The Difference in One Minute
llms.txt explains a website in readable Markdown. It helps an AI agent find the pages and documentation that matter without digging through every menu.
llm.json expresses a site summary as named JSON fields. That makes the information easier for software to validate, compare, and reuse.
One is a guide. The other is a data record. Publishing both can be useful when they are generated from the same current source.
| File | Best use | Format | Typical URL | Main risk |
|---|---|---|---|---|
| llms.txt | A readable site guide with selected links | Markdown text | /llms.txt | It becomes another stale summary when maintained by hand |
| llm.json | A structured site summary for software | JSON fields | /llm.json or /ai/llm.json | There is no universal public contract for every field |
What llms.txt Gives an AI Agent
The llms.txt proposal defines a Markdown file that introduces a site and points to useful material. It is closer to a focused table of contents than a technical permission file.
- A plain description of the website
- Links to the pages an agent is most likely to need
- Optional notes that explain how the material is organized
- A format that a person can read without special software
llms.txt is a community proposal. It does not force a crawler or AI product to fetch the file, trust it, or cite the website.
What llm.json Gives Software
There is no single industry-wide llm.json contract. A site must declare what its fields mean and keep the file tied to its real content.
AItoAIConnect uses the Digital Karma Federation contract. Its llm.json files are generated from the current manifest, catalog, and health records. That keeps the readable and structured summaries connected to the same site facts.
- A stable site identity and canonical URL
- Named machine-readable endpoints
- Current content and Dataset references
- A modification date that software can compare with the manifest
Neither File Guarantees AI Visibility
A clean endpoint can help an AI system find a better description of a site. It cannot make weak content useful or repair conflicting facts across the website.
The page content still has to answer a real question. The schema must match what the page says. The sitemap and internal links must point to live canonical URLs. The files only become valuable when the rest of the site agrees with them.
A Practical Four-Step Setup
Start with the site itself. Then generate the files from the facts that already exist instead of creating a second identity by hand.
- Choose the pages worth guiding an agent toward Use current canonical pages that answer important questions or expose useful data.
- Write or generate llms.txt Give the site a short readable introduction and link to the selected material.
- Generate llm.json from the same source Use stable fields and publish the file at a documented URL.
- Check both files after every content build Confirm that their URLs, dates, and content references still match the live site.
Choose the File by the Job
Use llms.txt when the next system needs context it can read. Use llm.json when the next system needs fields it can parse. Use both when one build can keep both accurate.