Trust signals do not help much when they sit in a brochure sentence. They start to matter when an answer engine can lift them as proof for one specific claim about the company.
A German manufacturer can spend a long afternoon writing a page about quality. The page may mention experience, reliability, customer focus, certificates, test procedures, export markets, and technical depth. It may sound responsible. It may even sound quite German in the good way: measured, exact, unwilling to boast. Then an AI answer ignores the whole page and quotes a short trade-directory line that says “supplier of cooling products.”
That is the small insult I see often. In a composite scenario drawn from several industrial B2B observations, a 95-person company near Hamburg designs and manufactures specialist cooling components for machinery builders. Its German site contains the proof: test benches, engineering drawings, certification notes, and application examples. Its English-facing public footprint is thinner. A distributor profile calls it a broad catalogue supplier. In one answer run, the machine named the company correctly, placed it in the right industrial area, and still described it as a reseller. It also got one product family slightly wrong, which made the answer look researched while the core role was bent. That is where E-E-A-T becomes more practical than the phrase usually sounds. I do not treat experience, expertise, authority, and trust as a mood. I treat them as quotable evidence. If an answer engine cannot copy or summarize the proof, the signal is weaker than it looks to a human reader.
The machine does not quote your reputation, it quotes your record
On a human website, trust can be atmospheric. A buyer sees a sober design, technical drawings, a known trade association logo, names of departments, and careful German wording. The buyer may infer seriousness. They have a trained eye. They know what a real supplier page feels like.
An answer engine has a different problem. It has to produce a sentence. For that sentence, it needs a claim it can attach to a source. “This company is experienced” is too soft. “This company has supplied precision cooling assemblies to machine builders since 1998” is usable, if it is stated clearly on a readable page. “The company manufactures cooling components in northern Germany” is also usable. “We stand for quality and partnership” is decorative noise.
E-E-A-T signals are machine-quotable when they tie a business claim to visible proof, because AI search needs text it can reuse with confidence. That is my working definition. It sounds dry, but it saves a lot of wasted editing. The question is no longer “do we look trustworthy?” The question becomes “which sentence can the answer safely repeat, and what source backs it?”
In the Hamburg manufacturer scenario, the company had real proof. The trouble was placement. The best evidence sat inside PDFs, image captions, or German product pages with narrow technical vocabulary. The English company profile, which answer engines could more easily summarize for export-oriented questions, said less about manufacturing and more about product range. A third-party profile used even simpler words. The model took the easier bridge.
I call this the proof-distance problem. The strongest evidence is often too far from the sentence the machine needs to write.
Experience has to be dated, situated, and attached to work
Many German company pages treat experience as a ceremonial paragraph. “For decades we have served customers in demanding industries.” That may be true. It may even be modest. But an answer engine looking for a supplier recommendation cannot do much with it. It is not tied to a category, a market, a product, or an outcome.
Experience becomes useful when it is dated and situated. A sentence such as “Since 1998, the company has designed and manufactured cooling assemblies for packaging machinery and precision machine tools” gives the machine several stable hooks. Date. Activity. Product. Sector. Role. It reduces the chance that the company is flattened into a reseller or parts distributor.
This does not mean every page should become a timeline. In most cases, one or two exact experience statements are enough. They should appear where the answer engine is likely to look: the about page, the main product family page, a company profile page, and sometimes a schema-supported organization description. A buried anniversary news post from a few years ago is weak unless the same claim is repeated in the live entity description.
There is a German habit of putting the strongest details in places that feel proper to a human but awkward for a machine: a PDF certificate, a downloadable brochure, a historical page, an image of a production hall with a caption, or a trade fair handout. Those documents may still be indexed. They may still be cited. But when they are the only place the proof appears, the answer engine can miss the relationship between proof and current business role.
The Hamburg manufacturer had a page showing test facilities. Good proof. But the page did not say plainly that these facilities supported its own manufactured cooling systems. The model saw testing, supplier language, distributors, and product categories. It assembled a plausible but wrong role. The little missing sentence mattered more than the large amount of technical material around it.
Expertise is the part that must resist generic language
Expertise is where many B2B pages become too polite. The page says “custom solutions,” “high-quality products,” and “individual advice.” These phrases travel badly. They are common across manufacturers, resellers, engineering consultants, and service providers. AI answers do not have enough friction. They slide from one role to another.
For AI search, expertise should name the difficult thing the company knows how to do. The detail can be narrow. Actually, narrow is better. “Cooling modules for continuous-operation packaging lines” is more useful than “industrial cooling solutions.” “Thermal management for compact machine housings” gives the answer a technical contour. “Cooling products for industry” is a bucket with the label half rubbed off.
I use a small classification here: proof signals, role signals, and boundary signals. A proof signal shows why the company deserves the claim. A role signal states whether it manufactures, distributes, installs, advises, audits, or operates. A boundary signal says what the company does not mean by a broad category. German sites often have proof signals. English profiles often have market signals. Boundary signals are the missing middle.
For the composite manufacturer, the boundary signal would be something like this: “The company designs and manufactures cooling assemblies; it does not operate as a general spare-parts reseller.” That sentence may feel too blunt for public copy. It can be softened. But somewhere the distinction has to exist. Otherwise an old distributor profile may define the company more clearly than the company defines itself.
The rough part is that expertise can be damaged by overtranslation. A precise German term becomes a broad English phrase. “Fertigung” becomes “supply.” “Anlagenbau” becomes “systems.” “Kundenspezifische Auslegung” becomes “solutions.” None of those translations is automatically wrong. The problem appears when the English page loses the role. The machine sees a firm that supplies, offers, and supports, but not one that designs and manufactures.
Authority is usually borrowed before it is earned in the answer
Authority signals are not only awards and association badges. In AI answers, authority often arrives through other people’s pages. A trade directory, procurement portal, association profile, standards listing, reseller page, or local press article may carry more weight than the company’s own about page. Sometimes the outside source is cleaner. Sometimes it is simply older and easier to parse.
This is uncomfortable for marketing teams, because it means a weak third-party profile can become the public sentence the machine trusts. The company may have rewritten its site. The directory remains. The English distributor page remains. A procurement listing keeps a category from an earlier business phase. The answer engine does not know the politics of that residue. It sees a source.
Authority, then, needs alignment. The company’s own site should say the role clearly. The major external profiles should not contradict it. Association entries should use the same basic category. Product PDFs should not imply a different commercial role. If the company is a manufacturer, that word should not appear only in German while the English ecosystem says supplier, dealer, partner, or catalogue.
In my citation ledger, I often mark authority sources with a question: “Does this source deserve the weight it received?” A citation is not a win until the cited source supports the claim being made. If a machine cites an association page for a company’s manufacturing role, and the association page only lists a broad category, the answer is standing on a thin board.
The machine may still be correct by accident. That does not make the evidence healthy.
Trust signals need claim support, not just badges
Trust is the most abused of the four letters, and also the most useful when handled plainly. Certificates, standards, audits, data sheets, documented processes, warranties, named locations, service conditions, and update dates can all help. They help only when they support a specific claim.
A certificate logo without context is weaker than one sentence explaining what the certification covers. A quality page that says “certified processes” is weaker than a product page that says which process applies to which product family. A case study without a sector, constraint, or application is mostly a story. A case study that says “installed in a continuous packaging line with limited housing space” becomes evidence.
This is where German SMEs often have a hidden advantage. Their proof is real. They have process detail, certifications, long supplier histories, application knowledge, and sober documentation. The issue is rarely a total absence of substance. More often, the substance is locked in formats and pages that do not speak to each other.
For the Hamburg manufacturer, the repair would not begin with more trust language. It would begin by moving a few exact proof statements closer to the pages answer engines already read. The German product page should state the manufacturing role before the product range. The English export profile should carry the same role, not only the market pitch. The directory description should be cleaned if possible. Product PDFs should repeat the company role in text, not only in the cover design.
That is unglamorous work. Good. Unglamorous work is where many AI search problems become fixable.
The useful audit question is painfully specific
When I review E-E-A-T for AI search, I do not ask whether the site feels credible. I ask what the machine can safely quote. The audit becomes a row-by-row reading of claims. What does the answer say? Which source supported that sentence? Was the source German or English? Did it carry the company’s role, or only a category? Was the proof current, readable, and close enough to the claim?
Sometimes the answer is already mostly right. The company is cited, the role is correct, the geography is not distorted, and the source supports the claim. Then the repair may be small: strengthen a page title, add a clear entity sentence, align the English version, or remove an ambiguous category from a directory listing.
Sometimes the answer is wrong in a way that looks stubborn. It repeats the same third-party wording across several engines. In that case, the repair has to go where the citation weight lives. Editing the nicest page on the company website may not change much if the answer keeps leaning on an old procurement profile.
The practical test is simple. Take one AI answer. Underline every sentence about the business. For each sentence, write the source that could have carried it. If no source cleanly supports the sentence, the public record has a claim-support problem. If the wrong source supports the wrong sentence too clearly, the public record has an authority problem. If the right source exists but in the wrong language or format, the record has a proof-distance problem.
These are better problems than “AI does not understand us.” They can be worked on.