Structured data for LLMs

LLMs process text, but not all text is presented to them equally. JSON-LD markup adds an explicit semantic layer telling the model: "this is an organisation", "this is a product", "this is a question and its answer". That information reduces ambiguity and improves extraction quality.

Additionally, the "sameAs" schema links your brand entity to external identifiers (Wikipedia, Wikidata, LinkedIn, Crunchbase), letting the model cross-reference sources and validate information. Without that web of links, your brand becomes an isolated node easy to confuse or skip.

Google has recommended JSON-LD for years and all major crawlers (including LLM crawlers) process it correctly. Compared to microdata or RDFa, JSON-LD lives in an independent <script> block separate from the visible HTML, which makes maintenance easier and reduces errors.

Best practice is to centralise JSON-LD generation in a reusable component of your CMS or framework. This avoids duplication, copy-paste errors and lets you update every schema at once when properties change.

The Organization schema defines what your company is, where it is, what it's called, which external identifiers it has. It's the first markup to implement and should be present on every page of your site, ideally in the main layout.

The most underrated field is "sameAs": an array of URLs linking your brand to its official profiles on Wikipedia, Wikidata, LinkedIn, Crunchbase, GitHub and social networks. For LLMs this array is gold: it lets them unambiguously identify your entity and aggregate information from multiple sources.

Every blog post, guide or editorial piece should carry the Article schema (or its subtypes: NewsArticle, BlogPosting, TechArticle). This tells the model which author signed it, when it was published, when it was last updated and which entity publishes it.

Particularly important for GEO are the fields "author" (with its own Person schema linked via sameAs to external identifiers), "datePublished", "dateModified" and "publisher". These are the signals the model uses to judge freshness and reliability.

FAQPage and HowTo are the highest-direct-impact schemas for GEO. FAQPage explicitly marks question-answer pairs, which models can extract almost verbatim. HowTo describes sequential steps with instructions, ideal for practical guides.

Implementing FAQPage on pages with real (not invented) frequent questions multiplies citation chances in generative engines. A caveat: Google has restricted FAQ rich results to official and health sites, but LLMs still read and value the markup even when it isn't shown in the SERP.

If you sell products or services, implement Product (for physical) or Service / Offer (for professional services). These schemas communicate what you offer, at what price, under what conditions and with what reviews, which helps the model recommend you when users ask for options in your category.

For Product, include name, description, brand, sku, offers, aggregateRating and review. For Service, define provider, serviceType, areaServed and offer. The more complete, the higher the chance of the model including you in comparisons and recommendations.

Implementing a schema badly can be worse than not implementing it. Always validate with Google's Rich Results Test and Schema Markup Validator (validator.schema.org). Common errors are incorrect date formats, missing required properties and badly closed circular references.

Set up a periodic schema review process. When you change content structure, update prices or reorganise categories, the schema must keep up. Outdated structured data generates distrust in the models and can hurt citation chances.