llms.txt and structured data: what really helps AI search?

For Google you don't need an llms.txt, or any special structured data: Google Search does not use such files, and an llms.txt neither helps nor harms your visibility ^[1]. None of ChatGPT, Claude or Perplexity officially confirms that they read the llms.txt of other sites. Structured data remains useful, but as hygiene, not as a lever.

The question often comes up after a LinkedIn post or a new button in an SEO plugin. If you hire an SEO agency, you can expect an honest answer, with sources.

What Google, the AI labs and the inventor say

The AI labs do publish an llms.txt themselves, for their developer documentation ^[9], ^[8]. That is not the same as saying their search functions use the one on your site. And the Chrome audit says nothing about search visibility: it belongs to the Lighthouse category "agentic browsing", which Chrome itself calls experimental ^[11].

Ahrefs looked at how it works in practice: 28% of the domains in its data published an llms.txt, and 97% of those received not a single request for it in May 2026 ^[12]. Ahrefs also found no major provider that had formally adopted it ^[12]. So articles claiming that GPTBot or ClaudeBot "regularly read" your llms.txt are not backed by the makers.

What llms.txt is

On 3 September 2024, Jeremy Howard proposed giving websites a Markdown file /llms.txt, with a short summary of the site and links to the pages a language model should read ^[9]. The file has an H1 with the name (the only required part), a blockquote with a summary and sections with lists of links; a model may skip a section called "Optional" ^[9]. Version 2 from August 2026 adds, among other things, links in HTML or in an HTTP header to the Markdown version of a page ^[10].

This is what a small file looks like, for a made-up company:

# Schrijnwerkerij Peeters

> Schrijnwerkerij Peeters installs wooden windows and doors in the Leuven region,
> for homeowners and contractors.

## Services
- [Wooden windows](https://www.example.be/windows/): wood types, prices per m², installation
- [Front doors](https://www.example.be/doors/): models, insulation values, lead time

## Optional
- [Projects](https://www.example.be/projects/)

When llms.txt does make sense

• Software documentation. If you have an API or a software package, an llms.txt and Markdown versions of your pages make things easier for coding assistants. That is also the use the inventor sees ^[9].
• Large sites with a lot of deep content, where a summary quickly takes an agent to the right page.
• An SME site of twenty pages: low priority. If you have it generated automatically, as Yoast does for free in WordPress ^[13], it costs little. An outdated file with dead links helps nobody.

Google says so literally: you can ignore tactics such as cutting your content into "chunks" or creating unnecessary AI text files ^[1].

Structured data: hygiene, not a lever

Structured data tells Google in a fixed format what a page is about; Google recommends JSON-LD ^[14]. It is not required for Google's AI features, and there is no special markup for AI ^[1]. Correct markup doesn't guarantee a rich result either ^[15].

Structured data does make you eligible for rich results ^[1], and it makes basic facts about your business unambiguous: name, legal name, address, VAT number and profiles elsewhere ^[16]. Google calls the VAT number an important trust signal, because users can look it up in a public register ^[16]. The condition: your markup faithfully reflects what is visible on the page ^[15].

FAQ schema no longer works. From August 2023 Google only showed FAQ rich results for authoritative government and health sites ^[17], and since 7 May 2026 it no longer shows them at all ^[3]. That markup does no harm, but it has no visible effect ^[17]. An FAQ in plain text is still useful, for visitors and for AI answers.

Which structured data an SME needs

An Organization block for our made-up company, with a fictitious VAT number:

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "Schrijnwerkerij Peeters",
  "legalName": "Schrijnwerkerij Peeters bv",
  "url": "https://www.example.be/",
  "vatID": "BE0000000000",
  "address": {
    "@type": "PostalAddress",
    "streetAddress": "Voorbeeldstraat 1",
    "postalCode": "3000",
    "addressLocality": "Leuven",
    "addressCountry": "BE"
  },
  "sameAs": ["https://www.linkedin.com/company/example"]
}
</script>

If you work locally, with a shop or workshop, LocalBusiness is the better type, together with a well-completed Business Profile. Benjamin explains how that fits together in local SEO.

You can also add it via Google Tag Manager, because Google reads JSON-LD that is added to the page with JavaScript ^[18]. But then the risk grows that your markup diverges from the page ^[18], so preferably put it directly in the HTML. Cédric explains how GTM works in Google Tag Manager explained.

Test your markup with Google's Rich Results Test and the Schema Markup Validator from schema.org.

What achieves more than either

1. Allow the right crawlers, such as OAI-SearchBot and PerplexityBot, including in your firewall or CDN.
2. Text in the HTML: prices, services and contact details as plain text.
3. One question per page, with the answer at the top and figures with a source.
4. Mentions elsewhere, in places where your customers read.

The step-by-step plan is in how to get found in ChatGPT. How AI search engines then choose between candidates is covered in how ChatGPT, Gemini and Perplexity choose their sources.

Want to know whether your site has the basics in order, from structured data to your visibility in ChatGPT, Gemini, Perplexity and Google AI Overviews? The free scan looks at it, and you get the findings and a conversation.