Dattva Research · July 2026

llms.txt for B2B Companies: What It Actually Does for AI Visibility and Whether It Is Worth Building

llms.txt is a proposed root-level text file that lists a website's key pages for AI systems to read, and as of mid-2026 no major AI platform, including ChatGPT, Claude, Perplexity, or Gemini, has confirmed using it to decide what gets cited. For most B2B companies it is a low-priority technical item, not the fix that produces AI visibility.

Does llms.txt Actually Change What ChatGPT or Perplexity Cites?

No, not currently, and the evidence for this is close to unanimous. Google's Gary Illyes confirmed in mid-2025 that Google does not support llms.txt and has no plans to, and John Mueller has compared it publicly to the discredited meta keywords tag. No major AI provider, including OpenAI, Anthropic, Google, or Mistral, has stated that llms.txt is read as a ranking or citation signal in production search or answer surfaces.

Search Engine Land tracked ten sites across finance, B2B SaaS, ecommerce, insurance, and pet care for 90 days before and after llms.txt implementation (Search Engine Land, January 2026). Eight of the ten sites saw no measurable change in AI referral traffic. One site declined. The two sites that did see traffic gains had also run PR campaigns and restructured product pages with extractable comparison tables in the same window, which makes llms.txt a poor candidate for the actual cause.

This does not mean the file is useless. It means it is not doing the job most agencies are selling it as. The confusion between AI crawler discovery and AI citation influence is where most of the wasted budget in this space comes from.

What llms.txt Was Actually Built to Solve

llms.txt exists to compress a website into a form a language model can read cheaply, not to influence search rankings. The specification, proposed in 2024, asks a site to publish a single markdown file at the root that lists its most important pages with short descriptions, plus links to page-level markdown versions of key content. The idea was that a model reading a website through an API call pays for every token it processes, and a bloated, JavaScript-heavy page costs far more tokens to parse than a clean markdown summary of the same content.

This is a real cost problem, but it is a narrower one than most GEO commentary suggests. It affects AI agents and coding tools that fetch a site on demand and need to understand it quickly, not the crawlers that build search indexes weeks or months in advance. Mintlify rolled out llms.txt automatically across its hosted documentation platform in November 2024, which put the file on thousands of developer docs sites overnight, including Anthropic, Cursor, Pinecone, and Windsurf (Elementera AI, May 2026). Stripe, Vercel, Cloudflare, and Supabase all publish one on their developer docs today.

The pattern across all of these examples is the same: llms.txt earns its keep on documentation and developer-facing sites, where AI coding agents such as Cursor, Claude Code, GitHub Copilot, Cline, Windsurf, and Aider fetch the file directly to understand a codebase or API surface before writing code against it. A B2B marketing site with no API, no developer docs, and no agent-facing integration surface is asking a file built for a different job to do work it was never designed for.

What the Adoption and Crawl Data Actually Shows

Adoption sits at roughly one in ten sites, and growth has been slow. An SE Ranking analysis of 300,000 domains found a 10.13% adoption rate after eighteen months of industry discussion about the standard (SE Ranking, cited in Limy, May 2026). A separate audit of the top 1,000 sites globally found a more conservative 8.7% confirmed adoption rate as of June 2026 (Rankability, June 2026).

Crawler interest in the file is close to negligible. Limy monitored more than 500 million AI bot traffic events across the brands it tracks over a 90-day window and found only 408 requests targeting llms.txt directly (Limy, May 2026). GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot, and Google-Extended overwhelmingly crawl HTML directly and skip the file, which matches what server log analysis has shown since the specification was first proposed.

The one documented case of direct impact involved a German digital agency that published a static llms.txt alongside JSON-LD structured data in February 2026 and submitted the URL through Google Search Console. Three days later, Google AI Mode cited the file as the primary source for a brand-name query (Elementera AI, May 2026). Even the agency that ran this test describes the effect as building an identity layer for a single branded query, not a general citation lift.

Where llms.txt Does Earn Its Place

Publish it if the company has developer-facing content. A B2B software company with API documentation, integration guides, or an SDK should treat llms.txt as part of the same infrastructure that already serves AI coding agents. This is the one use case with a documented mechanism and a plausible return.

Keep the file to genuine discovery, not marketing copy. The root file should list pages and one-line descriptions in plain language, the same way a sitemap lists URLs. Filling it with brand positioning statements or keyword-heavy descriptions defeats its purpose and adds nothing, since no citation-facing crawler is reading it as ranking input.

Do not redirect budget away from the fixes with a documented citation path to build this file instead. Bing Webmaster Tools verification, Brave Search index submission, robots.txt entries for GPTBot and ClaudeBot, and direct-answer content structure all have a traceable route to AI citation. llms.txt does not, at least not yet, and treating it as equal priority is a resourcing mistake.

Add a closing note on what the file deliberately excludes. If a company has confidential client reports, gated research, or authenticated sections, a short note in the llms.txt file stating that private material is out of scope is useful housekeeping. It does not function as an access control on its own, so sensitive paths still need separate handling in robots.txt or behind authentication.

Where This Is Headed

The likely trajectory is narrow, permanent adoption rather than either death or universal standardisation. Platform tools like Yoast and Webflow have already built one-click llms.txt generation into their publishing flows, which will keep adoption climbing on the low end even without any provider announcing support for it as a citation signal.

The more consequential shift is a separation Anthropic has already made explicit in its crawler documentation: ClaudeBot handles training data, a separate search-facing crawler handles citation retrieval, and llms.txt sits closer to the agentic and developer-tooling side of that split than the search side. That framing, infrastructure for agents rather than infrastructure for search, is the most accurate way to think about where it fits.

For a Dattva client, the practical takeaway does not change based on future speculation. The file is worth five minutes of a developer's time on a documentation-heavy or API-heavy property. It is not worth restructuring a content calendar around.

Conclusion

llms.txt is a legitimate piece of infrastructure for a narrow use case and a distraction when treated as a general AI visibility fix. The data across multiple independent trackers points the same direction: low adoption, negligible crawler interest from the bots that actually decide citations, and one credible case study that produced a single-query effect rather than a broad lift. Companies with developer docs or API surfaces should publish the file because AI coding agents genuinely use it. Companies without that surface area get more return from robots.txt configuration, Bing and Brave indexing, and content structured for direct-answer extraction. The file is not the problem to solve first, and it is rarely the problem at all.

Frequently Asked Questions

Written by the Dattva Research Team, the technical SEO and AI visibility research group at Dattva AI, an implementation-focused GEO agency.