llms.txt and Schema for AI: The Technical GEO Playbook

By Abhiijay Vinayak, co-founder of Beamcite. Last updated August 31, 2026.

TL;DR

llms.txt is a proposed Markdown convention for publishing a curated map of important pages at your domain root, introduced by Answer.AI co-founder Jeremy Howard in September 2024. Schema markup is structured data, usually JSON-LD, that can help Google understand supported page content and make valid pages eligible for supported Search features. Google says no special schema is required for its generative Search features, and Google has said it does not use llms.txt and is not planning to. Neither is a documented AI-citation switch. Beamcite runs a managed 90-Day Visibility Plan for founder-led software companies without a capable in-house organic team, with the technical work scoped alongside research, authority, human-reviewed content, publishing, measurement, and iteration where needed.

Most "technical GEO" advice online is one of two extremes: add an llms.txt file and watch the AI citations roll in, or schema markup is dead, just write good content. Both are wrong. The technical layer is worth maintaining only when each control retains its documented role. This playbook covers the llms.txt proposal, the evidence available for adoption, Google's structured-data guidance, and the boundary between crawl eligibility, supported Search features, and direct AI-citation measurement. Beamcite handles this layer for the companies it works with, and the method below is the same one it uses.

Quick navigation:

• Want the non-technical version of getting cited by AI? Read the full GEO method
• Just need to ship an llms.txt file? Jump to "How to write an llms.txt file" below.
• Comparing the whole technical setup? You are in the right place. Keep reading.

What is llms.txt?

llms.txt is a proposed standard: a single Markdown file at your site's root intended to offer language-model tools a structured, low-noise index of your site. The proposal packages a curated summary, what the site covers, and which pages the publisher considers important in a predictable Markdown format. That intended use is not evidence that a production assistant fetches or relies on the file.

The convention was proposed by Jeremy Howard, co-founder of Answer.AI and fast.ai, on September 3, 2024, and the specification lives at llmstxt.org. The proposal defines two files. The first, /llms.txt, is intended as a navigation aid: an H1 title, a short description, and lists of key pages with one-line descriptions and links. The second, /llms-full.txt, is an optional companion intended to collect the full text of important pages in one Markdown document. Neither format establishes that a production model will request the file.

The idea borrows from robots.txt and sitemap.xml: a predictable file at a predictable location that machines can find without guessing. The difference is what it is for. robots.txt tells crawlers what they may access; llms.txt tries to tell language models what matters and hand it to them in a format that wastes no tokens on navigation, ads, and markup.

The honest state of llms.txt adoption

This is where most guides stop being useful, so this section is blunt. No primary adoption source cited in this article establishes that a major AI or search engine uses llms.txt for ranking or content discovery in production.

Google has been the most direct. At Search Central Live in July 2025, Gary Illyes said Google does not support llms.txt and is not planning to, and John Mueller compared the file to the old meta keywords tag, the tag search engines stopped trusting decades ago because site owners controlled it and gamed it. Mueller's core point was practical: AI crawlers already download the full page, so serving them a separate, self-declared version of your content is both redundant and easy to abuse. A site could show one thing in llms.txt and another to readers, which is exactly the cloaking problem that killed meta keywords.

It is worth being precise about what that does and does not mean. Several companies publish llms.txt files on developer-documentation sites, but publication alone does not show that a production assistant consumes the file or changes an answer because of it. Documentation is the most plausible place to test the convention because the proposal is designed to provide a compact map of technical pages. That is not the same as "add llms.txt and AI will recommend your business," and anyone selling the second claim is ahead of the evidence.

So the file can be low-effort to test when it is generated from current content, but any benefit should be treated as potential rather than confirmed. It is not a documented crawl, indexing, ranking, or citation control. For ChatGPT specifically, prioritize the documented ChatGPT Search controls, including OAI-SearchBot access and crawl verification, before testing an optional llms.txt file.

When llms.txt may be reasonable to test

The proposal is most plausible to test when the maintenance cost is low and the result can be observed. These conditions make a test more practical, not more proven:

• You run documentation-heavy or developer-facing content. API docs, SDK references, and technical guides are a plausible fit for the proposal, but test whether any tools actually request the file rather than assuming they do.
• Your pages are heavy with navigation, scripts, and interactive elements. The proposed Markdown summary may offer a cleaner representation, but it does not replace making the canonical page crawlable and rendering its main content in HTML.
• You already maintain content in Markdown. Automated generation may make the test low-effort and reduce drift between the file and canonical pages.

The payoff remains speculative for a standard marketing or local-business site. If you test the file, generate it from current Markdown or sitemap data, validate the links, and record whether relevant agents request it. A stale file may contradict the canonical pages and would undermine the test.

How to write an llms.txt file

The format is deliberately simple. It is Markdown, it lives at yourdomain.com/llms.txt with no subdirectory, and it follows the structure the spec lays out. A minimal valid file looks like this:

# Acme Analytics

> Acme Analytics is a privacy-first product analytics tool for SaaS teams. This file maps the most useful pages for AI tools.

## Core pages

- [Product overview](https://acme.example/product): what Acme does and who it is for
- [Pricing](https://acme.example/pricing): plans, limits, and what is included
- [Documentation](https://acme.example/docs): setup, SDKs, and the API reference

## Guides

- [Getting started](https://acme.example/docs/start): install and send your first event
- [Self-hosting](https://acme.example/docs/self-host): run Acme on your own infrastructure

## Optional

- [Changelog](https://acme.example/changelog): recent releases

The proposal specifies one H1 with the site or product name, an optional blockquote description, then H2 sections that group links with short notes. It reserves an "Optional" section for lower-priority pages and defines llms-full.txt as an optional document containing page content rather than only links. Following that format does not establish adoption. Validate every URL, serve the file consistently, and keep it synchronized with the canonical pages if you choose to test it.

Schema markup for AI: what actually works

Schema is structured data embedded in a page, usually as JSON-LD, that labels supported content such as an Organization, BlogPosting, Product, or breadcrumb trail. It can help search systems understand a page and can make supported pages eligible for rich results when every field matches the visible content.

Google's current generative Search guidance is more limited than many GEO guides claim. It says there is no special schema required for AI Overviews or AI Mode. Structured data should remain accurate and useful, but it is not a citation switch and it does not prove that an AI assistant will select the page.

The SEO, AEO, and GEO comparison shows where schema fits in the shared technical foundation and where original evidence, internal links, and third-party corroboration do more of the work.

Supported schema types to evaluate

You do not need every type schema.org defines. Evaluate only types that match the visible page and a supported Search use:

Schema type	What it labels	Safe use
Organization	Your business name, logo, URL, and verified profiles	Keep one canonical entity and reuse its @id across page schema
BlogPosting / Article	A post's headline, author, image, and dates	Match the visible byline, cover, published date, and modified date
FAQPage	Visible question-and-answer pairs	Mark up only questions and answers shown on the page; do not expect AI placement
Product	A product's name, description, and supported attributes	Use only current product facts that visitors can verify on the page
BreadcrumbList	The page's position in the site hierarchy	Keep the hierarchy consistent with visible navigation and canonical URLs

Implement these as JSON-LD in the page head, keep every value consistent with what the page actually shows, and validate the markup before you ship it. Schema that claims something the visible page does not say is the structured-data version of cloaking, and it erodes the trust the markup is supposed to build.

llms.txt vs schema: which does what

The two are not competitors, they do different jobs, and the evidence behind them is very different. This is the comparison to keep in mind before you spend time on either.

	llms.txt	Schema markup
What it is	A Markdown index of your site at /llms.txt	Structured data (JSON-LD) embedded in each page
Officially documented AI-search role	None confirmed; Google says it does not use it	Google says no special schema is required for generative Search features
Supported or potential use	Proposed compact map for language-model tools; adoption must be verified	Machine-readable page context and eligibility for supported Google Search features
Effort to maintain	Low if automated, risky if hand-kept	Moderate; per-page, must match visible content
Verdict	Optional hygiene, narrow payoff	Keep supported markup accurate and matched to visible content

The short read: use supported structured data where it accurately describes visible content and serves an ordinary Search feature. Test llms.txt only if it is cheap to automate and you can verify whether it is requested or used. Neither is documented as an AI-citation signal.

What actually earns AI citations

Crawl and index eligibility matter when a system retrieves live web pages, but neither schema nor llms.txt is documented as an AI-citation lever. A Princeton-led paper presented at the KDD 2024 conference tested content interventions across 10,000 queries in 25 domains. In that controlled environment, adding authoritative citations raised visibility by about 30%, statistics by about 32%, and quotations by about 41%, with gains of up to 40% in some settings. Keyword stuffing did not help, and the study did not test schema or llms.txt.

The technical layer can remove avoidable access and consistency problems, but it does not make an engine read, attribute, or cite a page. The controlled study above supports testing source-backed content interventions such as citations, statistics, and quotations. Other practices, including answer-shaped openings, internal links, comparison tables, FAQs, and third-party corroboration, should be treated as operational hypotheses and measured directly. The complete workflow is in the guide to getting your business cited and recommended by AI.

Doing the technical setup yourself vs having it done

The technical pieces are not hard in isolation. The work is in doing all of it, correctly, on every post, and keeping it in sync as the site grows. Here is how the realistic options compare.

Approach	Who maintains the schema and files	Consistency across the whole site	Best for
Do it yourself	You, by hand or with plugins	Depends on discipline; drift is common	Teams with technical capacity and spare time
Raw AI tool	Nobody; markup is generated and shipped unread	Often inconsistent or invalid	Teams who will validate every output themselves
Beamcite	A managed, human-reviewed implementation	Agreed during scoping	Founder-led software companies without a capable in-house organic team

The honest trade-off: a one-time schema setup is manageable, but the markup can drift as pages and business facts change. Beamcite's managed plan can include technical fixes, internal links, human-reviewed content, publishing, measurement, and iteration. The exact output and commercial terms are scoped on the founder call, and no technical setup guarantees a ranking or citation.

Which technical GEO setup does your site need?

• Standard marketing or local-business site: Use only supported schema that matches visible content, keep the canonical pages crawlable, and do not treat llms.txt as an indexing or citation control.
• Documentation-heavy or developer-facing product: Use applicable supported schema and consider an automated llms.txt or llms-full.txt as a test, while verifying actual requests or downstream use where possible.
• Site that publishes content weekly: The bottleneck is consistency, not setup. Use a process (or a service) that ships correct markup on every post, not a one-time configuration.
• No capable in-house organic team: A managed implementation can connect the technical layer with research, authority, human-reviewed content, publishing, measurement, and iteration.

Conclusion

The technical GEO playbook starts with crawl and index eligibility on each surface. Use supported schema accurately and keep it consistent with visible content, but do not treat it as a special AI-ranking or citation signal. llms.txt is an optional proposal with potential documentation use, not a confirmed crawler directive or citation lever. Separately, the Princeton study found that citations, statistics, and quotations improved visibility in its controlled experiments.

Different sites need different amounts of this. A local business should keep supported schema accurate and move on. A developer tool may also find llms.txt useful for documentation. A company publishing regularly needs a process that keeps markup and visible content in sync. Founder-led software companies without a capable in-house organic team can talk to the founder about Beamcite's managed 90-Day Visibility Plan. Exact scope, output, reporting cadence, and commercial terms are agreed on that call.

Frequently asked questions

What is llms.txt in simple terms? llms.txt is a proposed Markdown file at your site's root (yourdomain.com/llms.txt) intended to present a curated list of important pages with short descriptions. Answer.AI's Jeremy Howard proposed it in September 2024 as a cleaner map for language-model tools, but the proposal does not prove that production assistants consume it.

Does llms.txt help you rank or get cited by AI? Not on current evidence. Google has said it does not use llms.txt and is not planning to, and no major AI engine has confirmed using it for ranking or content discovery in production. The proposal may offer a compact map for documentation tools, but publishing the file is not proof that a production assistant consumes it or changes an answer because of it.

Is schema markup worth it for AI search? It is worth maintaining for the ordinary Google Search purposes its documentation supports. Google says no special structured data is required for AI Overviews or AI Mode, and there is no markup shortcut to a citation. Use a schema type only when it accurately describes visible content and satisfies the applicable Search guidelines.

What schema types should a business add first? There is no universal schema checklist for AI search. Choose from Google's supported structured-data features based on the visible page and the Search feature you are eligible for. Organization, Article or BlogPosting, Product, FAQPage, and BreadcrumbList can be relevant in different contexts, but every property must match the page and the applicable guidelines.

Do I need both llms.txt and schema? No AI feature requires both. Maintain supported schema for its documented Search purposes, and test llms.txt only if it is low-effort and its use can be verified. Neither is documented as an AI-citation signal.

Can a managed service handle the technical GEO setup? Yes. Beamcite's 90-Day Visibility Plan can include research, authority architecture, technical fixes, internal links, human-reviewed content, publishing, measurement, and iteration. The exact implementation, reporting, and commercial terms are scoped on the founder call, and the work does not guarantee rankings or citations.

---

Written by Abhiijay Vinayak, co-founder of Beamcite. Beamcite runs a managed 90-Day Visibility Plan for founder-led software companies without a capable in-house organic team. You can talk to the founder.