From search to conversation: What we learned adding AI to technical documentation

David Garcia, founder of Biel.ai and TechDocs.studio, shares key insights from two years of integrating AI chatbots and agentic workflows into API documentation.

David explores how AI exposes gaps in existing documentation and how teams can prepare their docs for both public and controlled AI agents. He discusses the changing landscape where AI agents are now a major audience for API docs, outlines practical steps to improve discoverability and quality for AI consumption, and highlights lessons learned—such as the importance of observability, iterative improvement, and maintaining up-to-date, well-structured content.

The talk includes actionable tips for making documentation more accessible to AI, strategies for monitoring and closing content gaps, and considerations for integrating AI tools into developer and documentation workflows.

Transcript

1. FROM SEARCH TO CONVERSATION What we learned adding AI to technical documentation David Garcia Founder, Biel.ai & TechDocs Studio Write the Docs Barcelona | April 2026
2. David Garcia | Biel.ai | Write the Docs Barcelona 2 / 18 SOMETHING CHANGED Two years ago, devs needed authentication help. They opened your docs. Today, they type the question in Claude Code. The AI answers. BEFORE Google > Docs site > Read > Code NOW Ask AI in IDE > Get code > Ship
3. David Garcia | Biel.ai | Write the Docs Barcelona 3 / 21 THE DATA ~43% of documentation site traffic now comes from AI agents, not humans Source: Mintlify, 2026 AI-driven traffic Growing on docs sites Human pageviews Declining steadily New primary reader Doesn't browse. Queries.
4. David Garcia | Biel.ai | Write the Docs Barcelona 4 / 18 YOUR DOCS HAVE TWO AI AUDIENCES PUBLIC AIs You don't control these. • ChatGPT, Claude, Perplexity • AI coding assistants (Claude Code, Cursor, Copilot) • AI-powered search engines • Any service that crawls the web CONTROLLED ENVIRONMENTS You decide what they see. • AI chatbot on your docs site • MCP server for IDE tools • Internal AI assistants • Custom RAG pipelines
5. PUBLIC AIs The ones you don't control.
6. David Garcia | Biel.ai | Write the Docs Barcelona 6 / 21 PUBLIC AIs THE PROBLEM Internal docs invisible Confluence, Notion, PDFs in Google Drive, slides in SharePoint. AI can't see any of it. SEO lottery A 3-year-old Medium article may outrank your current API reference. Training data is stale New SDK version last month? Suggests deprecated methods. No product context Doesn't know your error codes, your custom config, or your CLI.
7. David Garcia | Biel.ai | Write the Docs Barcelona 7 / 21 PUBLIC AIs HOW TO WIN AI VISIBILITY Make it easy for AI engines to ingest your content. Be publicly crawlable No login walls, no JS-only rendering. Check robots.txt for GPTBot, ClaudeBot. Use IndexNow. Earn citations & backlinks Docs linked from Stack Overflow, GitHub READMEs, and blog posts carry more weight in AI training. Answer upfront Lead with the answer. AI prefers pages that get to the point. Structured data Schema.org for pricing, FAQs. Good OG tags, meta descriptions. Crawlers use these. llms.txt Like sitemap.txt but for LLMs. 73% of top API docs still lack one. Serve MarkDown URLs Pages at .md URLs. Agents prefer clean markdown over bloated HTML.
8. David Garcia | Biel.ai | Write the Docs Barcelona 8 / 21 PUBLIC AIs MEASURE WHERE YOU STAND Agent Score by Fern WHAT IT CHECKS llms.txt 5 checks Markdown availability 2 checks Page size & boilerplate 4 checks Content structure 3 checks URL stability 2 checks Discoverability 1 check Observability & auth 5 checks https://buildwithfern.com/agent-score
9. David Garcia | Biel.ai | Write the Docs Barcelona 9 / 18 PUBLIC AIs If your docs are invisible to public AIs, you're company is leaving revenue on the table. We improved our SEO We optimized our docs for search engine visibility. The results showed up in traffic and conversions. Docs drive signups Users discover Biel.ai through AI answers that link to our integration guides and API reference. AI cites our content When developers ask AI about chatbots for documentation, they reference our technical docs. DOCS ARE A SALES CHANNEL
10. CONTROLLED ENVIRONMENTS The ones you build and own.
11. Biel.ai chatbot widget on a documentation site A AI CHATBOT TRAINED ON YOUR DOCS CONTROLLED
12. David Garcia | Biel.ai | Write the Docs Barcelona 12 / 21 CONTROLLED A AI CHATBOT TRAINED ON YOUR DOCS WHAT WORKS Trained on your content only Better a "don't know" than a wrong answer from generic training data. Internal and public data. Users get instant answers on your site Reduce support tickets / Slack messages. Gives you observability See what users ask and improve your product. LIMITATIONS Web UI only Users must visit your site. Doesn't reach devs in their IDE. Only as good as your docs Missing content = missing answers. Day 1 will expose gaps. Needs maintenance Content must stay in sync with product changes.
13. MCP server querying docs from an IDE (Claude Code / Cursor) CONTROLLED B MCP SERVER FOR AI TOOLS
14. David Garcia | Biel.ai | Write the Docs Barcelona 14 / 18 CONTROLLED Direct connection to the IDE Claude Code, Cursor, Copilot query your docs. Best for internal setups Bundled with company tooling, adoption is natural. LIMITATIONS Optional for external users Devs have to choose to connect it. Adoption is not guaranteed. Works best in a package Standalone MCP is a hard sell. Bundled with company tools, it sticks. Still emerging MCP is new. Tooling and standards are still maturing. B MCP SERVER FOR AI TOOLS WHAT WORKS
15. LESSONS FROM THE FIELD What happens when you actually deploy an AI chatbot + MCP server.
16. David Garcia | Biel.ai | Write the Docs Barcelona 16 / 21 LESSONS DAY 1: DISASTER Support tickets went up, not down. The assistant kept saying "I couldn't find information about this" for basic flows: Export data Reset password Cancel subscription AI assistants don't fix documentation problems. They make them visible. Error codes Troubleshooting Examples
17. David Garcia | Biel.ai | Write the Docs Barcelona 17 / 21 LESSONS DON'T OVERPLAN. ITERATE. You don't need perfect docs from day 0. Listen and fix. THE TRAP • Months of docs strategy planning • Theorize about what users need • Launch the AI chatbot once "ready" • Still discover the same gaps THE LOOP 1 Deploy AI on your docs 2 See what users actually ask 3 Document the gaps 4 Watch AI improve 5 Repeat
18. Biel.ai dashboard: content gaps and most asked questions LESSONS DON'T OVERPLAN. ITERATE. David Garcia | Biel.ai | Write the Docs Barcelona
19. 19 / 21 LESSONS MICRO TIPS FOR BETTER AI ANSWERS CONTENT AND DISCOVERABILITY Every page is page one No "as mentioned above" Think in questions Cover what users ask Complete code examples Imports, config, file paths Consistent terminology One term per concept Avoid JS-heavy rendering Crawlers read source HTML Use keywords & metadata Tags, descriptions, OG. SEO for AI Describe images in text AI can't see screenshots Separate chatbots per product Don't mix, or context bleeds Less is more Keep blog posts, release notes, marketing pages … Full post with tips: blog.techdocs.studio
20. LESSONS MICRO TIPS FOR BETTER AI ANSWERS ITERATE Keep content up to date Stale docs = wrong answers. Sync with product changes. Build an eval set 50-100 real user questions. Grade answers. Find failures. Sandbox before deploying Test in staging first. Catch issues before users do. Track retrieval misses Log what AI can't answer. That's your docs roadmap. Flexible platform Models, prompts, restrictions: it all changes fast. Full post with tips: blog.techdocs.studio
21. THANKS! LET'S TALK. David Garcia david@techdocs.studio blog.techdocs.studio biel.ai Questions? Disagreements? War stories about your own docs?