Dynamic JSON-LD vs Hardcoded Schema

By Stas Levitan, CEO

· May 8, 2026

· 5 min read

The wall every schema project hits

The first 20 pages of JSON-LD on a site are fun. An engineer hand-writes Product, FAQPage, Organization. Everyone admires the rich result preview. Then the catalog grows to 2,000 SKUs, the FAQ database doubles, marketing launches a comparison hub, and the schema falls behind. Within a quarter, half of it is stale and a third is invalid.

This is the wall a dynamic LLM Discovery API was built to break through.

What "dynamic" actually means

Dynamic schema generation means a single service inspects each URL's rendered content and emits a fresh, valid JSON-LD graph at request time (or precomputed and cached at the edge). The page author writes content. The platform writes the schema.

The key properties:

One source of truth. Price changes in the CMS, schema follows.
Coverage. Every URL gets schema, not just the ones engineering had time for.
Validation built in. Invalid graphs never ship.
Versioned. When schema.org evolves (and it does, often), the platform updates centrally.

What hardcoded schema costs

Hand-rolled JSON-LD has three failure modes that compound over time:

Drift. Visible price says $49, schema says $39. Google demotes the page; AI engines stop trusting the source.
Coverage gaps. The blog has BlogPosting but the new product line ships without Product. AI engines extract nothing structured.
Invalid graphs. A missing required field silently kills a rich result. Nobody notices until traffic does.

A $20k schema retrofit from an agency fixes the snapshot. It does not fix drift. Three months later you're back where you started.

What changes when schema is dynamic

From the data we've seen across customer sites running the LLM Discovery API , three things move:

Pages extracted by AI engines: +12% on average within 30 days. This is the cleanest GEO metric we have.
Schema validity: ~100% — invalid graphs never reach production because the API refuses to emit them.
Engineering time on schema: near zero after initial integration.

When hardcoded is still the right call

For a 5-page brochure site, dynamic is overkill. Hand-write the Organization and one Service graph and move on. The break-even is roughly 50 URLs or any catalog that changes weekly.

How to migrate without breaking what works

Inventory existing JSON-LD. Note which pages have it and what types.
Stand up the discovery API in shadow mode — generate but don't serve.
Diff against existing graphs; resolve conflicts in favor of the dynamic source.
Cut over per-template, not per-page.
Monitor AI bot analytics for the next 30 days — extraction should rise, not fall.

Frequently asked

Does dynamic schema slow down pages? No, the JSON-LD is generated server-side or at the edge and cached aggressively.

What schema types are supported? Product, Article, FAQ, Service, Organization, BreadcrumbList, HowTo, Review — the types AI engines actually consume.

Does it work alongside existing JSON-LD? Yes; the discovery API can either replace or augment hand-written graphs per template.