Optimise your website for ChatGPT, Perplexity and Claude: a practical 2026 tutorial

llms.txt is a community proposal for a Markdown file, not the official equivalent of robots.txt. Google states that no special file is needed for its AI features. If you test it, present it as an optional documentation layer.

Create an `llms.txt` file at the root of your domain (accessible at `https://votredomaine.com/llms.txt`) with this sample content:

# [Votre marque]

> Description courte de votre activité en 1-2 lignes (qui, quoi, pour qui, où).

## Services principaux

- [Service 1] : courte description
- [Service 2] : courte description
- [Service 3] : courte description

## Expertise distinctive

[2-3 paragraphes sur ce qui vous différencie]

## Cas clients documentés

- [Client 1] : problème, intervention et preuve consultable
- [Client 2] : problème, intervention et preuve consultable

## Coordonnées

- Site : https://votredomaine.com
- Email : contact@votredomaine.com
- Téléphone : +32...
- Adresse : Bruxelles, Belgique
- BCE : BE 0XXX.XXX.XXX

An extended version is useful only if it stays up to date and does not duplicate contradictory information. Refer to our experimental example, without inferring any effect on indexing or citation.

Add an appropriate Organization block on the page that clearly represents the business. Properties must match visible, verifiable information:

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "[Votre marque]",
  "url": "https://votredomaine.com",
  "logo": "https://votredomaine.com/logo.svg",
  "description": "[Description factuelle de l'organisation]",
  "founder": {
    "@type": "Person",
    "name": "[Nom fondateur]"
  },
  "foundingDate": "2018",
  "address": {
    "@type": "PostalAddress",
    "streetAddress": "[Adresse]",
    "postalCode": "1180",
    "addressLocality": "Bruxelles",
    "addressCountry": "BE"
  },
  "sameAs": [
    "https://www.linkedin.com/company/[votre-marque]",
    "https://www.wikidata.org/wiki/[Votre-Q-ID]",
    "https://www.facebook.com/[votre-marque]"
  ]
}
</script>

The `sameAs` attribute can link official profiles actually controlled by the organisation. Do not add a Wikidata entry or third-party platform listing that does not exist or whose eligibility has not been established.

Choose types that actually describe the page: Article for editorial content, Service for a service offering or BreadcrumbList for breadcrumbs. When a page contains a genuine visible FAQ, FAQPage can describe its questions and answers:

Since 7 May 2026, Google no longer displays FAQ rich results, as its updates log explains. FAQPage markup remains a descriptive vocabulary, with no Google display benefit to promise. You should not create an FAQ or JSON-LD solely to seek a citation.

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [
    {
      "@type": "Question",
      "name": "Combien coûte [votre service] ?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "[Réponse visible, complète et sourcée]"
      }
    },
    {
      "@type": "Question",
      "name": "[Question 2]",
      "acceptedAnswer": { "@type": "Answer", "text": "[Réponse]" }
    }
  ]
}
</script>

An answer should be as short as possible and as long as necessary. No word count guarantees extraction. Add the context, limitations and sources the reader needs.

Check your root `robots.txt` and decide what you want to allow. OpenAI distinguishes, among others, its search bot and training bot. Perplexity also documents several types of access. Here is a sample policy, not a universal recommendation:

# Exemple de politique, à adapter à vos objectifs
User-agent: OAI-SearchBot
Allow: /

User-agent: PerplexityBot
Allow: /

User-agent: GPTBot
Disallow: /

Also check the firewall, CDN and server logs. An `Allow` rule makes access possible but guarantees neither a bot visit, indexing nor citation. User-initiated access may also follow different rules.

Length is not a goal. Content should answer the intent, distinguish facts from assumptions and allow readers to verify important information:

• Scope: a clear question and an answer covering useful nuances
• Structure: descriptive headings, lists or tables only when they make reading easier
• Sources: primary references placed near the claims they support
• Identified author: verifiable identity and expertise when the subject requires it
• Updates: `dateModified` updated with each major revision

Focus effort on usefulness and maintenance. Several pages can be relevant if they answer distinct intentions. A long page is not inherently better.

Define representative questions, then test them in ChatGPT and Perplexity. Keep the date, language, context, answer and displayed references. Repeat the same protocol before concluding that something has changed.

No timeframe or stability can be guaranteed. Platforms and their sources evolve, and an answer can change between tests. For a closer look at Perplexity specifically, read how to check your presence in Perplexity.

To go further: see the AEO / GEO / LLMO hub, which distinguishes facts, checks and hypotheses, or request an LLMO audit.