Pedro Dias says schema helps search, not LLM understanding of content

Pedro Dias is pushing back on one of the loudest promises in the AI search sales pitch: that schema markup can make large language models understand content. The former Google Search Quality team member, now a Search Findability Architect and AI Discovery and Growth Product Manager in London, says the current “technical GEO” framing often repackages standard SEO hygiene as a new breakthrough.

His critique lands in the middle of a crowded market for “AI search visibility” tools and playbooks built around structured data, entities and other technical signals. Dias argues that LLMs process language by tokenizing unstructured text, not by relying on a special parser that reads Schema.org tags and infers meaning sentence by sentence. In his view, schema still matters, but for familiar search functions such as rich results and entity disambiguation, not as a magic key to model comprehension.

That distinction matches Google’s own guidance more closely than many vendor decks do. Google says structured data helps it understand the content of a page and information about entities such as people, books and companies, and can make pages eligible for richer search appearances. It also warns that eligibility is not a guarantee: a result may qualify and still never show up. Google’s AI-features documentation says the same core SEO practices remain relevant for AI features including AI Overviews and AI Mode.

The timing helps explain why the argument has gotten so sharp. Google said AI Overviews rolled out broadly to users in the United States at Google I/O on May 14, 2024, after the feature had already been used billions of times in Search Labs testing. As search shifts from blue links toward generated summaries and citations, publishers and vendors are racing to define the new rules, and the word “ensure” has become especially loaded.

The SEO community remains split. Some practitioners say structured data can help LLM-based systems and AI search engines ground or interpret content more efficiently. Others say claims that schema markup directly improves LLM output go beyond the evidence. Dias is firmly in the second camp, arguing that technical cleanup is useful, but that strong unstructured content still does the heavy lifting. In a market hungry for shortcuts, his warning is simple: schema can help search systems see a page more clearly, but it does not make an LLM understand the page for you.