LLMs and Schema, how LLMs read Schema

Published: 10-25-2025
by Wayne Smith

Before considering how large language models interpret schema, it’s important to understand its purpose—and to do so without the lens of search features or the knowledge panel. Those functions belong to search engines; LLMs neither populate knowledge panels nor generate search features. Many of these features can exist without schema, but schema provides explicit clarity, resolving ambiguity that can arise from both language and the user interface.

An address—an entity representing a place—may appear at the top of a page, and while its purpose may be clear to human visitors because of design or placement, machines lack that contextual understanding. Schema makes the intent explicit: it defines whether the address belongs to a business, a venue, or an event location. By providing this structure, schema clarifies meaning and reinforces the relationships already implied by on-page language. It gives models explicit signals to interpret entities, context, and intent more reliably than through text alone.

Without schema, machines must rely on secondary signals to determine what an address—or any entity—represents, drawing inferences from how that information appears across other online sources. Both search systems and LLMs can perform this kind of contextual reasoning, but there is a clear advantage in explicitly providing the information through structured data. Schema removes the need for machines to guess -- or hallucinate --about the role and relevance of entities on a page.

Outside of its role in powering search features, schema also helps shape the perceived relevance of the entities described on a page ... an influence that ultimately affects how both search engines and AI systems assess topical visibility.

LLMs Can Use the Schema Data

There is considerable debate within the SEO community regarding schema, AI, and large language models. One viewpoint argues that LLMs do not process schema as structured data tables—and, from a purist perspective, without such processing, they cannot truly understand schema. However, this view overlooks how meaning can still be inferred. Just as humans can read a block of JSON-LD and comprehend its intent without converting it into a database table, an LLM can extract relationships, entities, and contextual signals directly from the structured text itself.

Consider this microdata markup:

Technically, this microdata is incorrect and cannot be processed into a formal data table because "Event-location" is not a valid schema property. Nevertheless, both humans and AI systems can infer from the surrounding context that the address refers to the location of an event. On the visual page, the image of the venue provides additional context, making the text, "Event-location," unnecessary for a human observer. Despite the incorrect markup, LLMs can still explicitly link the address to an event, extracting meaningful relationships directly from the page content.

Based on the page content, AI systems can answer questions such as where the event is taking place, demonstrating that schema is helpful but not strictly required for comprehension. In this example, the image of the venue is adjacent to the address, which reinforces understanding. However, other cases exist where the venue image is not adjacent. Historically, adjacency of related content has been a signal for SEO, helping both humans and machines establish relationships between entities.

There is no debate that LLMs can see and interpret microdata and JSON-LD

JSON-LD has been used specifically as an alternate language for generative AI systems because, in many cases, it is less ambiguous than human languages. People have also intentionally obfuscated content they do not want LLMs to use with such techniques. The real debate is not about visibility, but about how LLMs interpret and integrate the information.

Schema as trust but verify content

Within SEO, schema functions as a trust-but-verify signal. It allows machines and search systems to understand information more quickly without relying solely on cross-references from other pages or sources. However, the accuracy of schema is critical: incorrect or misleading data can undermine trust, reducing its effectiveness and the benefits it provides. In essence, schema accelerates comprehension while signaling reliability ... but it must be maintained carefully to preserve that trust.

Maintaining correct and accurate schema is therefore essential to ensure benefits across all SEO and AI-driven channels. Hence for part 2: "New Search Ecosystem: AI, Schema, and Entities" correct schema and microdata will be maintained, and examples provided with adjacent microdata ... not because it improves LLM interpretation, but to make the information clear and easier for humans reviewing the examples.