John Mueller recently answered a question a lot of SEOs are quietly obsessed with:

Does extensive schema markup help Large Language Models (LLMs) understand your entity better, or is it basically just for rich snippets?

He flagged it as his personal point of view — not official guidance.

Even though it’s “not official,” John's answer is one of the clearest reality checks we’ve seen in a while because it slices structured data into three buckets:

• Structured data that’s practically required for accurate machine use (high-fidelity facts like price/availability)

• Structured data that isn’t required, but makes comprehension and feature-enablement easier

• Structured data that’s mostly wishful thinking (where people try to “markup their way” into better rankings or “best” labels)

Bucket #1: “High-Fidelity” Facts — Where Machines Actually Need Structure

Mueller’s strongest point is that some data is basically impossible to extract from plain text accurately at scale — especially when it changes frequently. He called out shopping details like pricing, shipping, and availability.

This maps cleanly to how AI systems work in practice:

• LLMs don’t “crawl your page” the way SEOs imagine. AI answers are often generated from retrieval system...