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

Retrieval-Augmented Generation (RAG) powers most consumer AI search. The pipeline is conceptually simple: embed the query, retrieve top-k chunks, generate an answer with citations. Merchant visibility lives or dies at retrieval.

Navigation chrome, promo banners, and review widgets dilute chunks. A fixed 512-token window may capture "free shipping today" but miss wattage specs. JSON-LD lives in a predictable script block—many crawlers index it as high-confidence structured evidence alongside visible HTML.

1. Product + Offer (priceValidUntil, availability, shippingDetails) 2. FAQPage with natural-language Q&A 3. HowTo for setup, sizing, care 4. AggregateRating when authentically collected 5. BreadcrumbList for category context (ItemList on PLPs)

Markup contradicting visible price or stock. Fake reviews. Product schema on category pages. Multiple conflicting Product nodes from theme + apps.

• Price and currency match.
• availability reflects cart button state.
• name and description align with H1 and lead paragraph.
• FAQ answers verbatim or semantically equivalent to accordion text.

Does Google require schema for AI citations? No universal rule, but structured facts improve retrieval across systems that parse JSON-LD.

Is microdata enough? JSON-LD is easier to validate and less prone to template breakage.

Can I automate schema from metafields? Yes—Shopify metafields-to-Liquid is a common pattern; validate output in /audit.

Lewis et al. (2020) showed retrieval quality caps generation quality. For merchants, that means structured fact quality is the lever—Visora /audit scores validity, completeness, and HTML–JSON consistency without ML ops overhead.