David Karlsson (kapa.ai): Docs accessibility, for robots

This talk was presented at the AI The Docs 2025 online conference. We are thrilled to share the recording and the summary with you. 

Visit the talk summary page to see all of the presentations from the conference.

presentation

David Karlsson (Product Specialist at kapa.ai)

Summary

What are the primary methods by which Large Language Models answer questions? How can content be made most consumable and effective for AI systems, regardless of the LLM architecture? 

In his talk, David Karlsson (Product Specialist at kapa.ai) explains how to make content usable for machines, since AI assistants depend on content quality as much as on models. 

Key insights: 

• How LLMs answer questions:
  • Closed book (training data): broad reach, but hallucinations and cut-off issues.
  • Tool calling: dynamic data access, but immature decision-making.
  • RAG: controlled, reliable, and secure answers, but challenging to build well.
• Content-first approach: Fancy formatting and scripts get lost, plain text is what LLMs consume. Accessibility best practices (semantic HTML, headings, alt text, captions, simple markup) make content easier for both humans and machines.
• Principles for AI-ready content:
  • Make context explicit (avoid assumptions; FAQs are surprisingly effective).
  • Curate content for quality to reduce noise.
  • Use semantic information architecture (headings, URL paths).
  • Provide machine-friendly alternatives for visuals and tables.
• Emerging standards and tools:
  • LLM.txt helps models parse product and site structure.
  • MCP services (e.g., Microsoft’s doc server) enable direct agent queries. 

Takeaway: Better robot accessibility = better human accessibility. Content optimized for LLMs improves both automation and user experience.