How we keep AI-written descriptions factual
Language models are good at summarising and bad at not inventing things. Our pipeline is designed around that.
When we generate an llms.txt, two kinds of information end up in it:
- Facts: names, prices, currencies, SKUs, availability, contact details, policies
- Descriptions: one-sentence summaries of what a page covers
Facts are never generated
Facts only come from what your pages publish in machine-readable form: schema.org JSON-LD (Product, Offer, Organization, FAQPage), OpenGraph product tags, and explicit mailto:/tel: links. If a product has a price but no currency, we leave the price out rather than guess.
Descriptions are guarded
A local model (Qwen3 4B running on our own server) writes short descriptions. Every output passes a fact guard before it is used. We reject it if it contains:
- a number that doesn't appear in the source text
- a currency symbol the source didn't use
- an email address or URL
Rejected outputs fall back to your own meta description or first sentence. Descriptions are stored separately from facts, marked as AI-written in site.json, and listed with their source pages in the dashboard, where you can override them.
If the AI is unavailable
Generation never depends on the model. Without it you still get a complete, valid llms.txt built from your own metadata.