Blog / Agent experience (AX)

How AI assistants pick which developer docs to cite: 80 answers from Claude and Codex

If developers ask an AI assistant instead of a search engine, your docs need to be what the assistant cites. We watched two assistants answer the same 20 developer questions, and they read almost entirely different parts of the web.

Leonid Bugaev · Published 21 Oct 2026

What we asked

20 developer questions, such as “How to test an MCP server with AI agents”, “How many tools can an MCP server have before agents get worse?” and “How do I make my API agent-ready?”. Each went to Claude (in Claude Code) and Codex, once without web access and once with web search: 80 answers, all returned. Neither assistant knew who was asking.

Two assistants, two webs

In 11 of 20 questions the two assistants' cited sites didn't overlap at all. Of 124 cited sites, 91 were cited by only one answer. Optimising for “AI search” means optimising for at least two different behaviours.

Codex: site-restricted search

• Searched in all 20 with-search answers: 94 searches, median 5 per answer.
• 78 of 94 were restricted to one site: modelcontextprotocol.io 24 times, anthropic.com 7, code.claude.com 5, then framework and vendor docs.
• Its 103 links: 53 vendor sites, 26 spec sites, 21 GitHub, 1 elsewhere. No directories, papers or community posts.
• Queries were short (about 6 words) and only 1 had a year.

To get into a Codex answer, your page has to be on a site Codex already thinks to target: the spec, the big platforms' docs, GitHub.

Claude: cite from the result list

• Searched in 18 of 20 with-search answers, usually 2 searches, about 9 words each; 17 of 29 ended in “2026”.
• About 9 results per search (258 total); 43 had “2026” in the title or URL.
• 133 of 144 links came straight from results it never opened; 8 from pages it opened; 3 from memory.
• A second query often named tools from the first results, so a product enters the answer by being on the first result page.
• For the two Claude Code questions it went straight to code.claude.com.

What gets cited

Pages with concrete numbers were picked over opinion: a server's token count, an app's tool limit, “20 of 20 runs”. On “which tool?” questions, vendor-written “best X 2026” roundups shaped the picks.

What to do with your docs

1. Put a number in the title and first lines. That's what gets quoted from a result snippet.
2. Date it. Claude adds the year to queries; fresh titles match.
3. Use the question's words in the title and H1.
4. Get onto sites Codex targets: contribute to spec docs or examples, framework integration pages, GitHub.
5. Serve clean Markdown and an llms.txt, for assistants that do open pages. See what agents use: llms.txt, AGENTS.md, MCP.
6. Re-run the same questions after publishing to see whether anything changed.

Go deeper: Agent experience (AX), Agent-ready tools, Pass rates with ranges.

Back to agent experience (AX).

How we measured

• Run: 20 questions × Claude and Codex × with and without search = 80 answers, 21 Oct 2026, 18:56–19:22 (Istanbul time, UTC+3), one sample each.
• Extraction: searches, opened pages and cited links from each assistant's own transcript; brands by an automated extractor capped at 24 names per answer.
• Limits: one sample per question and mode; Codex doesn't expose its result lists, so “seen in results” is Claude-only; no Google AI Overviews, ChatGPT or Gemini in this pass.