llms.txt: We Shipped It Anyway — Here's What It Actually Does

We just added an llms.txt file to zenmasterworks.com. Before you assume that means we think it'll get us cited more by ChatGPT or Claude, here's the actual data — and it's less flattering to the file than most of the guides selling it to you.

What llms.txt actually is

It's a plain markdown file at the root of a domain — /llms.txt — listing a site's most important pages with one-line descriptions, proposed by developer Jeremy Howard in September 2024 as a way to help AI systems find the content that matters without wading through a full site's HTML. It's not a search-engine standard. It's not affiliated with any AI company. It's a convention, maintained by a small open-source project, that any site is free to publish or ignore.

What the 2026 data actually says

We didn't want to publish a page telling you this file will get your brand "seen by AI" without checking whether that's true. It mostly isn't — at least not yet, and not for the reason most of the marketing around it implies.

• A study across roughly 300,000 domains found llms.txt adoption sitting at about 10% — and among the fifty most AI-cited domains on the web, only one had the file at all.
• Google's own Gary Illyes and John Mueller have said on the record that Google doesn't read it and isn't planning to, comparing it directly to the old meta-keywords tag — a signal search engines stopped trusting decades ago because sites gamed it.
• Server-log studies of real AI crawler traffic — GPTBot, ClaudeBot, PerplexityBot, Google-Extended — show those bots overwhelmingly skip the file and crawl HTML directly, the same way they always have.
• No major AI provider — OpenAI, Anthropic, Google, or Meta — has publicly committed to reading or acting on llms.txt in a production search or answer system.

Where it does get real, routine use is a narrower lane than most of the SEO advice admits: AI coding assistants — Cursor, GitHub Copilot, Claude Code — fetch it when a developer points them at a documentation site, so the agent can pull the right reference pages without burning tokens parsing an entire docs tree. That's a real, working use case. It's also not ours — we're not a documentation site, and nobody's pointing a coding agent at zenmasterworks.com to integrate an API.

Why we shipped one anyway

Because the cost is a single static file, and because "verify everything, publish it plainly" is the whole premise of this studio. We're not going to claim llms.txt will get us cited more — the evidence says it probably won't, not this year. But it costs us nothing to publish a clean, curated map of our highest-signal pages, and on the small chance a major provider does start reading it, we'd rather already be there than scrambling to catch up.

• ClaimReality, per 2026 data
• "llms.txt boosts AI search citations"No measured correlation in independent studies
• "Google reads it for ranking"Google has said, on the record, it doesn't
• "Major AI crawlers fetch it"Server logs show they mostly don't
• "It helps coding agents use your docs"True — genuinely, for documentation sites
• "Costs nothing to publish"Also true — which is the actual reason we shipped one

If a claim about our own work wouldn't survive us checking it, we don't publish it. That standard doesn't stop at our PageSpeed scores — it applies to the tools we adopt, too, including this one.