Performance

GoatSEO does its expensive work at development time. Schema.org is parsed and turned into Python by the generator; the installed packages contain ordinary modules and never read the ontology.

Import cost

Measured on CPython 3.14 (Apple silicon):

Import	Time
from goatseo import SEO	about 0.1 s
import goatseo.schema	about 0.08 s
First access to a generated type (from goatseo.schema import Article)	about 0.7 s, once

Most of the first two numbers is Pydantic itself. The third one is the creation of 789 Pydantic classes, dominated by Pydantic copying inherited fields down deep Schema.org hierarchies.

To keep that cost out of applications that do not use Schema.org, goatseo.schema loads its generated modules lazily: the package imports only the base classes and the JSON-LD serializer, and a module-level __getattr__ loads the generated types on first access. The SEO builder depends only on those base classes, so rendering metadata, Open Graph and Twitter never triggers the generated modules.

Validation cost

Schema.org is a deeply recursive graph: Thing refers to CreativeWork, which refers to Person, which refers back to Thing. Letting Pydantic build a full validation schema for one model would build it for most of the vocabulary.

Generated fields carry a prop marker instead. It validates each value with an isinstance check against the declared types and never asks Pydantic for the schema of nested models. Models also use defer_build=True, so a class builds its (small) validator only when first instantiated. Creating a first Article takes a few milliseconds.

Rendering cost

• Metadata layers are merged once per SEO object and cached until a layer changes.
• Settings (GOATSEO_DEFAULTS, extension defaults) are validated once at startup, not per request.
• JSON-LD is serialized once, directly from the models, without an intermediate model_dump.
• No adapter performs database queries; model metadata comes from objects your view already loaded.