Leaderboard / Headless CMS
Headless CMS: what the agents install
Which headless CMS gets installed? 30 answers from 3 models to 5 build tasks, run 2026-11-02. Sanity was installed first in 37% of answers.
Install share by product
| Product | All models | 95% band | Claude | GPT | Gemini | Mentioned |
|---|---|---|---|---|---|---|
| Sanity | 37% | 22%–54% | 20% | 60% | 30% | 47% |
| Git-based CMS | 13% | 5%–30% | — | 10% | 30% | 30% |
| MDX files (Contentlayer/Velite) | 10% | 3%–26% | 20% | — | 10% | 13% |
| Strapi | 10% | 3%–26% | 20% | 10% | — | 10% |
| Storyblok | 3% | 1%–17% | — | — | 10% | 10% |
| Directus | 3% | 1%–17% | — | 10% | — | 7% |
| Puck | 3% | 1%–17% | — | — | 10% | 3% |
| Payload | 0% | 0%–11% | — | — | — | 10% |
| Contentful | 0% | 0%–11% | — | — | — | 3% |
| None / built by hand | 20% | 10%–37% | 40% | 10% | 10% | — |
Install share counts the product each answer installed first. Mentioned counts answers that named it anywhere, picked or not. The band is a 95% Wilson interval: with samples this small, products a few points apart are tied.
Prompt by prompt
Each chip is one answer (2 samples per model). None means the model built it by hand or didn't choose.
“Add a CMS to my Next.js marketing site so non-developers can edit pages and blog posts. Give the exact install commands and the key code.”
ClaudeSanitySanityGPTSanitySanityGeminiSanitySanity“Build a blog for my Astro site where the marketing team can write posts without touching code. Give the exact install commands and the key code.”
ClaudeNoneNoneGPTGit-based CMSNoneGeminiGit-based CMSGit-based CMS“I need a headless CMS for a multilingual company website built with Nuxt. Give the exact install commands and the key code.”
ClaudeStrapiStrapiGPTDirectusStrapiGeminiSanityStoryblok“Add editable content (landing pages, testimonials, FAQ) to my React app with a nice editing UI. Give the exact install commands and the key code.”
ClaudeNoneNoneGPTSanitySanityGeminiPuckNone“Set up a CMS for a docs and changelog section in my Next.js SaaS. Give the exact install commands and the key code.”
ClaudeMDX files (Contentlayer/Velite)MDX files (Contentlayer/Velite)GPTSanitySanityGeminiGit-based CMSMDX files (Contentlayer/Velite)
What the models wrote
ser-friendly for non-developers.
### 1. Install & Initialize Sanity
Run these commands in your Next.js project root:
```bash
npm install next-sanity @sanity/image-url @sanity/vision sanity styled-components
npx sanity@latest init --env
```
*(When prompted, choose "Yes" to use TypeScript, "Yes" to embed the Studio in your Next.js app, and route it to `/studio`.)*
### 2. Define Your Content Schema (Blog Post)
Creadatabases. Perfect for Docs and Changelogs.
### 1. Install & Initialize
Run this in the root of your Next.js project:
```bash
npm install outstatic
npx outstatic init
```
*This command automatically creates the `/app/outstatic/` routes and API endpoints for the CMS dashboard.*
### 2. Configure Collections
Start your dev server:
```bash
npm run dev
```
Go to `http://localhost:3000/outstatic`.
1. Log in with yourcontrolled, requires no external databases, and is standard for Next.js docs/changelogs.
### 1. Install dependencies
```bash
npm install next-mdx-remote gray-matter
```
### 2. Create your content directories
Create these folders at the root of your project:
```bash
mkdir -p content/docs content/changelog
```
Add a sample file `content/docs/getting-started.mdx`:
```mdx
---
title: "Getting Started"
date: "2023-10-Questions
Which headless cms tool do AI coding agents install most?
Sanity. It was installed first in 11 of 30 answers (37%) from Claude, GPT and Gemini, run 2026-11-02. Next: Git-based CMS 4, MDX files (Contentlayer/Velite) 3, Strapi 3.
What do Claude, ChatGPT and Gemini pick for headless cms?
Claude: Sanity (2 of 10); GPT (ChatGPT): Sanity (6 of 10); Gemini: Git-based CMS (3 of 10).
Which headless cms tools do AI agents mention but not install?
Named but never installed in this run: Payload (3 mentions), Contentful (1 mentions). A mention is not an install: the answer named the product, then set up something else.
How was this measured?
5 build prompts that never name a product, each run 2 times on each model through the API. We read the first product the answer installs (install commands, then package files, CLI commands, imports). 30 answers is a small sample: products a few points apart are tied.
Is a tool missing, or a pick wrong?
A tool we missed, a pick that looks wrong, a prompt we should add? Tell us. We read every note and fix the data in the next run.
Is your product losing these picks?
Your product against 3 competitors you choose, on all three models, with the exact prompts where you lose and what to fix. $19, delivered within 24 hours.