Making WordPress sites agent-ready

The more people visit a site, the more likely an agent shows up. On the busiest sites, it is close to certain.

Within that, reads concentrate on a few sites. 2,775 sites got 100 or more.

The people who used to search for a business now ask an assistant about it. The assistant follows them to the same sites.

This one we did not expect.

More than half of the AI traffic came from builders: owners who connect a coding agent to their site and let it edit pages, update content and change settings.

Codex sent almost 9,000 requests per site it touched. That is not a visitor checking opening hours. That is an agent rebuilding a site.

And it grows every week: sites with a coding agent working through MCP went from 29 a week to 277 in eight weeks.

When an assistant reads for a visitor, it reads what a person would. 86% of its requests went to home pages and inner pages.

llms.txt was asked for 317 times. In four weeks. Across 27,088 sites. WordPress even links a machine interface, the REST API, from 96% of home pages, and assistants barely touch it.

So the page is the product. A price inside an image, opening hours inside a slider, a business with no address in its markup: whatever the page hides, the agent never sees. We found the same across the wider web in What agents actually reach.

Traffic shows where agents go, not whether they get what they came for. So we sent agents to 40 Elementor sites with five tasks: decide whether to buy, list the newest items, quote a page word for word, start an order or a booking, and ask what the site offers the owner’s own assistant. Claude and Claude Code each ran all 200.

They fully answered 39% of requests and partly answered another 34%. Deciding whether to buy went best, at 57%. Asking what the site offers an owner’s assistant went worst, at 12%.

When links run out, agents guess an address. In at least 46% of Claude Code’s runs, it asked for a page that did not exist. The most common misses were the files agents expect a site to publish.

Meanwhile, 21 of the 40 sites run an MCP server. It lives under /wp-json/, where no visiting agent thinks to look. The door was there. Nobody knew the address.

That costs more than a missed visit, because an agent that cannot read a site answers anyway. In an earlier study of 1,056 businesses, answers built from other sites were 3.7× as likely to contain none of the facts the user asked for.

These sites already had a better format for agents: ask any page for markdown and you get a clean copy, about 20 times lighter than the HTML. 96% of them serve it, against 23% of everyone else in ora.ai’s index.

Agents never found it. None of the 275 home pages we checked said it existed, and across 400 agent runs it was used zero times.

The fix is a line in the page head and an HTTP header. Done once at the platform level, it reaches every site.

Before the release, we ran ora.ai’s checks on the 16,586 Elementor sites in its index. The pattern is sharp. Elementor sites passed nearly everything the host provides, and missed much of what each site provides.

Without those, an agent landing on the page has to guess what the business is, where it is and how to reach it.

This is the starting line. These scans predate the release, and every gap in the lower half of that chart is something one platform can close for every site at once.

For every request an assistant made for a person, the same company’s crawlers made many more.

Most of what AI companies take from WordPress is still training and indexing, not live reading. That is why reading and training should be separate choices: a site can welcome the assistant answering a customer without opting into the crawler that trains on it.

AgentReady is the open standard Ora launched for what a site should offer agents, and the Ora ranker is how a site measures itself against it: every check, with the fix for each one that fails. Here is what each finding calls for.

Owners already hand their sites to agents, so a fix does not have to be a to-do list. It can be a prompt the owner’s agent runs.

Elementor, the AgentReady committee member for WordPress, is bringing these fixes to its sites with a readiness score powered by the Ora ranker (announcement).

Everything above is the baseline. These fixes can reach more than 23 million websites at once: a rare chance to measure a change across a whole platform. In a follow-up study we will rerun the same traffic study and the same agent tasks on the same sample, and report what moved.

In the meantime, see where your own site stands. Scan a WordPress site on Elementor with its checker, or any site on the full ranker at ora.ai: every check in this post, with the fix for each one that fails.

• Traffic. Four full weeks, October 12 to November 8, 2026 (UTC), from Elementor’s edge logs, on a sample of 27,088 custom-domain sites hosted by Elementor, chosen by Elementor from the sites that sent traffic. Every figure in this post describes that sample. The 142,387 elementor.cloud subdomains are left out; the median one had 15 requests.
• A live agent read is a page (home page, inner page or file) returned to an assistant fetching it for a person.
• Scanners removed. A scanner posed as AI assistants and crawlers while probing for secret files. We removed it from every figure. On these sites it sent almost all of the traffic carrying Perplexity’s and Meta’s fetcher names, and about two thirds of what claimed to be ChatGPT.
• Builders are Codex, Claude Code and Cursor, MCP and API calls from assistants, and the Claude desktop app on sites where it also worked in wp-admin or the site API.
• People are browser requests with no bot signature. Leaving out the 1% of sites with the most browser traffic, assistants are 0.208% of people and agents.
• Samples. MCP figures come from 20% of sites and the crawler cleanup from 1% (286 sites). Every other traffic figure counts every site in the sample. The MCP trend covers September 8 to November 2.
• Upper bound. User agents can be copied, and this traffic arrives without IP verification, so assistant figures are an upper bound.
• Scores cover the 16,586 Elementor sites in ora.ai’s index, mostly from one scan run in September. The markdown comparison uses the other 114,371 sites ranked on November 4. That baseline is not a random sample of the web: most of it is self-submitted, and it leans toward software companies.
• Journeys. 400 read-only runs on 40 Elementor-hosted sites on November 4: five tasks per site, each run through Claude, as consumers use it, and Claude Code. The agents were told not to submit anything, sign in or contact anyone. Claude Code’s fetches through its shell are not all graded by status code, so its 404 share is a minimum.