GEO Technology June 3, 2026 5 min read

Structured Data for AI Agents: How ChatGPT Reads Your Website

How structured data works in 2026: JSON-LD, schema types, and practical steps so AI agents understand and cite your content correctly.

Thomas Leonhardt Founder & AI Orchestrator, FORGE

Why AI Agents Read Differently Than Google Bots

Anyone who has optimized their website for search engines knows the principle: Googlebot crawls pages, evaluates links, indexes content. AI agents like ChatGPT, Perplexity, or Gemini work fundamentally differently — and that has direct consequences for how you structure your content.

Google crawlers work continuously. They build an index that is then searched when queries come in. AI agents, by contrast, combine two sources: a pre-trained language model and — in an increasing number of cases — a live web search or direct page requests. What the agent extracts from a page depends heavily on how clearly and unambiguously the information is available in machine-readable form.

The decisive problem: running text is meant for humans. A language model can understand running text, but it prefers structure. If an agent wants to know whether your company is based in Munich, it has to read that out of a paragraph — or it finds it immediately in the schema.org/address field. The difference is not academic; it determines whether your content appears in an AI answer or not.

This concept is called Generative Engine Optimization (GEO) — optimizing content specifically for generative AI systems. Structured data is not a nice add-on here but the foundation. AI agents can parse structured data directly and reliably, with no room for interpretation. That makes it the most precise form of communication between your website and an automated system.

The Most Important Schema Types for GEO 2026

Not every schema is equally relevant to AI agents. Four types have proven especially effective:

Organization

schema.org/Organization is the fundamental record about your company. Name, address, URL, logo, founding date, contact details — everything that enables an AI agent to answer factual questions like: "Who is behind this website?" or "Where is the company located?". Without this schema type, your company remains a black box to AI agents.

Article

schema.org/Article — or more specifically BlogPosting, NewsArticle, TechArticle — gives an article context: author, publication date, modification date, headline, description. AI agents need to know when a piece of content was created in order to classify it correctly. Outdated information without a date is either ignored or attributed incorrectly. The dateModified field is at least as important as datePublished.

FAQPage

schema.org/FAQPage is currently probably the single most valuable schema type for GEO. Why? Because AI agents reproduce exactly this format — question, answer, done. If you structure your most common user questions as an FAQPage, you hand the agent a ready-made answer unit that it can use directly. No interpretation effort, no uncertainty.

HowTo

schema.org/HowTo works with similar precision for instruction-based content. Each step is marked up separately as a HowToStep, with a name and a description. When a user asks an AI agent: "How do I install X?", and you have a correctly structured HowTo page, the chances rise considerably that your content is used as a source — or at least flows into the answer.

JSON-LD vs. Microdata vs. RDFa: What AI Models Prefer

There are three established methods for embedding structured data in HTML. The differences are technically relevant and influence how well AI systems can process the data.

JSON-LD

Weiterlesen — kostenlos

Den vollständigen Inhalt freischalten

Trag deine E-Mail-Adresse ein und bestätige sie: Du abonnierst den Signal-Forge-Newsletter von FORGE und erhältst sofort Zugang zu diesem und allen weiteren registrierungspflichtigen Inhalten. Die Abmeldung ist jederzeit möglich.

Schon registriert? Der Link aus deiner Bestätigungs-Mail schaltet dieses Gerät wieder frei.

JSON-LD (JavaScript Object Notation for Linked Data) is embedded in the <script type="application/ld+json"> tag in the head or body. It is completely separate from the HTML markup — that is its biggest advantage. An AI agent or crawler does not have to parse the entire DOM but finds all structured information in one place, in a clearly defined format.

Google explicitly recommends JSON-LD, for the same reason it is optimal for AI agents: it is robust, low-maintenance, and easy to generate automatically. A parsing error in the HTML markup does not affect JSON-LD.

Microdata

Microdata embeds attributes directly into HTML tags (itemscope, itemtype, itemprop). It is tightly linked to the visible content — that can be an advantage when you ensure that markup and content are always in sync. In practice, though, it is more maintenance-intensive and more error-prone. AI systems can process microdata, but the tighter coupling to the DOM makes it less reliable.

RDFa

RDFa (Resource Description Framework in Attributes) is the most powerful but also the most complex format. It allows very precise semantic markup but is rarely implemented correctly in practice. For most websites, the effort clearly outweighs the benefit.

Conclusion: For GEO 2026, JSON-LD is the clear choice. It is the format processed most reliably by all relevant AI systems, the easiest to implement and test, and the one Google and OpenAI prefer in their retrieval systems.

Hands-On Checklist: Implement and Test Structured Data in 30 Minutes

The following checklist is prioritized. Work through it from top to bottom — the first three points have the biggest effect.

  1. Add Organization schema on the homepage

    Create a JSON-LD snippet with @type: "Organization" and fill in at least: name, url, logo, address, contactPoint. Place it in the <head> of your homepage. This step takes under 10 minutes and is the basis for everything else.

  2. Add Article schema on all blog posts

    Every blog post needs @type: "BlogPosting" with headline, author (as a Person object with name), datePublished, and dateModified. If you use a CMS, this can usually be done once for all posts via a plugin or a template adjustment.

  3. Create FAQPage schema for your most important FAQ page

    Take the five most common questions from your users. Create a JSON-LD object with @type: "FAQPage" and a mainEntity array. Each entry is a Question object with name (the question) and acceptedAnswer (an Answer object with text).

  4. Test with the Rich Results Test

    Go to search.google.com/test/rich-results and enter your URL. The tool shows you immediately whether your JSON-LD is syntactically correct and which schema types were detected. Errors are output with line numbers — fix all errors (not just warnings) before the next step.

  5. Use the Schema Markup Validator

    At validator.schema.org you can check a URL or code directly. This validator is schema.org-native and checks more strictly than the Google tool. Especially useful for finding missing recommended properties.

  6. Follow up with HowTo schema for instruction pages

    If you have step-by-step content: now implement HowTo. At minimum: name (title of the guide), a step array with name and text for each step.

  7. Set canonical URLs in schema objects

    Every schema object should contain a url property that points to the page's canonical URL. This prevents attribution problems when content is shared or crawled.

  8. Monitor structured data in Search Console

Signal Forge · Free playbook

Get the free hands-on playbook.

Put AI agents to work. Plus monthly hands-on AI insights via Signal Forge.

Double opt-in · no spam · cancel anytime · The playbook arrives straight to your inbox

Next step

A GEO strategy for your business?

Get in touch →

More articles