llms.txt: 844,000 Implementations, Zero Confirmed AI Usage

Why llms.txt is on every DACH B2B agenda right now

Few technical proposals shaped the GEO debate in 2025 as much as llms.txt. The idea sounds compelling: a single Markdown file in the root directory that gives AI systems such as ChatGPT, Perplexity, or Google AI Overviews a curated map of a website's most important content. Plugins generate it automatically, conference talks recommend it, and many requirement specifications for website relaunches now list it as a standard item.

This is exactly where the problem starts for B2B decision-makers in the DACH region: adoption suggests effectiveness, but the evidence does not support that expectation. No major AI provider has confirmed that it retrieves or evaluates the file. If you allocate GEO budget responsibly, you need a sober assessment rather than an article of faith: what llms.txt can deliver today, what it cannot, and which tactics produce measurable AI visibility instead. This article separates verifiable facts from marketing promises.

What llms.txt is and how it differs from robots.txt and sitemap.xml

llms.txt was proposed on September 3, 2024, by Jeremy Howard, co-founder of the research lab Answer.AI. It is a Markdown file at /llms.txt that bundles brief background information, guidance, and links to more detailed Markdown files. The only mandatory element is an H1 heading with the project or site name; the original states verbatim: "This is the only required section". All other sections, such as an optional block with secondary resources, are not required.

A clear distinction matters because the three files are often lumped together:

• robots.txt: controls what crawlers may and may not do. It is an access and exclusion control, not a content offering.
• sitemap.xml: lists all indexable URLs in machine-readable form. Its job is complete discovery, not prioritizing or preparing content.
• llms.txt: is a curated table of contents in Markdown format, readable by humans and machines alike. It is not meant to list everything but to highlight what matters most and provide it in an LLM-friendly form.

The key difference: robots.txt and sitemap.xml are established standards that search engines actually evaluate. llms.txt is a proposal without official support from the AI providers.

The problem llms.txt was originally meant to solve

Howard justifies the proposal in two places. In the specification on llmstxt.org, he writes that context windows are still too small for most websites in their entirety, and that converting HTML pages with navigation, ads, and JavaScript back into clean text for language models is difficult and imprecise. In the Answer.AI blog post, he stresses that assembling the right context from a website is ambiguous. The underlying assumption is that site owners know best which content is central and can provide this curated selection in machine-readable form.

The basic idea is understandable and technically sound. It explains why the proposal caught on so quickly with developers and tool vendors. What it does not explain is whether AI search systems actually use the file in practice. This gap between a plausible idea and real-world use is the core of the issue.

The adoption paradox: high absolute adoption, low reach

Adoption of llms.txt is often communicated in absolute numbers, which creates the impression of a de facto standard. On October 25, 2025, for example, BuiltWith counted more than 844,000 websites with llms.txt, as documented by Publii. The live count in the BuiltWith tracker has fluctuated since then. The relative view is more reliable. An analysis by SE Ranking shows: of around 300,000 domains examined, only 10.13 percent had implemented an llms.txt file. In other words, almost 90 percent of the domains examined do without it.

This paradox is central to any strategic assessment: a large absolute number of implementations does not mean that the authoritative sites relevant to AI visibility have the file. Many implementations come from automated plugin generation, not from a deliberate editorial decision. Adoption is a weak signal here. It measures how easily something can be generated, not how effective it is.

The evidence against active use by AI search systems

Three independent pieces of evidence contradict the assumption that AI search systems actively use llms.txt today.

First, Google's statement. Google's John Mueller said that no AI system currently uses the llms.txt file. For AI Overviews, Google's guidance is to "use normal SEO practices", and according to Google you need neither GEO nor LLMO nor anything else on top. In addition, Gary Illyes made clear that Google does not support llms.txt and has no plans to.

Second, the reality in server logs. An experiment by OtterlyAI analyzed real crawler access. The result: of more than 62,100 AI bot visits over 90 days, only 84 requests targeted the /llms.txt file, a share of around 0.1 percent of AI bot traffic. For comparison: an average page on the same site received about 265 AI bot visits over the same period, so llms.txt, with 84 visits, was requested about three times less often. Crawlers do not request the file systematically.

Third, the correlation analysis. SE Ranking examined whether the presence of an llms.txt file is related to AI citation frequency. Both the statistical analysis and the machine learning model showed no effect of llms.txt on how often a domain is cited by LLMs. Particularly revealing: when the llms.txt variable was removed from the model, the predictions actually improved. The file contributed nothing to explaining AI visibility.

What actually drives AI visibility

If llms.txt does not explain AI citations, the really relevant question is: what does? Here, the data is much clearer and directly actionable for GEO practice.

Brand mentions over backlinks: An Ahrefs analysis of 75,000 brands shows that brand mentions on the web correlate much more strongly with visibility in AI Overviews at r=0.664 than backlinks at r=0.218, a ratio of roughly three to one. In practice, this means digital PR, expert articles, industry directories, and consistent mentions across the web have a stronger effect than classic link building.

Statistics and source citations in content: The paper from the KDD '24 proceedings, often cited as the Princeton GEO study, shows that targeted optimization can increase visibility in generative answers by up to 40 percent. The most effective methods are adding statistics, including credible quotations, and citing reliable sources, with relative improvements of 30 to 40 percent on the position-adjusted word count metric and 15 to 30 percent on the subjective impression measure. According to the study, keyword-stuffing-style methods bring hardly any improvement. The benchmark comprised 10,000 queries from different domains.

Front-loading and answer islands: Content that delivers the core answer in the first sentences and is organized into clearly delimited blocks that can be cited on their own is easier for language models to extract. This is not a file format trick but an editorial discipline that works on every single page.

The conclusion is clear: the levers for AI visibility lie in the content itself and in brand presence, not in a separate directory file.

Where llms.txt is actually useful outside search

The lack of impact on AI search visibility does not mean the file is worthless. It has legitimate use cases; they just lie outside the search context.

• Agentic workflows and tool use: When a user gives an AI agent a specific URL and asks it to process it, a curated Markdown map can make the entry point easier. Here, the file is accessed deliberately, not discovered by a search crawler.
• Internal RAG pipelines: Companies building their own retrieval systems over their own documentation benefit from a clean, pre-structured Markdown source. The benefit arises in a controlled internal system, not on the open web.
• Documentation consumption for dev tools and SaaS: In the developer and SaaS segment, where users actively point language models at product documentation, llms.txt can improve access to technical content. This segment is the most plausible real-world benefit.

For a typical B2B company in the DACH region whose goal is visibility in ChatGPT Search, Perplexity, or Google AI Overviews, these use cases usually do not apply.

Weighing risk, effort, and benefit realistically

An honest assessment: low risk, low cost, low upside. The risk is low because the file doesn't break anything. The effort is low as long as you only create it once. Based on current data, the expected benefit for AI search visibility is practically zero.

Three aspects still deserve attention:

• Maintenance effort: A manually maintained llms.txt goes stale. If content changes but the file does not, contradictory snapshots emerge.
• Outdated snapshots: A wrong curated map is worse than none, because it paints a distorted picture of the site.
• Spam and cloaking risks: A file that shows different content than the actual pages can drift toward cloaking. Consistency between llms.txt and the real site is therefore mandatory.

If you implement llms.txt, treat it as an automated, low-maintenance add-on, not as a GEO measure with a promise of impact.

Relevance for GEO strategy in DACH and Austria

The urgency of AI visibility in the DACH region is real, regardless of llms.txt. In Austria, according to Statistics Austria, 31 percent of 16- to 74-year-olds have used a generative AI tool such as ChatGPT at least once, and 28 percent also did so in the three months before the survey. The data comes from the ICT Usage in Households 2024 survey.

For B2B GEO strategies, this has two implications. First: almost a third of 16- to 74-year-olds in Austria have already used generative AI tools, so AI visibility is not a topic for the future. Second: the right response is verifiable, well-structured content with sources and statistics, not a technical file without proven effect. Especially in the German-speaking region, where high-quality specialist content is often sparsely represented in LLM training corpora, substantial content is the more effective lever.

How to measure AI referral traffic and conversions correctly

AI-driven research is fundamentally changing click behavior, which makes measurement more demanding. A Pew Research Center analysis based on US browsing data shows: users who saw an AI summary clicked on a traditional search result in only 8 percent of visits, compared with 15 percent without an AI summary. Only 1 percent clicked on a link within the summary itself. In addition, 26 percent of users ended their session after a page with an AI summary, compared with 16 percent without. These figures come from the US market, but the direction applies to DACH as well.

In practice, this means classic click metrics underestimate the value of AI visibility, because part of the effect occurs through citation-based impressions without a click. To capture the remaining referral traffic properly, use a two-pronged approach:

• GA4 segmentation: Isolate AI sources such as ChatGPT, Perplexity, or Gemini as separate referral sources and evaluate their behavior, such as engagement and conversions, separately.
• Server log analysis: Analyze AI bot crawler access directly in the logs to see which pages are actually retrieved. This exact method showed in the OtterlyAI experiment that llms.txt is hardly ever requested.

If you combine both levels, you measure AI visibility reliably instead of assuming it based on a file format.

Recommendation: the llms.txt decision matrix

The decision can be broken down into a simple matrix.

• Automated generation available, for example via your CMS or a plugin: implement it. The effort is minimal, the risk is low, and in a dev tool or SaaS context there is a plausible side benefit. Expectation: no measurable effect on AI search visibility.
• Only manual maintenance possible and no agentic or documentation use case: skip it or treat it as a strictly low priority. The maintenance effort exceeds the proven benefit.
• Limited GEO budget: don't invest in llms.txt; invest in the tactics backed by evidence.

Based on the data, the levers with proven impact take priority: consistent brand mentions across the web instead of pure link building, front-loading answers and using statistics and source citations directly in your content, and clean measurement via GA4 segmentation and server logs. At best, llms.txt is a harmless add-on. The real GEO work happens on the page itself, not in a directory file.