What Is llms.txt, and Does Your B2B SaaS Site Actually Need One for AI Answer Engines?

llms.txt is a plain-text file at the root of your domain (yoursite.com/llms.txt) that hands large language models a curated map of your site: what you sell, who buys it, which pages matter, and where the clean documentation lives. Jeremy Howard proposed it in September 2024 as a sort of robots.txt for AI, and it has picked up real adoption among SaaS and developer-tool companies.

Does your B2B SaaS site need one? Add it. It takes an afternoon and costs nothing. Just do not call it your AEO project. No major answer engine has publicly confirmed it reads the file, and three other things decide whether ChatGPT, Claude, Gemini, and Perplexity can cite you at all.

We put a number on that judgment instead of hand-waving it. Our AI Readiness score runs 0 to 100 across four weighted dimensions: crawler access at 30 percent, content structure at 30 percent, structured data at 25 percent, and llms.txt at 15 percent. That last weight is deliberate. llms.txt is the smallest slice because it is the only one of the four with no confirmed consumer on the other end. The other three carry 85 percent, and they are the ones with evidence behind them.

What is actually in an llms.txt file?

Markdown, in a file with a .txt extension. An H1 with your company name. A blockquote summary of what you sell and who buys it. Then grouped link lists: documentation, product pages, pricing, a handful of blog posts, each with a one-line description of what a reader finds there.

That is the entire format. No schema to validate against, no registry, no verification step, nobody checking your work. A companion convention, llms-full.txt, dumps your whole documentation set into a single markdown file so a model can ingest it in one fetch. A dev-tool company with 400 pages of API reference gets something out of that. A fifteen-page SaaS marketing site does not.

Do AI answer engines actually read llms.txt?

No major engine has said publicly that it does. ChatGPT, Perplexity, Google's AI Overviews, and Copilot have never named it as a crawl input or a ranking signal. Thousands of sites publish one. Nobody has shown an answer that demonstrably changed because of it.

So treat it as an option on the future, not a lever today. The convention is cheap, it is machine-readable, and if adoption lands, the sites that already have a clean file are a day's work ahead of everyone else. That is worth a Friday afternoon. It is not worth a sprint, and it is not worth doing before the three things engines demonstrably do read.

The stakes are not hypothetical. G2's 2026 report found 71 percent of B2B buyers now rely on AI chatbots for software research. That shortlist conversation is already happening inside the engines. Whether your name appears in it depends on fundamentals, not on a file nobody has confirmed reading.

What actually gets a B2B SaaS site cited

Crawler access, 30 percent. If GPTBot, ClaudeBot, PerplexityBot, and Google-Extended are blocked in robots.txt, there is nothing to quote. Same outcome if your content only exists after client-side JavaScript runs: a marketing site on a client-rendered framework serves an empty div to a crawler that does not execute scripts. This is the most common fatal issue we see, and it is almost always accidental. A blanket robots.txt rule somebody added during a staging push and never removed. A rendering default nobody revisited after launch. The site looks perfect in a browser and is invisible to the thing you are trying to win.

Content structure, 30 percent. Question-shaped headings and self-contained answers. A paragraph that only makes sense after reading the two above it cannot be lifted into a generated answer, and engines lift. Lead each section with the complete claim, then support it. We went deeper on that structural difference in answer-first content versus traditional SEO.

Structured data, 25 percent. Valid JSON-LD for Organization, Product, Article, and FAQPage tells an engine what your business is in a format it does not have to infer from prose. Broken schema fails quietly. Nothing errors, nothing alerts, the markup just stops being usable. Our site audit crawl flags it alongside the rest of the blockers: accidental noindex, robots.txt over-blocking, canonical conflicts, sitemap-versus-crawl mismatches. Crawlability and schema are the AEO layer of an audit, not generic hygiene.

If you want the mechanics underneath all three, we broke down how an AI answer engine actually decides which brand to cite.

What to do this week

Check the heavyweights first. Then add the file.

• Pull your robots.txt. Confirm GPTBot, ClaudeBot, PerplexityBot, and Google-Extended are allowed. Two minutes.
• View source on your best blog post. Search the raw HTML for a sentence from the middle of the body. If it is not there, your content is JavaScript-only and most crawlers never see it.
• Validate your JSON-LD. Organization and Article at minimum, FAQPage where the page genuinely answers questions.
• Then generate an llms.txt and ship it with the rest of the work, not instead of it.

Our AI Readiness check runs all four dimensions against your URL and returns a prioritized fix list, free, no account needed. That is one snapshot. Clark runs it continuously, and ties the leads that follow back to the exact engine and post that earned them.

Everyone will tell you to add llms.txt. Add it. Then go prove which pages actually pay.