, show the proof nearby: document types, sectors served, process responsibilities, audit constraints, languages handled, systems interfaced with, and the kind of buyer who uses the service. A machine can quote concrete scope more easily than it can infer expertise from a general service paragraph.
In the composite case, the company had real operating detail. It knew how supplier files break: missing certificates, inconsistent vendor names, expired declarations, language mismatches, duplicated forms, procurement systems that accept one format and reject another. But the public page said “we simplify supplier onboarding.” That was too smooth. It gave away the texture that would have protected the role.
The texture matters. A sentence such as “The team checks supplier declarations, certificate validity, document completeness, and onboarding evidence for mid-sized industrial procurement teams” is not poetic. It is hard to misclassify. It pushes the answer away from software-only, consulting-only, and generic outsourcing readings.
I often think of proof as grit in the sentence. Without it, the sentence slides.
The fifth mistake is reporting mentions without checking claim accuracy
A company appears in an AI answer. Someone takes a screenshot. The mood improves for half an hour. Then I ask what claim the answer made. Did it call the company the right kind of business? Did it cite a source that supports that role? Was the answer in German or English? Did it recommend the firm for the right buyer situation?
This is where many B2B teams discover that a mention is not the same as visibility. A mention with the wrong role can create bad demand. A mention with the wrong geography can shrink a company’s market. A mention beside the wrong competitors can teach sales a painful lesson. A mention sourced from a weak directory can be unstable.
The service-company scenario showed all four risks. In one answer, the firm appeared in a supplier-onboarding context, but the neighboring companies were software platforms. In another, it was grouped with local consultants. The company was present. The category was moving. A screenshot would have hidden the problem.
A useful AI SEO report separates presence, source, role, and claim support. Presence asks whether the company appeared. Source asks what carried weight. Role asks how the business was classified. Claim support asks whether the cited source actually backs the answer. Those four fields are enough to stop a lot of false comfort.
This is also why I do not recommend repairs from a single prompt run when the pattern can be checked. A single bad answer may be noise. Repeated role drift across engines, languages, and related queries is evidence.
The repair starts with one sentence, then moves outward
The temptation is to begin everywhere: schema, blog posts, directory cleanup, about page, service pages, PDFs, FAQs, and reporting dashboards. Sometimes that work is needed. But the first repair is usually one sentence. It should answer the question a machine is trying to settle: what is this company?
For a German B2B site, that sentence needs a role noun, a client type, a service boundary, and one proof-bearing detail. It should appear on the homepage, about page, main service page, and major public profiles. It should survive translation. It should be boring enough to stay stable.
Then the surrounding evidence can be aligned. Directory profiles should stop using old category labels. German and English pages should carry the same role facts. Case studies should show the actual work, not only the benefit. Schema should support the entity, not decorate the page. PDFs should repeat the company role in text. Reports should track whether AI answers repeat the repaired claim.
In the North Rhine-Westphalia scenario, the company did not need to pretend to be simpler than it was. It needed to be explicitly complex in the right way. Managed documentation work, compliance process advice, supplier onboarding support: those pieces can coexist. The page just has to tell the machine which piece is the role and which pieces are supporting functions.
If the company does not say that, the next directory will.