0 of 193 sites used WordPress: AI coding agents are now picking your software
Seven coding agents were asked to build websites for bakeries, clubs and small businesses. Not one of the 193 sites ended up on WordPress, even though it powers around 40 percent of the web. Together with other studies, the picture is clear: more and more, the agent decides which software wins, not the human giving the instructions.

What this is about
Getting a website used to go like this: the baker googled “create a homepage,” landed on a web host or a site builder, and usually ended up on WordPress. According to W3Techs, it powers around 40 percent of all websites. Today she's more and more likely to type “Make me a site for my bakery” into a coding agent, an AI like Claude Code or Codex that creates files, installs programs and writes code on its own. From that moment on, the agent decides on the software.
Several measurements from recent months show the same pattern: the agent has become the buyer. It picks the host, the payment provider and the database, usually without anyone asking. And it practically never picks WordPress.
193 websites, zero WordPress
Andy Peatling spent 17 years at Automattic, the company behind WordPress.com, and now builds an AI design agent called Miles. He has published an open benchmark, with code and raw data on GitHub and a dashboard at wpab.view.fast. He presented the results at WordCamp Canada. According to the methodology page, the October run used seven agents with default settings: Opus 5.5 in Claude Code, GPT-6.1-Sol in Codex, Gemini 3.8 Flash in Antigravity, Grok 4.7 in Grok Build, Kimi K3 in Kimi Code, and DeepSeek V4 Pro and GLM 5.3 in OpenCode. They got 28 short, everyday requests with no platform specified, including a few control cases where WordPress isn't a fit at all.
The 19 requests that would have suited WordPress produced 193 websites. Not one ran on WordPress. According to the WPMore newsletter, the agents wrote custom code 93 percent of the time, “mostly plain HTML or React.” Asked what else they would have considered, they named WordPress in 51 percent of cases. But according to Peatling's own analysis, the agents brought it up during the actual decision for only 9 sites.
The raw data shows what this looks like in practice. Claude Code justified a plain-HTML bakery site like this:
“The folder was empty and the request was short, so I went straight to the simplest option.”
Codex built a React app for the same kind of request, because React makes “menu filters and a cart straightforward.” Antigravity went straight for React 19, Vite and Tailwind for a “highly interactive” cake configurator with live pricing. We tallied the first 42 runs in the raw data ourselves. The result is an approximation, since some runs built more than one thing:
| What was built (sample, first 42 runs) | Count |
|---|---|
| Plain HTML/CSS/JavaScript | 17 |
| React (mostly with Vite) | 11 |
| Node.js server | 3 |
| Astro, Eleventy, Decap CMS (each) | 2 |
| Hosted/Shopify-based | 2 |
| Next.js / custom generator (each) | 1 |
| Nothing built | 1 |
The reasons, and which ones hold up
Peatling's analysis lists the reasons the agents gave against WordPress after the fact. A site could have several reasons:
| Reason (share of 193 sites) | Fact check per benchmark |
|---|---|
| Hosting and maintenance: 82% | – |
| Security: 66% | partly true: plugins accounted for 91% of vulnerabilities in 2025, per Patchstack |
| Content should live in Git: 46% | – |
| Needs PHP and MySQL servers: 44% | misleading for local setup (Node and SQLite are enough), true for production |
| Cost of paid hosting: 42% | misleading: WordPress.com has a free plan |
| Slow without caching: 30% | partly true |
The most important number isn't in this table. For 80 percent of the sites, the agent never asked who would maintain the site later. That would be the strongest argument for a CMS (content management system, software with an interface for editing text and images). According to the benchmark, it would have been the deciding factor for 54 percent of the sites.
There's a methodological weakness, though. Five of the seven agents ran on a developer machine without PHP or a database. Some promptly cited that as a reason: “Only Node.js and Python were installed in this environment.” Claude Code and Codex, by contrast, worked in their makers' cloud environments, where PHP was available. They never cited it as a reason and still didn't pick WordPress a single time. According to WPMore, several replies on X questioned Peatling's test setup. And Peatling is an open WordPress advocate.
Even when WordPress is already there: 0 of 42
The second test looks at existing sites. Claude Code and Codex were asked to handle 21 typical site-owner requests, from “Get my business online” to “Migrate my site.” That's 42 runs per access method. The report was published by Respira:
| Access method | Tasks completed |
|---|---|
| Official WordPress MCP adapter | 0 of 42 |
| Owner's application password | 33 of 42 |
| Respira plugin | 33 of 42 |
MCP (Model Context Protocol) is a standard AI agents use to call a piece of software's functions. The adapter makes WordPress functions, called Abilities, available to them. But according to the WordPress developer blog, only basic functions like page and user info are included by default, and even those aren't automatically enabled for MCP. The agents connected, gave up after three or four calls, and simply had no tool for the actual task. The password works because it grants every permission at once. The agent then made 677 changes in 37 of 42 runs, a security risk of its own. The report's author sells Respira, one of the three methods being compared. He discloses that.
According to The Repository, the WordPress AI team wants to build MCP into WordPress core with version 7.3. Matt Mullenweg reportedly agreed, asking that they do everything possible to prevent new security holes.
The agent as buyer: the data at a glance
| Study | Key finding |
|---|---|
| Armature, Sept. 2026 | 16,893 sessions, 5,292 analyzed; Claude Code, Codex and Cursor pick the same tool only 42% of the time |
| Amplifying: Claude Code, Feb. 2026 | Vercel as host for JS frontends in 86 of 86 cases; AWS, Google Cloud and Azure 0 first picks in 112 responses |
| Amplifying: Codex vs. Claude Code, March 2026 | Codex picks Cloudflare Workers 49% of the time, Claude 9%; Claude prefers Vercel |
| Vercel, April 2026 | over 30% of deployments triggered by agents, 75% of them by Claude Code |
Vercel's April 9 blog post cites more than 30 percent, up 1,000 percent in six months. It now even points to “more than 50 percent” in September. The breakdown by agent isn't in the blog post but in Vercel's social media post at the time, as reported by Sacra and CodeZine.
How the agent chooses
The studies point to four levers:
- The project's language. At Armature, the same email task went to Resend in TypeScript, SendGrid in Python, Postmark in Go and Azure in Java. Vercel won every Next.js project and was never recommended in Python projects.
- The fine print on pricing pages. Mailgun lost to Postmark after agents read that the free tier only retains data for one day. Supabase lost pure database tasks to Neon because its pricing bundles extras.
- Training vs. web search. According to Armature, Codex searched the web in 94 percent of sessions, Claude Code in only about 30 percent. So Claude relies more on what it learned in training, and in 19 percent of cases it builds something itself instead of integrating a product. One critic on Hacker News argues this may have less to do with the model than with Claude Code asking for permission before every web search.
- Friction at the start. Drupal founder Dries Buytaert had Claude Code recommend a tech stack. Drupal was the best fit on paper but only came third because of “risk within the session,” behind Next.js and WordPress. If there's friction in the first few minutes, you lose.
Brand recognition alone doesn't help: PayPal was mentioned 139 times and never chosen, while Stripe won 124 of those sessions. LangChain got 194 mentions but only 4 picks, Netlify 152 mentions and only 6.
Who's already cashing in
Armature is a Y Combinator startup that sells vendors exactly this: getting “discovered and used by coding agents.” By its own account, the study is part of that work, and every judgment was made by an AI model acting as referee. One Hacker News user slammed the business model as manipulation. The co-founder countered that otherwise only the incumbents from the training data would get picked. Tools like Promptwatch now measure how often Claude Code and Codex recommend a brand. Google engineer Addy Osmani calls the discipline “Agentic Engine Optimization,” SEO for agents instead of humans. Among other things, he recommends an llms.txt, a plain-text content overview made specifically for AI models.
So far there's no evidence that this changes what gets picked. The paper Osmani cites on nine coding agents does put an llms.txt on the test server, but doesn't report whether any agent fetched it. According to Ahrefs, 97 percent of roughly 38,000 llms.txt files got zero requests in May 2026. Notably, Claude Code was the second most active named AI bot among those that did fetch them. Promptwatch itself says llms.txt has had no measurable effect so far. Peatling is focusing on it anyway. His key recommendation is to publish a one-command WordPress setup in llms.txt, the docs and on npm.
What this means for the bakery
A site built in plain HTML or React has no admin interface. If the baker wants to change her opening hours, she needs a developer, or the agent again. The French-language portal info-lux therefore advises testing before you commission a site whether you can change a price, an opening time and a photo without AI. You should also register the domain in your own name and work out the costs for twelve months. None of the tests checked whether the agents remember legally required German pages like the Impressum (legal notice) and the privacy policy. The requests were in English.
Web hosts are responding in different ways. On November 16, Cloudways (DigitalOcean) launched a builder that turns a single sentence into a WordPress site made of standard Gutenberg blocks. The company says it worked in 98 percent of cases during testing, in about four minutes, starting at $7.99 per site per month plus hosting. GoDaddy released the Airo AI Builder the same day. It generates websites and web apps with their own database, and you can download the code. The press release doesn't mention WordPress.
Our take
Our take: the individual numbers are open to challenge. Armature sells agent marketing, Respira sells a plugin, Peatling champions WordPress, and some of the test environments are artificial. But the studies come from different directions and reach the same conclusion. Agents have firm preferences. They differ a lot from one another, and they decide based on what meets the least resistance in their sandbox, not on what the customer will need in two years. Peatling's line “Code is the path of least resistance” sums it up well.
For software vendors, this means a one-command setup counts for more than any campaign, even against 20 years of brand building. For non-technical users, things are trickier. They get a slick site fast, but a dependency comes with it. It's an open question whether agent makers will adjust course. An agent that asks “Who will maintain the site later?” would likely reverse some of these decisions. But there's no evidence for that yet.
Sources
- GitHub: apeatling/wp-agent-benchmark github.com
- WordPress Agent Benchmarks – Methodik wpab.view.fast
- Benchmark-Auswertung Oktober 2026 (advice.md) raw.githubusercontent.com
- Benchmark-Rohdaten Oktober 2026 (runs.json) raw.githubusercontent.com
- Respira: WordPress agent benchmark respira.press
- WPMore Issue 65 wpmore.substack.com
- Armature: Which tools do Claude Code, Codex and Cursor choose? armature.tech
- StackOne: Coding agents disagree on tool choice stackone.com
- Amplifying: What Claude Code picks amplifying.ai
- Amplifying: Codex vs. Claude Code picks amplifying.ai
- Vercel: Agentic Infrastructure vercel.com
- Addy Osmani: Agentic Engine Optimization addyosmani.com
- arXiv: Developer Experience with AI Coding Agents arxiv.org
- Ahrefs: llms.txt-Studie ahrefs.com
- Promptwatch: Coding Agent Visibility promptwatch.com
- The DropTimes: Dries Buytaert zu Drupal und Coding-Agenten thedroptimes.com
- WordPress Developer Blog: MCP Adapter developer.wordpress.org
- PR Newswire: GoDaddy Airo AI Builder prnewswire.com
- DigitalOcean: Cloudways AI Website Builder investors.digitalocean.com
- info-lux: Créateur de site IA info-lux.com
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.