Gap: both agent prompt files are served with Content-Type: text/markdown. ChatGPT's web reader refuses to open that type. Observed twice on 2026-09-16 against nextcommerce.com/agent-eval/prompt.md (same content type, same behaviour): "the web reader here is rejecting that endpoint because it serves text/markdown directly". The URL was confirmed 200 to ChatGPT-User, OAI-SearchBot and GPTBot user agents, so it is the type, not access.
Codex CLI, Claude Code and Cursor fetch these fine (plain HTTP clients), which is why the persona × agent probes never hit it. ChatGPT web is the surface a prospect on the marketing site reaches through the hero strip, and the www consultation file hands the developer persona to /evaluate/prompt.md, so a ChatGPT user gets one hop and then a wall.
Fix: serve /evaluate/prompt.md and /agent-setup/prompt.md as text/plain; charset=UTF-8, as llms.txt and the domain bundles already are. Paths unchanged; they are referenced from llms.txt and from the www file. nextcommerce-site did the same in NextCommerceCo/nextcommerce-site#176 with a Netlify header rule.
Verify: curl -sI https://developers.nextcommerce.com/evaluate/prompt.md | grep -i content-type.
Gap: both agent prompt files are served with
Content-Type: text/markdown. ChatGPT's web reader refuses to open that type. Observed twice on 2026-09-16 againstnextcommerce.com/agent-eval/prompt.md(same content type, same behaviour): "the web reader here is rejecting that endpoint because it servestext/markdowndirectly". The URL was confirmed 200 toChatGPT-User,OAI-SearchBotandGPTBotuser agents, so it is the type, not access.Codex CLI, Claude Code and Cursor fetch these fine (plain HTTP clients), which is why the persona × agent probes never hit it. ChatGPT web is the surface a prospect on the marketing site reaches through the hero strip, and the www consultation file hands the developer persona to
/evaluate/prompt.md, so a ChatGPT user gets one hop and then a wall.Fix: serve
/evaluate/prompt.mdand/agent-setup/prompt.mdastext/plain; charset=UTF-8, asllms.txtand the domain bundles already are. Paths unchanged; they are referenced fromllms.txtand from the www file. nextcommerce-site did the same in NextCommerceCo/nextcommerce-site#176 with a Netlify header rule.Verify:
curl -sI https://developers.nextcommerce.com/evaluate/prompt.md | grep -i content-type.