To make a help center readable by AI, do four things: let AI crawlers in with robots.txt, publish an llms.txt that lists your articles, serve each article as clean Markdown, and keep articles answer-first with structured data. None of it replaces good writing. It removes the reasons a machine cannot read what you wrote.

We run all four on our own help center, and every file and number below was checked against the live site on 5 November 2026.

Why bother

More people now ask an assistant "how do I connect GitHub to FeatureOS?" before they open a help center. If the assistant cannot read your docs cleanly, it answers from guesses or from a competitor's page.

One honest caveat. Google's documentation says AI Overviews and AI Mode need no special files, because they run on normal Search. So these steps do not guarantee a spot in Google's AI answers. They matter most for assistants and agents that fetch pages directly, and they cost very little.

1. Let the crawlers in

Check robots.txt first. A single rule blocking an AI user agent makes everything else pointless. Decide separately about each type of bot: search bots that find pages, bots that fetch a page when a user asks, and bots that collect training data. Blocking one does not block the others.

This is the block at the top of our help center's robots.txt, with the comment shortened:

User-agent: *
Allow: /
 
# AI crawlers and assistants are explicitly welcome. A clean Markdown copy of
# every page is available via /llms.txt and the .md URL suffix.
User-agent: GPTBot
User-agent: OAI-SearchBot
User-agent: ChatGPT-User
User-agent: ClaudeBot
User-agent: Claude-User
User-agent: anthropic-ai
User-agent: PerplexityBot
User-agent: Perplexity-User
User-agent: CCBot
User-agent: Google-Extended
User-agent: Applebot-Extended
Allow: /

Stacking several User-agent lines above one Allow applies the rule to all of them. If you want to opt out of model training, remove the training bots from that list and give them their own Disallow: /. Do that without touching the search and fetch bots.

2. Publish an llms.txt

llms.txt is a Markdown file at your site root that tells AI systems what you have and where it lives. Ours is at help.featureos.app/llms.txt. It starts with the site name, then lists every collection as a heading and every article under it as a link to its Markdown copy:

# FeatureOS Help Center
 
## Collections
 
### [Setting up FeatureOS](https://help.featureos.app/en/c/setting-up-featureos.md)
- [Brand Customization](https://help.featureos.app/en/articles/brand-customization.md)
- [Mapping custom domain to FeatureOS](https://help.featureos.app/en/articles/mapping-custom-domain-to-featureos.md)

Keep it simple:

  • A one-line description of the site.
  • Collections as headings, articles as links.
  • Link to the Markdown copy, not the HTML page, so the assistant fetches the clean version.
  • A line of description per link only if the title is not clear.

It is not a ranking trick, and no major engine has confirmed it boosts anything. Think of it as a sitemap a language model can read in one fetch.

3. Serve Markdown, not just HTML

A help article in HTML carries navigation, scripts and styling that an assistant has to strip out. Markdown is the same content without the noise.

On our help center, add .md to any article URL:

help.featureos.app/en/articles/what-is-featureos.md

It returns text/markdown. For the "What is FeatureOS?" article, the HTML page is about 47,900 bytes. The Markdown copy is 1,760 bytes, about 96% smaller. Most of that gap is scripts and markup an assistant would otherwise download and discard. It is one article measured once, so read it as an example, not a benchmark.

Generate the Markdown from the same source as the HTML. Two hand-maintained versions of every article will drift apart.

4. Write answer-first, and add structure

Assistants lift passages, not pages. So:

  • Start each article with the answer in the first two sentences. (How to write great help center articles covers the rest.)
  • Use headings that match how people ask the question.
  • Keep one idea per paragraph, with steps as numbered lists.
  • Put the date on it. Stale docs lose to fresh ones.

Add structured data as well. Our articles carry Article, BreadcrumbList and Organization markup, with the author as a Person. Nothing requires it, but it tells a machine who wrote the page and when.

Check it works

Run these two commands against your own site:

curl -s https://your-help-center.com/llms.txt | head
curl -sI https://your-help-center.com/article-slug.md | grep -i content-type

The first should print your Markdown index. The second should print content-type: text/markdown. If you get text/html, your Markdown route is not working.

Then ask ChatGPT, Claude and Perplexity five real questions about your product, and write down whether they cite you. Run each more than once, because answers change run to run.

If you are also writing the articles your own support AI answers from, that is a different job. We cover it in How to optimize your knowledge base for AI.

Related: FeatureOS knowledge base and keeping your knowledge base updated with Claude Code.