llms.txt may become a useful signpost for AI systems, but it is not a remote control for how answer engines understand a German business.
A composite manufacturer near Hamburg asked a blunt question after seeing a wrong AI answer: can we add llms.txt and stop this? The answer had framed the company as a reseller of industrial cooling components, even though its German pages showed design and manufacturing work. The marketing lead had read about files that tell AI crawlers what to use. It sounded tidy. Put the right page in the right file, block the bad paths, and the machine should behave.
I understand the appeal. A public web record is untidy and slow to repair. Old distributor profiles do not vanish on command. Procurement entries take time. Directory categories have their own forms and rules. A small text file in the root of the website feels like a door handle. Turn it, and the building opens correctly. But AI visibility does not work like a locked office. It works more like a yard full of documents after a windy day: the file near the gate may help, but it does not gather every sheet.
llms.txt belongs to access, not meaning
The idea behind llms.txt is straightforward enough. A site can publish a machine-readable file that points language models or AI tools toward useful content, policies, summaries, or preferred documentation paths. The exact adoption and interpretation depend on the systems reading it, and no German SME should assume universal obedience. That uncertainty is important. A file can be well written and still not be used by the answer engine that produced the problematic answer.
For AI SEO work, I put llms.txt in the access layer. It can help describe which pages are useful to AI systems. It may guide tools toward clean documentation. It can sit beside robots.txt, sitemaps, structured data, and crawlable page architecture as part of a wider technical record. It does not decide whether the company is a manufacturer, reseller, platform, agency, or local service firm. The pages still have to say that.
Here is the working definition I use with clients: llms.txt is a machine-facing site guide, because it can point AI systems toward selected public content but cannot make unsupported claims true.
That definition is deliberately modest. Modesty prevents bad repair work. If a business is misclassified because an old English distributor page says “supplier of a wide cooling range,” then an llms.txt file on the company domain may not neutralize that external source. It may help a cooperative system find the company’s preferred pages. It will not rewrite the distributor profile, the procurement portal, or the association entry that an answer engine also retrieved.
The Hamburg composite manufacturer needs this distinction. Its own site can add a neat file listing the current German and English product pages, the company profile, and a technical overview. That is sensible. Yet the wrong role came partly from sources outside the domain. Access to the right pages is only one piece. The meaning of those pages, and the public contradiction around them, still has to be repaired.
The false promise of a single switch
German SMEs often have a strong appetite for controlled systems. A good technical page, a clean product taxonomy, a certification table, a precise PDF, a maintained sitemap: these are useful habits. The problem comes when the same desire for order meets answer engines. The system is not a single database waiting for a corrected field.
An answer may draw on the company site, an English export page, a distributor profile, a directory, a PDF, an association page, a procurement entry, and pages about nearby competitors. Some engines show citations. Some do not. Some answer from live retrieval. Some blend retrieval with model memory, query interpretation, and ranking choices that outsiders cannot fully inspect. A root file on one domain cannot control that whole environment.
This does not make llms.txt useless. It makes it smaller than the rumor around it.
I have seen the same pattern with robots.txt, schema, sitemaps, and other technical signals. A tool becomes a charm. Someone says, “We added it,” and the conversation stops before anyone checks the answer claim. But the machine did not misread the business because one technical file was missing. Usually it misread because the public evidence was split, vague, stale, or easier to summarize from the wrong source.
In the Hamburg composite, the German pages have stronger proof, but the role sentence comes too late. The English trail contains older channel language. The distributor profile is cleaner than the manufacturer’s own export page. An llms.txt file could point to the better pages. Still, those better pages should become better evidence: clearer opening role, stronger manufacturing verbs, aligned German-English entity description, and a current PDF that says what the company is before it says what it sells.
A signpost is not a road repair.
Where crawler guidance helps in practice
Crawler guidance helps most when the preferred content is already strong. If the site has a clear company page, a precise product overview, accessible technical PDFs, and consistent German and English descriptions, then an llms.txt file can act as a useful index for AI tools that choose to read it. It can say, in effect: these are the pages that explain the business; start here.
For a German manufacturer, the file might point to a company profile, a manufacturing capability page, key product families, certification pages, and selected English export pages. It might avoid dumping every low-value page into the guide. A machine does not need to be invited into every calendar notice, old campaign landing page, or duplicated PDF. Selection matters.
For a service company, the file might point to the main service explanation, a role-specific about page, case-study summaries, industry pages, and policy pages. Again, the file is only as good as the pages behind it. If those pages cannot support the claim, the guide becomes a neat map to weak evidence.
I use a small diagnostic for this called the access-meaning split. Access asks whether AI systems can find and read the preferred page. Meaning asks whether that page supports the claim we want the answer to make. Many llms.txt conversations stop at access. AI SEO work has to finish the meaning half.
The split is useful because it prevents technical and editorial teams from blaming each other too quickly. The developer may have made the content crawlable. The writer may have written a good paragraph. The problem may sit in an old PDF, a directory, or a missing bridge between German proof and English market language. Without the split, everyone points at the nearest tool.
In the manufacturer case, access could be improved by listing the current product and company pages in llms.txt. Meaning needs a separate repair: the pages should identify the firm as a designer and manufacturer before naming distributors, partners, product lines, or sales markets. If the role remains implicit, a guided crawler can still arrive at a page that expects too much prior knowledge.
What llms.txt cannot clean up outside the domain
The external record is where many German AI visibility problems become stubborn. A company can edit its own site quickly. It cannot instantly edit the distributor profile, the procurement portal, the trade directory, or the local press article. Some profiles are abandoned. Some are controlled by partners. Some require slow administrative changes. Some cannot be changed at all.
llms.txt does not remove those sources from the web. It does not tell Perplexity to ignore a directory. It does not force ChatGPT to prefer the newest English page. It does not guarantee that Google AI Overviews will quote the company’s own role sentence. Different systems behave differently, and their visible source behavior changes across queries.
The repair plan therefore has to rank sources by reach and claim damage. If an old distributor profile repeatedly carries the reseller framing, that profile deserves attention. If a procurement entry has broad but harmless wording, it may wait. If a product PDF on the company domain is being read and contains outdated partner phrasing, that is easier to fix and probably urgent. If a directory category is wrong but rarely appears in answers, it may be less urgent than the main English company page.
The imperfect detail in this composite is that the company’s own German page is not bad. It is technically respectable. It simply talks like a page written for people who already know the category. AI systems often need the obvious sentence earlier than specialists do. That can feel insulting to a precise company. It is still necessary.
A clean llms.txt file pointing to an implicit page is like a label saying “main entrance” on a door that opens into a dark room. Better than no label. Still not enough.
A sensible llms.txt place in the audit
I do not tell clients to ignore llms.txt. I tell them to place it correctly in the audit sequence.
First, record the answer. Exact query, language, date, engine. Second, identify the claim that worries the company. Third, inspect the sources that support or distort that claim. Fourth, repair the preferred pages so they state the entity role and proof clearly. Fifth, make those pages technically accessible through normal means: internal links, sitemap, indexable pages, readable PDFs where needed, structured data where useful. Then consider llms.txt as a guide to the best public evidence.
That order can feel slow. It is faster than decorating weak evidence.
For the Hamburg manufacturer, I would include llms.txt only after the core role pages are cleaned. The file might point to the German company page, the English manufacturer profile, the main product families, and a technical proof page. I would not expect it to erase the old distributor page. I would separately try to correct that profile or make the company’s current English page so clear that the old wording loses influence over repeated answers.
The next check must be done across repeated prompts. One run after a technical change proves very little. I would watch German and English queries, supplier phrases, manufacturer phrases, and comparison phrases. Did the role claim improve? Did the wrong source still appear? Did the preferred page get cited? Did the answer still use reseller language even when citing the company domain?
If the current trend in AI tooling holds, site owners may gain more ways to publish machine-facing guidance. That will be useful. It will not cancel the need for public evidence that says, plainly and repeatedly, what the business is.