Markdown for AI Search: llms.txt Best Practices & GEO

Research checked October 22, 2026 · For website owners, marketers and WordPress publishers

Markdown for AI is a way to publish your website’s information in a clean, structured format that assistants can retrieve and use. Businesses are publishing Markdown versions of pages, linking them through llms.txt, and sometimes delivering Markdown directly to recognized agents. Their aim is to make products, services and expertise easier to discover, describe, compare and recommend.

That is a real strategy to test. It is also more involved than uploading one text file and expecting to rank in ChatGPT. This guide explains the different implementations, what companies have actually tried, what the evidence supports, and how to build a useful pilot on WordPress.

View the infographic: The Markdown GEO PlaybookEight illustrated steps from your website to an AI answer

What is Markdown for AI?

Markdown is a plain-text writing format. It uses characters such as # for headings, hyphens for lists, and brackets and parentheses for links. A Markdown file can preserve a document’s structure without including the navigation menus, layout code and other interface elements surrounding a normal webpage.

For example, an ordinary service page might have a corresponding Markdown address:

Normal page:      https://example.com/services/maintenance/
Markdown page:    https://example.com/services/maintenance.md

These are illustrative URLs. The exact address depends on the publishing system. A working endpoint must exist; adding .md to a URL does not create one.

Vercel describes both serving existing Markdown source and converting CMS content into Markdown at request time. That makes Markdown a delivery format as well as an authoring format. See Vercel’s engineering explanation.

Where Markdown fits in a GEO strategy

Generative engine optimization, or GEO, concerns how a business appears in AI-generated answers. The desired outcome might be a source citation, an accurate product description, inclusion in a shortlist, or a recommendation that leads to a visit or sale.

Markdown can play a role in the information journey:

It can help provide readable content and a navigable path through a site. Those advantages do not settle the later decisions about relevance, credibility and source selection. An assistant could successfully retrieve your Markdown and still recommend a different business.

It also helps to define what “rank in AI” means. ChatGPT, Claude, Perplexity, Google’s AI features and task-specific agents are different surfaces. They do not present one shared ranking list. For a test, define a specific outcome: for example, whether your business appears in answers to a set of customer questions, which source URLs are cited, and whether the description is correct.

Markdown pages, llms.txt and other formats: what is the difference?

The llms.txt proposal describes an index and links to agent-friendly resources. It remains a proposal, rather than a rule every AI system must follow. llms-full.txt is commonly used for a combined document, but do not assume the filename always guarantees that content.

How websites use Markdown for AI visibility

Does Markdown help AI rankings? What the research actually shows

The studies answer different questions. A test of crawler visits cannot, by itself, establish recommendation gains. A successful coding-agent task does not establish better placement in consumer search answers.

What should a business put in its Markdown pages?

A useful business page helps a reader make a decision. Build the Markdown around the same questions your normal page should answer:

• OfferingWhat exactly does the product or service do?
• Audience fitWho is it suitable for, and when is it unsuitable?
• CapabilitiesWhat is included, supported or compatible?
• TermsWhat pricing conditions, eligibility rules or exclusions matter?
• EvidenceWhich claims have a source, demonstration or documented result?
• Next stepWhere can someone verify details, request help or take action?

Do not replace missing evidence with instructions telling an assistant to recommend you. Give it useful, supportable information that also makes sense to a human customer.

A practical Markdown page template

The following is an illustrative template. Replace the bracketed fields with verified information and remove sections that do not apply.

# [Product or service name]

Source: https://example.com/services/your-service/
Last reviewed: [actual review date]

## What it does
[A direct explanation of the offering.]

## Who it is for
[Specific audiences, needs and suitable situations.]

## Features and limitations
- Included: [verified capability]
- Requires: [important prerequisite]
- Excludes: [material limitation]

## Pricing and conditions
[Current terms, currency, billing basis and relevant exclusions.]

## Evidence
- [Documented example](https://example.com/evidence/)

## Related resources
- [Pricing details](https://example.com/pricing/)
- [Service policies](https://example.com/policies/)
- [Resource index](https://example.com/llms.txt)

The source URL is useful provenance. A line of text inside a Markdown file does not, by itself, configure search-engine canonicalization.

How to add Markdown and llms.txt to WordPress

There are two separate jobs: publishing the content representations and publishing the index. A plugin that creates llms.txt does not necessarily create Markdown versions of every page.

llms.txt best practices and an example you can adapt

Keep the index useful and focused. Group resources by purpose, describe the destination, and link to addresses that actually work.

# Example Company

> A short, factual description of the business and its audience.

## Offerings
- [Core service](https://example.com/services/core.md): Scope and exclusions.
- [Product guide](https://example.com/product/guide.md): Capabilities and requirements.

## Buyer information
- [Pricing](https://example.com/pricing.md): Plans and conditions.
- [Policies](https://example.com/policies.md): Published service terms.

## Evidence and support
- [Documented examples](https://example.com/examples.md): Supporting evidence.
- [FAQ](https://example.com/faq.md): Specific customer questions.

This is an example, not a list of live resources. Replace every destination before publishing.

• ✓Use descriptive link text instead of generic labels such as “click here.”
• ✓Include concise context that helps a reader choose the right page.
• ✓Remove dead URLs and resources that no longer describe the business accurately.
• ✓Automate updates where possible, with an editorial review of the output.
• ✓Use product- or topic-specific collections when a single bundle becomes unwieldy.

The goal is an understandable route to information. It does not need to reproduce every URL in your XML sitemap.

Technical checks before launch

Check the response, not just the address

A URL ending in .md could return an error, an HTML wrapper or an unrelated page. Confirm a successful response and inspect the body. If you support negotiation, test it explicitly:

# Test a dedicated Markdown resource
curl -i https://example.com/services/core.md

# Test Markdown at the ordinary page URL
curl -i -H "Accept: text/markdown" https://example.com/services/core/

These commands are inspection examples for your own implementation. For negotiated Markdown, verify the expected Content-Type and appropriate cache variation, commonly Vary: Accept.

Advertise alternatives accurately

The current llms.txt proposal recommends alternate-format and index relationships. For example, a developer can add the following to the HTML head, using your real URLs:

<link rel="alternate" type="text/markdown"
      href="https://example.com/services/core.md">
<link rel="describedby"
      href="https://example.com/llms.txt">

Equivalent discovery information can also be supplied through HTTP link headers. A discovery hint helps a compatible client find a resource; it does not force a crawler to fetch it.

Separate search access from training access

Classify the traffic you measure. For example, OpenAI documents different roles for OAI-SearchBot, GPTBot and ChatGPT-User: search support, potential training collection and user-triggered retrieval. A training crawler visit does not demonstrate inclusion in a current answer. See the official crawler documentation.

Review existing crawl permissions and access controls alongside the new endpoints. Publishing an index does not override a firewall, a login requirement or crawler restrictions.

How to measure AI citations, recommendations and business impact

Write down what success would look like before deployment. “More AI traffic” is too imprecise if the intended result is appearing in a product shortlist.

Choose customer questions such as “What options suit [specific need]?” and “Does [product] support [requirement]?” Include category questions without the brand name as well as branded verification questions. Do not use only prompts that tell the assistant which business to mention.

Repeat the same question set, record the platform and date, and preserve answer transcripts. Where practical, compare matched page groups while keeping content, audience targeting and other major changes consistent. If you change the format, content and bot access simultaneously, you cannot isolate the format’s contribution.

Common mistakes that make the strategy less useful

• ✕Publishing an index without useful destinations. A map cannot compensate for missing or inaccessible content.
• ✕Stripping away important context. Exclusions, table relationships and citations matter as much as the headline claims.
• ✕Letting the versions disagree. Keep prices, features and policies synchronized.
• ✕Making one enormous content dump. Offer manageable, clearly described resources instead.
• ✕Confusing requests with results. Separate retrieval, attribution, recommendation and conversion.
• ✕Generalizing from one AI platform. Record what each tested system actually does.
• ✕Forgetting ordinary publishing quality. Clear explanations, accurate claims and working links remain essential.

Frequently asked questions about Markdown for AI and GEO

A practical launch checklist

1. 1Select a small set of useful, decision-relevant pages.
2. 2Correct the public information before converting it.
3. 3Record a baseline and a repeatable customer-question set before launch.
4. 4Publish working Markdown representations with consistent facts.
5. 5Preserve navigation and provide a concise resource index.
6. 6Verify access, content, headers and caching.
7. 7Compare recommendations, citations, accuracy and business actions over time.