What the data shows when docs meet LLMs

There are few benchmarks to know if you or your users are successful using AI with documentation. I delved into the Mintlify platform data to evaluate how people discover, read, and maintain documentation using AI.

I presented this information at Write the Docs Portland 2026 and built a companion site for the data.

Research

I pulled and analyzed agent-mediated interaction data across 17,784 Mintlify-hosted sites using ClickHouse and Helicone.

Rather than present a flat list of stats, I organized the findings into three stages of documentation interactions: discovering, reading, and writing. Agents find docs differently than people, use them differently, and increasingly help produce them.

Findings

Discovering — At the time of presenting, AI conservatively accounted for roughly 35% of documentation readership.

Reading — across 95,000 AI chat interactions, 55% were a single question with no follow-up. Humans lean on embedded AI assistants for quick answers. Conversations with a lot of back-and-forth generally had more negative sentiment.

Writing — Pull requests created by agents are merged about half the time. Agent-authored doc changes are good enough to trust in production workflows, but can be improved.

Presenting

My main goal was for the audience at Write the Docs (mostly technical writers) to have actionable data. I didn’t want to just dump numbers on people. So I built benchmarks for people to measure their documentation against.

• AI readership above 35% so that you know agents can discover your content
• 55%+ of AI chat interactions finishing in a single turn so your users are finding quick success instead of getting frustrated with a chatbot
• 55%+ of agentic PRs merged so that you’re not wasting tokens on content you don’t use