Do LLMs Actually Use Schema? The Duck Test That Broke SEO

·12 min·1 clip

Mark Williams Cook's experiment reveals whether LLMs actually use schema or just treat it as text.

This episode examines whether large language models actually parse and use structured data markup, or schema, as intended for SEO and AI visibility. Host Edward explores a recent experiment by SEO expert Mark Williams Cook, who will be a guest on an upcoming show. The central question is if schema markup genuinely influences how LLMs like ChatGPT and Perplexity interpret and cite web content. Mark Williams Cook created a fictional company called Duck Yeah T-shirts for his test. He deliberately omitted the company's physical address from the visible text on the webpage. Instead, he inserted a completely fabricated JSON-LD schema containing invented property types like "block name" and "waddle style." The experiment tested if LLMs would extract and output the fictional address data embedded only within this made-up schema. Both ChatGPT and Perplexity accessed the page and subsequently used the address from the schema in their responses. This outcome mirrors common "proof" cited by proponents of Generative Engine Optimization, who claim LLMs parse structured data in real-time. Cook's experiment directly challenges that assertion by demonstrating LLMs treat schema markup as ordinary page text. He concludes the models are not specifically interpreting schema as structured data but simply consuming all text on the page. The most striking insight is that an LLM can output information from a schema containing entirely nonsensical and invented property types. This suggests the model is not validating the schema's structure or meaning against a known vocabulary. The experiment implies that observed correlations between schema use and AI citations may not be due to special parsing. A surprising claim is that the common SEO practice of hiding data only in schema to test LLM uptake is methodologically flawed. The memorable moment is the creative use of a fictional duck-themed T-shirt company to frame the test. The core takeaway is that schema might not provide a unique or privileged signal to large language models for information retrieval. The tone is investigative and conversational, blending a direct explanation of an experiment with the host's enthusiastic commentary. The style is educational, walking through a specific test setup to challenge a widely held SEO belief. SEO professionals and digital marketers interested in AI search optimization would find this episode highly relevant. Listeners seeking definitive, proven best practices for technical SEO might be frustrated by its speculative and experimental nature.

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from The Edward Show