An llms.txt file is easy to install and easy to overread. The field question is narrower: after the file exists, do German business answers cite the site differently, paraphrase it differently, or keep choosing the same outside sources?
A composite test around Object B, the Leipzig B2B maintenance and industrial services firm, starts with a plain annoyance. The company’s German service pages explain maintenance scope, emergency availability, and industrial service limits. AI answers still lean on a local directory, an older brand profile, and occasional uncited wording that makes the firm sound like an equipment supplier. A technical fix looks attractive because the public evidence is messy.
So the site condition changes. An llms.txt file is placed at the domain root, pointing to the pages the company wants machines to read first. The lab does not celebrate the change. It logs it. Then the same German and English prompt sets are run again, with company-name prompts, service-category prompts, regional modifiers, and buyer-intent questions. The interesting result is not whether the file exists. The interesting result is what, if anything, changes in the answer.
Treating llms.txt as a condition, not a lever
An llms.txt test is a bounded observation of answer behavior before and after a visible guidance file appears on a site. That definition keeps the claim small enough to be useful. The lab is not measuring the whole web. It is not proving crawler behavior. It is asking whether recorded answers show a different source path after one site condition changes.
The distinction matters because llms.txt attracts checklist thinking. A file can be added, verified, and shown to a client. It feels like a clean intervention. German AI visibility rarely behaves that cleanly. An answer may still cite a directory because the directory is easier to summarize. It may still use an English trade profile because the query is in English. It may paraphrase the owned page more closely without citing it. It may do nothing visible at all.
The lab therefore records four kinds of behavior after the file appears. The site may be cited. The site may be paraphrased without a visible citation. The site may be ignored while other sources carry the answer. Or the site may appear beside competing sources that still shape the category. None of these outcomes is dramatic by itself. The pattern across repeated runs is what matters.
This approach also protects the reader from a false binary. The question is not “does llms.txt work?” as if every answer engine shared the same rules and every query exposed them. The better question is whether the file changes observable answer behavior for a particular German business, inside a defined prompt set, while other source paths remain in view.
What the lab records before the file
The before state is often more valuable than people expect. If Object B is already being described from a local directory, the lab needs to know which claim the directory supplies. Is it the company name, the Leipzig location, the broad industrial category, or the old brand wording? A later change cannot be interpreted unless the earlier source roles are clear.
For German business queries, the before record usually includes more than one path. The owned site may appear for company-name prompts. A directory bridge may appear for category prompts. An English or translated source may appear for English variants. An uncited assertion may show up in comparison answers. That mixed baseline is not noise. It is the field.
The lab uses its qualitative anchor here: four citation paths in German AI visibility — native source, translated source, directory bridge, and uncited assertion. The llms.txt file is then tested against those paths. Does the native source become more visible? Does the translated source lose influence? Does the directory bridge remain the carrier for category prompts? Do uncited assertions become better aligned with the owned pages?
This is also where the team checks whether the owned pages deserve to be preferred. A guidance file pointing to vague service pages will not fix vague evidence. If a German page says “solutions for industry” while a directory says “industrial maintenance in Leipzig,” the directory may still provide the clearer category. The lab does not assume owned means better. It inspects the wording.
What counts as a visible change
A clean change would look like this: before the file, category prompts cite a directory bridge and describe Object B as a general industrial supplier; after the file, repeated prompts cite the German service page and describe the firm as a maintenance and industrial services provider. That would be worth recording. It still would not prove that llms.txt caused the change, but it would show a source-path shift after the condition was introduced.
A weaker change may be more common. The answer still cites the directory, but it adds a sentence closer to the owned German service page. The visible citation path has not changed, while the answer wording has. The lab marks that as possible paraphrase movement, not citation gain. The distinction is small and important. A business may care about being described accurately even when the citation line remains unchanged.
Another outcome is source crowding. The owned site appears, but the directory bridge and older profile remain beside it. The answer becomes a small negotiation among sources. If the owned page supplies maintenance scope but the older profile supplies a broader supplier label, the final paragraph may still drift. More sources do not automatically produce a clearer business story.
The quiet outcome is no visible change. The file exists. The prompts are repeated. The same sources appear, or the same uncited wording returns. This is not a failed experiment if it is recorded properly. It tells the lab that, inside that prompt set, llms.txt did not visibly alter source selection or answer wording. That negative result is more useful than a cheerful case study with no baseline.
The file matters only if answer behavior changes in a way the record can see and describe.
That is the lab’s restraint in one line. A technical condition has to meet the same evidence standard as any other observation.
German-English runs complicate the test
A German-only test can make llms.txt look clearer than it is. The file may point to German service pages, and German prompts may already favor native sources. If the answer improves, the lab has to ask whether the file mattered or whether the prompt language naturally kept the system close to German evidence.
English prompts expose a different problem. If Object B has limited English material, answer engines may still reach for directories, older profiles, or translated summaries. An llms.txt file written in German or pointing mainly to German pages may not change the English answer path in a visible way. Or it may help the system find the German page but still lead to an English paraphrase that blurs the category.
The lab does not treat this as failure. It treats it as source-path competition. The query language shift may be stronger than the site condition. If English prompts keep selecting English-readable third-party sources, the issue may belong to source alignment rather than technical discovery. That belongs near work on English profile overrides, but the llms.txt material stays focused on the file as a tested condition.
There is another complication: the file can be well structured while the public source set around the business remains contradictory. A local directory may call the company one thing. An old brand profile may call it another. The German site may use a third phrasing. Answer engines often assemble from the surrounding evidence, not from one preferred document. The file may make one path more legible without removing the others.
That is why the lab does not test llms.txt in isolation from source review. It records the file, the linked pages, the prompt set, the answer wording, and the cited or implied sources. Then it asks whether the file appears to change the path. If the answer still drifts, the drift may be coming from contradiction rather than absence of guidance.
What a cautious result can say
The most responsible positive result is conditional. The lab may say that after an llms.txt file was added and the linked pages were clarified, repeated German prompts more often used native source wording for a particular business category. That is not a universal claim. It is a bounded observation with a before state, an after state, and a prompt set.
A cautious neutral result is also possible. The lab may find that the file coexists with the same citation paths as before. The owned site may appear in company-name answers, while directory bridges still dominate service-category prompts. That finding would suggest that the file did not overcome the role of third-party summaries for discovery-style questions.
A cautious negative result avoids drama. The file may produce no visible change because the answer engines in the sample do not use it, because the prompts do not trigger the linked pages, because competing evidence is stronger, or because the interface does not expose the relevant path. The lab cannot choose the most convenient explanation. It can only mark what the record shows.
This restraint is especially important for agencies. A client may ask whether llms.txt is worth adding. The honest answer is that it may be worth testing as a low-cost site condition, but it should not be sold as a citation guarantee. The value is in the measurement discipline around it: baseline runs, linked-source review, repeated prompts, and careful comparison of wording.
Limits of the llms.txt record
The lab cannot prove that an answer engine read the file unless the engine or its documentation exposes that behavior in a verifiable way for the case at hand. Recorded answers can show correlation with a site change. They cannot always show mechanism. A visible citation to a linked page after the file appears is suggestive, not proof of causation.
The method also cannot separate llms.txt from simultaneous edits if the site changes too much at once. If the file is added, service pages are rewritten, directories are repaired, and English profiles are updated in the same period, the after-state becomes muddy. The lab prefers one change at a time when the research question is narrow. Real businesses do not always have that patience, so the record marks the complication.
Prompt sets remain bounded. Company-name prompts, regional prompts, service-category prompts, buyer-intent questions, and German-English variants may each behave differently. Citation share is described only inside the set. The lab does not invent a market score from a handful of observations.
The clearest conclusion is modest and still useful. An llms.txt file can be part of a German AI visibility experiment when it is treated as an observable condition, not as a control panel. The fieldwork asks whether answer engines cite the site, paraphrase it, ignore it, or keep surrounding it with older and broader sources. Anything beyond that belongs in the uncertainty column until the answer record earns a stronger claim.