Everything in detail.

Share this section

Agent-Ready describes a state in which a website is fully accessible not only to human readers but also to machine readers, in particular Large Language Models and the agents they drive. A website earns the badge if it meets three core requirements:

• /llms.txt sits at the root and describes the site in clear Markdown: purpose, key subpages, contact details, citation rules. Following the proposal by llmstxt.org (Jeremy Howard, Answer.AI, September 2024).
• Structured data in JSON-LD according to schema.org: at minimum Organization and WebSite, and depending on content also Article, TechArticle, FAQPage, Product, Event, or Person.
• Semantic HTML with a clean heading hierarchy, real <nav>/<main>/<article>/<section> tags, descriptive alt text, and meaningful link anchors instead of "click here".

No new framework, no dependency, no fees. Pure craft, back to the way HTML was originally meant to be, before JavaScript frameworks and cookie banners made pages inaccessible to machines.

Share this section

In thirty years, the internet has produced two universal quality seals. Both began as an insider topic and became the standard within a few years. Agent-Ready stands on the threshold of becoming the third:

1. HTTPS (since around 2014): the padlock in the address bar. Indicates: the connection is encrypted. Today: a baseline requirement; without HTTPS, browsers flag a site as insecure.
2. Mobile-Friendly (since 2015): Google rewards mobile-optimized pages in search. Today: a baseline requirement; non-mobile sites have effectively disappeared.
3. Agent-Ready (from 2026): the seal indicating that a page is structurally readable for AI agents. Within a few years, the standard against which credibility and visibility on the web will be measured.

Whoever adopts Agent-Ready today gains the same head start that early adopters of HTTPS and Mobile-Friendly enjoyed, with the difference that AI-powered search is spreading very quickly right now.

Share this section

Search is shifting. Anyone in 2026 looking up a fact, a product recommendation, or a medical interpretation no longer automatically types into the Google search box, but puts the question directly to a language model. ChatGPT, Claude, Perplexity, Gemini, and the AI assistants built into browsers, IDEs, and operating systems pull answers from the web, but they deliver them not as a list of links, but as a summarized answer with attribution.

This fundamentally changes the rules of the game. For ten years, SEO meant keyword density, backlinks, page speed, meta descriptions. That is no longer enough. A language model needs clear prose, machine-readable structure, and concrete facts with sources. It ignores text that is merely keyword-optimized and rewards text that explains something to an intelligent reader.

This new optimization approach already has a name: Generative Engine Optimization, or GEO for short. Agent-Ready is the visible label for that approach. Anyone who carries the badge signals: this site is built not for SEO tricks, but for substantial answers, and that is exactly what models pull into their responses.

Share this section

An AI agent has only a few seconds to determine whether and how a page is relevant to a specific user question. The process typically follows the steps summarized under the term Retrieval-Augmented Generation (RAG):

1. Discovery: The agent finds the page via a search engine, a link in another answer, an entry in an llms.txt file, or a direct user link.
2. Fetch: It retrieves the page, ideally as static HTML. Pages that load their content only through JavaScript are invisible to many agents.
3. Parse: The agent extracts headings, paragraphs, lists, tables, and structured data. The cleaner the HTML, the more reliable the result.
4. Chunk & select: The text is split into sections or handed to the model whole, and the passages that fit the user question are selected. Search indexes often use vector embeddings for this.
5. Cite: The matching sections flow into the generated answer, usually with attribution and a link back to the originating page.

Agent-Ready targets steps 2 through 5. A cleanly structured page is found more reliably and read more accurately, which can raise the chance of being cited. Sites without a clear structure are more likely to be misunderstood by AI assistants, because the model has to guess what the page is actually saying.

Share this section

Behind Agent-Ready sits a larger shift in online marketing. The industry has a name for it, coined in a 2023 research paper: GEO, Generative Engine Optimization. SEO optimizes for appearing high in Google's results list. GEO optimizes for being named and cited directly in the finished answers of ChatGPT, Perplexity, Gemini, and Claude.

The difference sounds small but is strategically large. A user who receives a finished AI answer often does not click on the classic result at all. If you appear in the answer itself, you are visible. If not, you disappear, even while your Google rankings still hold.

Where do SEO and GEO differ?

SEO and GEO differ in eight points: the goal, the form of the search output, the platforms, the ranking signals, the type of query, the shape of the user journey, the stability of the results, and the success metric. The table below places the differences side by side.

Why is Agent-Ready the technical foundation for GEO?

GEO is the strategy. Agent-Ready is the prerequisite. For an AI to cite a page at all, it has to read the page's structure cleanly. That works more reliably the more clearly three building blocks are in place: an llms.txt as a table of contents for language models, schema.org structured data as a machine-readable label per page, and semantic HTML instead of div soup.

Anyone serious about GEO without implementing Agent-Ready is optimizing text for a machine that does not reliably understand the page. The two belong together: GEO decides what is on the page. Agent-Ready makes it easier for the AI to find it and read it correctly.

SEO remains the entrance. Without indexing, no AI agent reaches the page either. GEO and Agent-Ready build on top of it. Whoever serves all three layers cleanly is visible in the AI era.

Share this section

Google published its own guide to optimizing for generative AI features in May 2026. Its core message: classic SEO and "people-first content" are enough; an llms.txt is not necessary; special AI markup is a myth. The message sounds tidy. To read it correctly, you have to know where it comes from.

Google is not neutral on this question. With AI Overviews, Google itself is a provider of an AI answer product and therefore a direct competitor to ChatGPT, Claude, and Perplexity. Google's strength is the web index it has built for almost 30 years. Google does not need an llms.txt to understand a page. The other AI platforms do not have that index. They depend on website operators delivering content in a clean structure. That is exactly what llms.txt does. If it becomes the norm, Google's role as middleman between website and AI shrinks. That may explain why Google lists it among the unnecessary measures in its guide. This is our assessment.

The other AI providers see it differently. Bots from OpenAI and Anthropic demonstrably fetch the file, though rarely. That it increases citations has not been shown so far.

In practice: Google considers schema.org structured data and semantic HTML useful, for example for rich results and screen readers, but says explicitly that neither is required for AI answers. Only llms.txt is contested, and the dispute is explainable as business strategy, not as a technical fact. An llms.txt costs one text file and ten minutes of work. It does not hurt with Google and may help with other AI platforms. Whether it increases your chances of being cited has not been shown so far. That is exactly the bet Agent-Ready is making.

Share this section

Perplexity is one of the best-known "answer engines": instead of a list of hits, it delivers a summarized answer and names its sources with direct links. Whoever appears there as a source has a good chance of getting visitors.

The mechanics behind it work differently from Google and from Claude or ChatGPT. According to observations from the SEO industry, not officially confirmed, Perplexity pays particular attention to relevance, authority, freshness and structure. llms.txt plays a minor role here. The real levers are:

• Freshness. Perplexity favors content updated within the last 60 to 90 days. A visible "Last updated: October 2026" plus a maintained dateModified in the JSON-LD show that the page is current.
• Clear structure. Each heading framed as a concrete question, followed by a 40-to-80-word answer, then the explanation. Tables and lists for comparisons. Perplexity extracts such passages cleanly.
• External authority. Mentions in Reddit discussions, in "best of" lists from reputable industry publications, and in Wikipedia. Perplexity actively reads these signals.
• Original data and studies. Primary sources that others must cite show up disproportionately often in Perplexity's top-three citation positions.
• PerplexityBot not blocked in robots.txt. If you block it, Perplexity will not crawl the page. On Fiperly this is in place.

In practice: anyone aiming for Perplexity visibility does not write primarily for an AI but establishes an update discipline, builds comparison and review content with a clear question-answer structure, and ensures the brand is mentioned outside its own domain. Agent-Ready remains the technical foundation (semantic HTML, schema, readable structure), but the bulk of the work for Perplexity visibility lies in editorial discipline and external presence.

Share this section

To complete the picture: here is an honest map of the major AI answer engines and their respective levers. What works at Anthropic Claude only helps at the margin with Perplexity, and Google ignores still other things. The good news: whoever cleanly serves all three Agent-Ready building blocks and maintains content with structure, freshness, and sources covers every platform at once.

The single most important measure from this overview: allow all relevant AI crawlers, in your robots.txt and in your host's settings. On Cloudflare, new default settings block some AI bots before the robots.txt is even read, and allowing them there does not lift that. Anyone who misses this shows up much less often in AI answers. Fiperly has all the bots listed above enabled in its robots.txt.

Share this section

Four additions have established themselves in 2026 as extra leverage. They are not required for the badge, but sites that take 2026 seriously have them in place:

/llms-full.txt: the full-text variant

According to BuiltWith, more than 844,000 sites publish llms.txt, but it is rarely read. The next differentiator is llms-full.txt: same Markdown structure, but instead of navigational links it contains the consolidated full text of your most important pages in a single file. LLMs get the entire brand context in one fetch, without having to parse dozens of HTML pages. Especially useful for sites of manageable size: marketing hubs, documentation, brand landing pages.

Anthropic: three bots plus Claude Code

Anthropic officially names three bots, each with its own purpose: ClaudeBot (training), Claude-User (fetches for user questions) and Claude-SearchBot (search). In server logs you will also see the identifier claude-code when the coding tool Claude Code fetches a page. If you want fine-grained control, list all four individually in your robots.txt.

Strategic decision: search bots vs. training bots

One of the most important business decisions in 2026: should training bots like ClaudeBot, GPTBot, CCBot, and Google-Extended be allowed to ingest your content for model training? Sites that need to protect paid content (premium articles, exclusive data, subscriber-only knowledge) block them and only allow the retrieval bots (Claude-SearchBot, OAI-SearchBot, PerplexityBot, ChatGPT-User, Claude-User). So content appears in AI answers but doesn't flow into training. Sites that prioritize reach and have no paid content allow both groups. Both strategies are legitimate in 2026. They just need to be a deliberate choice, documented as a comment in robots.txt.

The name tag: a small file for assistants

Think of the name tag as a door sign that only assistants read. It is a tiny file at a fixed spot: /.well-known/ai-catalog.json. It lists which doors (MCP servers) the site offers and where their description lives, the so-called Server Card. The Server Card has been an optional MCP extension since 2026; the catalog itself is a separate open format that is still a draft. Many assistants do not read either yet.

The honest part: the tag alone does nothing magical. It only says where the door is. An entry belongs in it only if a real door runs there (see stage 2). If there is none, you need no sign. A tag that promises a door that does not exist does more harm than good.

One rule: place the tag only once your site is publicly visible, not while it is still under construction and blocked for search engines.

Share this section

Here's what a minimal llms.txt looks like. Place the file in the root directory of your website (i.e. at https://your-domain.com/llms.txt) and adapt the contents:

# Company Name

> Short description: What the site is and who it is made for.

## Key pages

- [Home](https://your-domain.com/)
- [About us](https://your-domain.com/about)
- [Products](https://your-domain.com/products)
- [Contact](https://your-domain.com/contact)

## Contact

- Email: info@your-domain.com
- Responsible: First Last

## Citation rules

Please cite this site with full name and URL.

That's it. No build step, no framework: a simple text file in Markdown.

Share this section

Add this code block to the <head> of every page. It tells search engines and AI agents the basic information about your company and website:

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "Company Name",
  "url": "https://your-domain.com",
  "logo": "https://your-domain.com/logo.png",
  "description": "What your company does.",
  "email": "info@your-domain.com"
}
</script>

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "WebSite",
  "name": "Company Name",
  "url": "https://your-domain.com",
  "inLanguage": "en-US"
}
</script>

Test the result with the Google Rich Results Test. You'll see immediately whether Google detects the data correctly.

Share this section

Agent-Ready is not a concept on paper. It is already running in production on these websites:

• fiperly.de: Parent brand and initiator of the Agent-Ready standard. llms.txt, JSON-LD (Organization, WebSite, TechArticle, HowTo, FAQPage), semantic HTML, hreflang DE/EN. Since October 2026 also Level 2: its own MCP door for the website offers.
• mutify.de: event calendar for all of Germany. Level 1 plus its own MCP door (Level 2) with three tools and a link to buy tickets, tested in ChatGPT as a custom connection. Submitted to OpenAI, review in progress.
• kipode.de: Fact-based information about Germany's situation in 2026. llms.txt, JSON-LD (Organization, WebSite, FAQPage), pillar articles with TechArticle schema.

Your site is missing from the list? Drop us a line. We're happy to include every Agent-Ready website.

Share this section

llms.txt

A Markdown file at the root of a website, proposed in September 2024 by Jeremy Howard (co-founder of Answer.AI and fast.ai). It explains in plain text to a language model what the site is about and which subpages are most important. The format is deliberately simple: an H1 with the project name, a short summary as a blockquote, then H2 headings with lists of relevant URLs. Citation rules are an additional Agent-Ready rule. In short: a business card of the website for AI assistants. It is not an official standard but a voluntary proposal, and many AI crawlers rarely read it so far. It complements clean HTML and structured data, but does not replace them.

robots.txt

The house rules for bots, in use since the 1990s and fixed as an internet standard since 2022 (RFC 9309). It sits in the root directory, for example at https://example.com/robots.txt, and tells crawlers which areas they may visit. User-agent names the bot (* for all), Disallow blocks paths, Allow permits them, Sitemap points to the sitemap. Without a robots.txt or without a matching entry, everything may be crawled. The file is a request, not a technical block: reputable bots follow it, others do not, and fetches made at a user's direct request often do not either. It matters for AI because you can allow or block individual bots such as GPTBot or ClaudeBot.

GPTBot

OpenAI's crawler that collects public pages for training the GPT models. OpenAI uses three separate identifiers: GPTBot for training, OAI-SearchBot for search in ChatGPT and ChatGPT-User for fetches a user triggers in a conversation. User-agent: GPTBot with Disallow: / in robots.txt blocks only training. The page can still appear in ChatGPT answers as long as OAI-SearchBot is allowed. Without an entry, GPTBot may crawl.

schema.org

A shared vocabulary from Google, Microsoft, Yahoo, and Yandex for structured data on the web, launched in 2011. Through types such as Organization, Person, Article, or Product, content is annotated in a machine-readable way. The best way to embed schema.org is as JSON-LD in the <head>.

JSON-LD

JavaScript Object Notation for Linked Data, Google's recommended way to embed schema.org. A JSON block inside <script type="application/ld+json"> doesn't disturb the HTML but is immediately processable for crawlers and LLM agents.

Generative Engine Optimization (GEO)

The evolution of SEO. While SEO aims to appear at the top of result lists, GEO aims to be cited in the generated answers of AI systems. The key levers are clear facts, verifiable sources, structured text, and machine-readable markup.

Retrieval-Augmented Generation (RAG)

A method in which a language model does not answer purely from its training knowledge but retrieves matching sources at runtime and incorporates them into the answer. RAG is the standard for modern AI assistants and the reason why websites are becoming increasingly important as a supply source for agents.

Semantic HTML

HTML that conveys its meaning through tag names. Navigation belongs in <nav>, the main content in <main>, a self-contained article in <article>. For machines, the difference between a <div class="nav"> and a real <nav> is significant. Only the latter is intelligible without guessing class names.

LLM bot / AI crawler

An automated retrieval agent operated by an AI provider. Examples: ClaudeBot, Claude-User, Claude-SearchBot (Anthropic), GPTBot and OAI-SearchBot (OpenAI), PerplexityBot (Perplexity), Google-Agent (Google), plus the control token Google-Extended for Gemini training, DuckAssistBot (DuckDuckGo), MistralAI-User (Mistral), Meta-ExternalAgent and Meta-WebIndexer (Meta), Amazonbot (Amazon), CCBot (Common Crawl). Access is mostly controlled through robots.txt, but user-triggered fetches do not always follow it.

Share this section

In its first thirty years, the internet produced two great quality seals: the HTTPS lock next to the address bar, indicating that a connection is encrypted, and the Mobile-Friendly notice, with which Google from late 2014 distinguished mobile-ready pages from non-mobile ones. Both seals became the standard within a few years and made the web measurably better.

With the breakthrough of language models, a third seal is missing: one for sites that AI agents can cleanly understand. There are good individual building blocks, llms.txt, schema.org, semantic HTML, but no shared, visible marker that brings them under one roof. This is precisely the gap that Agent-Ready closes.

The idea is explicitly open: the badge belongs to no one. Fiperly provided the impulse, formulated the criteria, and supplied the logo file, but the standard only lives if many adopt it. Every agency, every startup, every blog, and every public institution is invited to embed the badge, without registration, without fees, without strings attached.

The specification, example files, and badge are also published openly on GitHub at github.com/michaelbertelsenmedia-commits/agent-ready. The repository is licensed under CC0 1.0 Universal (public domain) and accepts pull requests, translations, and implementation reports.

We believe that within a few years Agent-Ready will be as self-evident a part of web craftsmanship as HTTPS or Mobile-Friendly. The web has proven in every new era that it can renew itself technically without losing its open character. The AI era is no exception. It just needs someone to take the first step.

Share this section

Fiperly is an independent AI development company from Germany. We build brands that put artificial intelligence to work where it measurably helps people, in public information, health, events, and everyday life. That a website in the AI era should also be readable for agents is, for us, a duty rather than a nice-to-have. We make the badge freely available because there has been no free seal for this so far.

The badge belongs to no one. It is in the public domain. Use it, copy it, modify it. If you'd like, drop us a line letting us know where you've put it. We're collecting examples.