Schema.org and RAG: why structured data changes LLM retrieval

Retrieval-Augmented Generation (RAG) powers most AI search products. When a user asks "best waterproof hiking boots under $200," the system retrieves chunks, then synthesizes an answer with citations.

Why HTML alone underperforms

Unstructured pages mix navigation, promos, and specs. Chunking algorithms (fixed-size, semantic, or heading-based) often split product facts across irrelevant segments. Schema.org JSON-LD sits in a predictable `<script type="application/ld+json">` block — many crawlers index it as high-confidence structured evidence.

Merchant-leveraged schema types (ranked by citation impact in our dataset)

1. Product + Offer (priceValidUntil, availability) 2. FAQPage (natural-language Q&A) 3. HowTo (setup, sizing, care guides) 4. Review / AggregateRating (when authentic) 5. BreadcrumbList (context for category pages)

Anti-patterns

- Markup that contradicts visible content (Google and AI systems penalize mismatch) - Fake reviews or aggregateRating inflation - Product schema on category pages (use ItemList instead)

Research note

Lewis et al.'s RAG framework (2020) established that retrieval quality caps generation quality. For merchants, *retrieval quality* increasingly means *structured fact quality*. Visora audits schema validity, field completeness, and HTML-JSON consistency — the three variables you control without ML expertise.