Two in five docs readers are now AI
We looked at three months of traffic to documentation sites hosted on DeveloperHub. Here's who reads docs now, what the AI readers actually do, and what docs teams should change.
of the readers on docs sites are AI. Docs used to have one kind of reader. Now they have three.
Each square is 1% of docs readers, July to October 2026.
Reading in a browser.
Reading ahead of time, so AI search and models can use the page later.
Fetching a page right now, because someone asked ChatGPT, Claude or their coding agent about your product.
The AI share held at two in five. What's inside it changed.
Crawlers grew from 23% of readers in July to 34% in September. Assistants fell from 14% to 6%, because ChatGPT, which made most of their fetches, started fetching about 70% fewer pages on 21 August.
Claude's share of AI assistant traffic to docs, July to October.
Claude's own fetches doubled, and about 9 in 10 came from Claude Code: a developer asks their coding agent how an API works, and the agent goes and reads the docs. ChatGPT shrinking did the rest: since 21 August it fetches about 70% fewer pages.
Other assistants (DuckDuckGo, NotebookLM, Perplexity, Mistral, Meta AI) together made under 3%.
more people arrived at docs from ChatGPT in September than in July.
ChatGPT is sending more people to docs. But most AI reading never becomes a visit. When a developer asks Claude Code how your API works, it reads your docs and answers right there in their editor. It fetches the page's text without opening it in a browser, so most analytics tools, like Google Analytics, never count it.
requests calling itself an AI assistant was a vulnerability scanner in disguise.
Anyone can put "ChatGPT" in a user agent. These came from addresses that were also probing for WordPress and PHP files that docs sites don't have. For AI crawler names it was nearly 2 in 5.
of Claude's fetches were for docs pages that don't exist. For ChatGPT, it was 0.4%.
People click links. Agents often guess: plausible paths like /reference/how-the-api-works that were never there, or old URLs from before a page was renamed.
Coding agents read docs in tokens. As markdown, a docs page is half the size: 15 KB instead of 29 KB.
Claude Code is the fastest-growing AI assistant in this report, and every page it reads takes room in a context window that's also holding the developer's code. The lighter the page, the more of your actual content fits.
AI crawler fetches asked for llms.txt. It's a file for developers' tools, not for crawlers.
Nearly all llms.txt requests came from developers' own tools: scripts, command-line clients and browsers. 77% of docs sites on DeveloperHub publish one.
as many people read docs on a weekday as on a weekend day. AI assistants keep office hours too. Crawlers never sleep.
Every AI assistant fetch has a person behind it, so assistants follow the working week. Crawlers run flat around the clock, and are slightly busier at weekends.
The dashed line is 1×: the same every day of the week.
of AI crawler traffic in October came from Exa, an AI search engine that was barely in the logs in July. AI products move that fast.
Exa isn't alone. About 30 AI companies sent bots to read docs this quarter, and at least 8 of them first appeared partway through it, among them Reflection AI, Safe Superintelligence and the opencode coding agent. Each new AI product that reads your docs is one more place developers can find them.
more questions typed into the AI chat on docs sites in September than in July.
This is the chat box on the docs site itself, where a reader types a question and gets an answer from the docs.
How we measured
- Data. Requests to documentation sites hosted on DeveloperHub, 8 July to 7 October 2026, from server logs.
- People are counted once per session, not once per page view.
- AI assistants and crawlers are counted per page fetch, and identified by user agent.
- Fakes removed. Addresses using AI names while probing for attack paths are left out of every AI figure. The scanner figures come from them.
- Also left out: uptime monitors, our own rendering service, and sites served through a customer's own CDN.
- Rounding. Shares are rounded to whole percentages and may not add up to exactly 100.