Check if AI crawlers can read your site

My website stripe.com scored 75/100 on Good for Bots (Good). Good for Bots measures one thing: whether a language model can read a site and cite it.

I need you to: serve markdown through content negotiation on all public HTML routes.

The check it fails is "Serves markdown when asked for it", worth 10 points. What the scanner saw: None of the 4 pages read pass. Asking the homepage for markdown still returns text/html.

The scan checked a sample of 4 pages. It does not establish which other pages have the same issues. Use the findings below to locate the responsible templates, components or configuration, and apply the fix wherever the same issue occurs. Pages in the sample with findings:
- https://stripe.com/en-de: Asking the homepage for markdown still returns text/html.
- https://stripe.com/en-de/blog/introducing-our-agentic-commerce-solutions: Asking the page at /en-de/blog/introducing-our-agentic-commerce-solutions for markdown still returns text/html.
- https://stripe.com/en-de/pricing: Asking the page at /en-de/pricing for markdown still returns text/html.
- https://stripe.com/en-de/products: Asking the page at /en-de/products for markdown still returns text/html.

Implement content negotiation across all public HTML routes, including page types
outside the scan's sample. Put it in the shared response handling or the handlers
for each page type, so new pages receive it too. Agents need the content of any
page they follow, including pages the report did not name.

When a request arrives with `text/markdown` in its `Accept` header, the same URL that
serves HTML to a browser should return a markdown rendering of that page's content.
Requests without that header must keep getting the HTML exactly as they do now.

The response must:

- set `Content-Type: text/markdown; charset=utf-8`;
- include `Accept` in the `Vary` header, so shared caches never serve the markdown to a
  browser or the HTML to an agent (RFC 9110 says a server SHOULD send `Vary` whenever
  the representation depends on a request header);
- contain real markdown, not HTML with a different header on it;
- drop `ETag` and `Last-Modified` for the converted response, since a conditional request
  cannot be honoured against a body generated on the fly.

Work out how to do this in whatever stack this project uses. If the site is behind
Cloudflare on a paid plan, the "Markdown for Agents" zone setting does all of the above
and is the least invasive option. Otherwise implement it in the framework: middleware
inspecting the `Accept` header, a route handler that returns the markdown source, or
whatever fits the architecture here.

Strip navigation, headers, footers and cookie banners from the markdown. The point is the
content, without the page furniture around it.

Verify HTML and Markdown responses for each public page type, including URLs
outside the reported sample. Check that each Markdown response contains that
page's content and that ordinary browser requests still receive HTML.

The whole report, as markdown, is at https://goodforbots.com/r/stripe.com.md?scan=cmvque06j000101rvr6x8jj3t. Read it first: this prompt covers one finding, and the report has everything the scan saw.

When you are done, summarise what you changed and how to check it against a running instance. Do not try to run the Good for Bots scan yourself: it only sees what is already deployed, it is rate-limited, and re-running it is the site owner's call once the change ships.

My website stripe.com scored 75/100 on Good for Bots (Good). Good for Bots measures one thing: whether a language model can read a site and cite it.

I need you to: state a position on AI crawlers in robots.txt.

The check it fails is "States a position on AI crawlers", worth 5 points. What the scanner saw: robots.txt names no AI crawlers, so every operator has to guess.

A `robots.txt` that says nothing about AI crawlers leaves every operator to guess what
this site wants, and they guess differently. The point of this task is to state a
position, whichever position that is.

The thing most people get wrong: the major operators run **separate crawlers for
separate purposes**, and blocking one does not block the others. OpenAI runs GPTBot for
training, OAI-SearchBot for search indexing, and ChatGPT-User for pages a user asks
about. Anthropic runs ClaudeBot, Claude-SearchBot and Claude-User for the same three
jobs. Decide about each job, not about the vendor.

A common and reasonable position is to refuse training while staying visible in AI
answers:

```
# Training: no
User-agent: GPTBot
Disallow: /

User-agent: ClaudeBot
Disallow: /

User-agent: Google-Extended
Disallow: /

User-agent: CCBot
Disallow: /

# Search and user-requested fetches: yes
User-agent: OAI-SearchBot
Allow: /

User-agent: Claude-SearchBot
Allow: /

User-agent: PerplexityBot
Allow: /

User-agent: ChatGPT-User
Allow: /

User-agent: Claude-User
Allow: /
```

Work out what this project wants before copying that. Allowing everything is
equally valid and earns exactly as much here; what is being measured is whether there is
an explicit answer, not which answer it is.

Blocking a documented search or user-fetch control declares a refusal of that route.
Some user-triggered fetchers may still retrieve a page; do not promise complete invisibility.
Google-Extended also controls grounding, while CCBot publishes a reusable dataset.
Their training opt-outs remain neutral in our score.

Two things about how this is judged. The list of crawler tokens we recognise is a
snapshot of a moving target: operators add crawlers faster than any scanner tracks them.
A rule for a token we have not seen yet is a good rule that earns nothing here until we
catch up. And a position stated in a vocabulary rather than in rules, such as
`Content-Signal`, is a different check; this one is about the `User-agent` groups.

Use the operator sources in the report's crawler catalogue. Unconfirmed or retired names
do not count; aliases for one control count once. One mixed-use control still
counts as one engagement group.

The whole report, as markdown, is at https://goodforbots.com/r/stripe.com.md?scan=cmvque06j000101rvr6x8jj3t. Read it first: this prompt covers one finding, and the report has everything the scan saw.

When you are done, summarise what you changed and how to check it against a running instance. Do not try to run the Good for Bots scan yourself: it only sees what is already deployed, it is rate-limited, and re-running it is the site owner's call once the change ships.

My website stripe.com scored 75/100 on Good for Bots (Good). Good for Bots measures one thing: whether a language model can read a site and cite it.

I need you to: mark the content up so a parser can find it.

The check it only partly passes is "Content is marked up so a reader can find it", with 5 points of it still on the table. What the scanner saw: 2 of 4 pages read pass; 1 partly. The page at /en-de/pricing is barely structured: No <main> element and nothing carrying role="main".

The scan checked a sample of 4 pages. It does not establish which other pages have the same issues. Use the findings below to locate the responsible templates, components or configuration, and apply the fix wherever the same issue occurs. Pages in the sample with findings:
- https://stripe.com/en-de/blog/introducing-our-agentic-commerce-solutions: The page at /en-de/blog/introducing-our-agentic-commerce-solutions is partly structured: No <main> element and nothing carrying role="main".
- https://stripe.com/en-de/pricing: The page at /en-de/pricing is barely structured: No <main> element and nothing carrying role="main".

Review the site's public page templates and shared layout components. Fix missing
content roots, heading structure and landmarks wherever they occur, including
templates outside the reported sample. Check the server's HTML response from each
page type: a shared layout fix can introduce a second `<main>` on pages that
already have one.

Semantic elements help a reader of raw HTML find the page's content and distinguish
it from navigation and other surrounding material.

Three things, all from the HTML standard. Change the markup; do not change what the page
looks like.

**1. One content root.** Wrap the page's dominant content in a single `<main>`.

```html
<body>
  <header>…site header…</header>
  <main>…the article, the docs page, the product copy…</main>
  <footer>…site footer…</footer>
</body>
```

The standard allows exactly one `<main>` per document that is not marked `hidden`, and
its ancestors may only be `<html>`, `<body>`, `<div>`, an unnamed `<form>` or a custom
element, which makes a `<main>` inside `<section>` or `<article>` invalid. Put the real
content inside it: a `<main>` that holds a sliver of the page while the article sits in a
sibling `<div>` marks nothing. If a framework template makes the element impossible,
`role="main"` on the wrapper is an acceptable substitute.

**2. A heading outline that holds together.** Exactly one idea per level, starting at
`<h1>`, and never skipping a level on the way down.

```html
<h1>What this page is about</h1>
  <h2>First section</h2>
    <h3>A detail</h3>
  <h2>Second section</h2>
```

`<h1>` → `<h3>` is invalid; `<h3>` → `<h2>` on the way back up is fine. Use heading
elements for headings and CSS for size: a `<div class="text-2xl">` is invisible to the
outline. If a heading is an image or an icon, give it real text and hide it visually
rather than leaving the element empty.

**3. Landmarks around the content.** At least two of these, outside `<main>`:

```html
<header>   <!-- the site header: becomes the banner landmark -->
<nav>      <!-- navigation -->
<footer>   <!-- the site footer: becomes contentinfo -->
<aside>    <!-- related content -->
<search>   <!-- a search form -->
```

Two rules that surprise people. `<header>` and `<footer>` only count as the page's banner
and contentinfo when they are children of `<body>`; inside `<article>` or `<section>` they
are that section's own header and footer and are not landmarks at all. And a `<form>` is
only a landmark when it has an accessible name, so give a search form
`aria-label="Search"`.

Do not add landmarks that describe nothing: five unlabelled `<nav>` elements are worse
than one. And do not wrap the whole body in `<main>`: that puts the header and footer
inside the content root, which removes the landmarks this asks for.

The report's content and navigation breakdown is diagnostic. It measures extracted
text in marked regions, not content quality. Keep useful navigation; do not remove
links or add filler to change a percentage. If much of the text is labelled "other",
check whether the real main content has a meaningful wrapper. The reported start of
main text follows HTML order, which can differ from the visual layout.

The whole report, as markdown, is at https://goodforbots.com/r/stripe.com.md?scan=cmvque06j000101rvr6x8jj3t. Read it first: this prompt covers one finding, and the report has everything the scan saw.

When you are done, summarise what you changed and how to check it against a running instance. Do not try to run the Good for Bots scan yourself: it only sees what is already deployed, it is rate-limited, and re-running it is the site owner's call once the change ships.