If you have sat through a conference session on AI and tourism in the past year, someone has told you to “add an llms.txt file”. This article explains what that file is, where it came from, who reads it today, and how to think about it as one part of making a destination website legible to AI, not as a fix on its own.

What it is

llms.txt is a plain markdown file that sits at the root of a website, at /llms.txt, in the same way robots.txt does. It was proposed on 3 September 2024 by Jeremy Howard of Answer.AI (Answer.AI, September 2024⁠ (opens in a new tab)), and the specification lives at llmstxt.org⁠ (opens in a new tab), where a second version was published in August 2026.

The idea is modest. A language model answering a question has a limited window of attention and no patience for navigation menus, cookie banners and script tags. A sitemap lists every address but says nothing about what matters. llms.txt is a curated map written for a machine reader: here is what this site is, here are the pages that matter most, and here is what each one contains.

The format is deliberately strict so that it can be parsed reliably:

  • an H1 with the name of the site or organisation, the only required element;
  • a blockquote with a one-paragraph summary;
  • optional short paragraphs of context;
  • H2 sections, each a list of links with a one-line description per link;
  • an “Optional” section at the end for material a reader can skip.

The proposal also recommends publishing a clean markdown version of each important page, reachable by adding .md to the page address, so that an agent following a link gets content, not markup.

Who reads it today, honestly

In June 2025, Google’s John Mueller wrote on Bluesky: “FWIW no AI system currently uses llms.txt.” (Search Engine Roundtable, June 2025⁠ (opens in a new tab)). The large consumer assistants fetch pages, not the map.

Two things follow. First, llms.txt is not a ranking lever, and anyone who sells it as “get cited by ChatGPT” is overselling. Second, it is close to free, carries no downside, and is read by a growing population of agents and tools that fetch sites on demand instead of crawling them: the assistants people point at a page, research and coding agents, and the tools that let a model work directly with a site’s own content. Developer documentation sites adopted it first for exactly that reason.

The position we recommend to destination teams: publish it, keep it accurate, measure who fetches it, and don’t build the strategy on it.

What matters more than the file

If an assistant fetches your pages and not your map, then the pages are what count. Four things have more effect than llms.txt.

Readable pages. Clean HTML with real headings, short paragraphs and facts stated in sentences. In February 2026, Cloudflare launched Markdown for Agents, which converts a page to markdown on the fly when an agent requests it with an Accept: text/markdown header (Cloudflare changelog, 12 February 2026⁠ (opens in a new tab)). Content negotiation of that kind, serving markdown to agents and HTML to people from the same address, is becoming the practical pattern.

Structured data. Marking up question-and-answer content with schema.org’s FAQPage type (schema.org⁠ (opens in a new tab)) gives a machine the question, the answer and the boundary between them. Be clear about what this does and doesn’t do: Google stopped showing FAQ rich results in its search results in 2026 (Google Search Central⁠ (opens in a new tab)), so the payoff is machine readability, not a decorated search result.

Crawler permissions. Your robots.txt decides which agents may read you, and the agents have split into two kinds. OpenAI’s GPTBot gathers training data while OAI-SearchBot surfaces sites in ChatGPT search (OpenAI⁠ (opens in a new tab)). Anthropic separates ClaudeBot, for training, from Claude-SearchBot and Claude-User, for search and on-demand fetching (Anthropic⁠ (opens in a new tab)). PerplexityBot indexes for Perplexity’s search and respects robots.txt (Perplexity⁠ (opens in a new tab)). Google-Extended controls whether content is used to train Gemini models, and Google says it has no effect on Search inclusion or ranking (Google Search Central⁠ (opens in a new tab)). Many sites block the training agents and allow the search agents. That is a defensible policy. Blocking everything by default, as some security configurations do, quietly removes a destination from AI answers.

Current facts. None of the above helps if the assistant reads an Operator’s opening hours from three summers ago. Currency is the quality signal both people and machines reward, and it is the one a destination team controls most directly.

What a good llms.txt looks like for a destination

A destination’s file should read the way a well-briefed visitor information centre volunteer talks: specific, current, organised around the questions people ask. For a regional destination it might look like this.

# Visit [Region]

> The official guide to [Region], [distance] from [capital]: [two or three things it is known for], across [N] towns.

[One short paragraph: who publishes the site, how often Operator details update, what the site covers.]

## Towns and areas
- [Town A](/towns/town-a.md): [one line on what it is for: markets, galleries, the start of the rail trail]
- [Town B](/towns/town-b.md): [one line]

## Things to do
- [Walks and trails](/see-and-do/walks.md): grades, distances, seasonal closures
- [Cellar doors](/see-and-do/wine.md): [N] Operators, opening days, which have restaurants

## Where to stay
- [Accommodation](/stay.md): [N] Operator Products, updated daily from ATDW

## Events
- [What's on](/events.md): dates, venues, booking links

## Plan your trip
- [Getting here](/plan/getting-here.md): drive times, train and coach options
- [Visitor information centres](/plan/visitor-centres.md): locations and hours

## Optional
- [About the organisation](/about.md)
Example llms.txt

Three habits make it useful. Describe every link in a line that would help a stranger choose it. Point only to pages that are themselves readable as text. And regenerate it whenever content changes: a hand-written file that still lists a trail closed since autumn does more harm than no file at all.

On Roam

On Roam, this is part of AI Discovery, which is live on every Roam site at no extra cost. Each site automatically generates and maintains its llms.txt file, and when an AI agent visits, it receives clean, structured markdown instead of raw HTML. Improvements roll out platform-wide, with no developer dependency. See how AI Discovery works.

A short checklist

  1. Does /llms.txt exist, parse cleanly, and say what the site is for in one paragraph?
  2. Does every link in it resolve to a page that reads well as plain text?
  3. Does robots.txt allow the search and on-demand agents you want to be found by?
  4. Are your most-asked planning questions answered in question-and-answer form, with structured data?
  5. Who regenerates the file when a town page, a trail or an Operator Product changes?

For the wider picture, including how AI answers are changing the traveller’s path to a destination, read How AI search is changing the way travellers find destinations.

Ask us to check how your current site reads to AI: book a demo and we will walk through it with your destination’s own pages.