Miriam Schwab

Head of WordPress @elemntor. Cofounder @StratticWeb. Opinions are my own and do not necessarily represent those of my employer.

[Miriam Schwab @miriamschwab]

Part of getting a site ready for AI agents is helping them get the facts on your pages right, and structured data is one of the main ways to do that.

Without it, an agent reading a product page has to work out from the layout which number is the price. Structured data labels those facts for it, usually in a block of JSON-LD that says what the product is, what it costs, and whether it's in stock.

There's one thing to watch for, though. Many of the tools agents use to fetch pages turn the HTML into plain text or markdown and drop JSON-LD on the way, so keep the facts that matter in the visible text too.

I go through this and the other agent-ready approaches catching on in my post on the Elementor blog 👇

https://t.co/RE9ZRoAQmm

[Miriam Schwab @miriamschwab]

I love @claudeai artifacts. They're a great way to run mini apps that are fully customized to how my crazy brain works and how I operate without having to worry about backends, hosting, deployment, security, etc. I have a few that I use on an ongoing basis and they're fab.

Anyway I introduced my daughter to them and now she's non-stop creating artifact apps for her work and personal health tracking and having the best time 😂

Welcome to the wild and wonderful world of AI my child 🤖

[Miriam Schwab @miriamschwab]

AI has made security a much bigger job for everyone who maintains software, and the WordPress security team is keeping up with it while getting better at the same time.

For about a decade, WordPress received 20 to 30 security reports a month. This past August it received 773! The team has since shipped 4 security releases in 41 days, with every fix going back to every branch still receiving security updates, all the way to 4.7. They also built an AI-powered system to help triage incoming reports, and they're now pointing the same tooling at core to find vulnerabilities before anyone reports them.

Thank you to the security team and everyone else doing this work for all of us 👏

Here's where it all stands, and how you can help 👇

https://t.co/BXQA0JzOPv

[Miriam Schwab @miriamschwab]

@jameswlepage 💯

[Miriam Schwab @miriamschwab]

Getting MCP into WordPress core is great, but it's only truly useful for end users if an agent can do something useful once it connects. That's why I'm so glad the WordPress AI team is talking about putting useful Abilities in core, starting with a set for everyday site management, so people can use AI in WordPress out of the box instead of adding extra plugins first.

The target is WordPress 7.3, and they're looking for people to help 👇

https://t.co/7DFrWYSkwj

[Miriam Schwab @miriamschwab]

@euvoor Best is to join the Make WordPress Slack community, then join the #core-ai channel and ask how you can help. I think they'll appreciate the extra hands!

[Miriam Schwab @miriamschwab]

@Lafe8088 I am tracking that.

[Miriam Schwab @miriamschwab]

AI agents are now regular visitors to websites, so site owners are starting to look at how to make their content easier for those agents to read. One approach that has caught on is offering markdown versions of your pages alongside the regular HTML.

The reason comes down to size. Every page an agent reads takes up space in its context window, and most of a page's HTML is stuff it doesn't need, like navigation, scripts, styling, and ads. In an example shared by @Cloudflare, one page dropped from 16,180 tokens as HTML to 3,150 as markdown, which leaves a lot more room for the actual question.

You can offer markdown at a separate .md address or through content negotiation on the normal URL, and it's a good idea to do both, so an agent gets markdown regardless of its preference.

I go through this and the other agent-ready approaches catching on in my post on the Elementor blog 👇

https://t.co/RE9ZRoAQmm

[Miriam Schwab @miriamschwab]

@Lafe8088 Creating markdown pages is simple - if you’re already creating one, create them all.

[Miriam Schwab @miriamschwab]

@heyamberhinds @andrewhoyer @elemntor Oh that is really not a helpful rule! Past participants would probably be grateful to get notified that another one is coming up, and anyone who doesn't want the notification can opt out.

[Miriam Schwab @miriamschwab]

WordCamp Canada is this Thursday and Friday, November 5 and 6, in Vancouver, and we at @elemntor are proud to be sponsoring it. The organizers built the event around connection, community, and contribution, and to get newcomers in the room with people who've been doing this for years, students get 50% off their tickets.

Community events are a big part of how WordPress keeps going, and we're proud and grateful to support this one. Thank you to the organizers and volunteers 🙏

https://t.co/gs5p1JG0fs

#WordPress

[Miriam Schwab @miriamschwab]

@claudeai I have a question about Claude Desktop. Since this morning, the ArtifactData tool disappears a few seconds into every session and gets replaced by a new built-in claude-artifacts server with its own database tools.

I have a daily scheduled task that updates an artifact’s database. It still works with the new tools, but it’s much slower, because the old tool could load a batch of writes from a file and save its reads to files, and the new tools can’t. Now every write has to be sent in full inside the request.

Is this switch permanent? If so, will the new tools get file support? Thanks!

[Miriam Schwab @miriamschwab]

@_the_faisal @jasontheadams Of course the Abilities will be structured and leveled.

[Miriam Schwab @miriamschwab]

The WordPress AI team has laid out the next big steps for AI in WordPress (written up by @jasontheadams), and it looks really promising! The MCP Adapter finally got into the plugin repo recently, and the plan is to bring it into core very soon 👏

That would mean an agent can connect to a WordPress site without having to install and configure a plugin first. More on this 👇

https://t.co/7DFrWYSkwj

[Miriam Schwab @miriamschwab]

@openwaf Hello AI bot commenter: yes, many agent visits pretend they're something they're not. That's why it's important to keep an updated list of verified IP ranges so you know which agent is real, and which is something else.

[Miriam Schwab @miriamschwab]

Being agent ready means letting the right agents in. A site would want to welcome a shopper's assistant and block a scraper, and today it's hard to tell them apart.

A bot only introduces itself with its user-agent string, which is a line of text it writes about itself. Nothing stops a script from calling itself GPTBot or Chrome. On my own site, one address called itself Chrome on a Mac, then node, then GPTBot, all within 12 seconds.

When names can't be trusted, sites end up blocking based on guesses, which backfires. An agent readiness scanner tested my site while pretending to be OpenAI's GPTBot, Cloudflare saw that the request hadn't actually come from OpenAI and blocked it, and the scanner then reported that my site blocks AI agents!

The fix in the works is Web Bot Auth, an IETF draft where an agent signs each request with a private key that the site can check against a public one. I wrote about where that stands on the Elementor blog 👇

https://t.co/mU5NVYWJD4

[Miriam Schwab @miriamschwab]

There's agents that change the site for builders, and that's where the Elementor MCP and Angie come into play:

https://t.co/Xn24iXR11B
https://t.co/9O2oDbCQK2

Then there's external agents that interact with your site to try to understand it, query it and even act on it. These are the agents that post is about.

[Miriam Schwab @miriamschwab]

There's not an official rulebook for agent-ready sites, but a handful of proposals have made it off the drawing board and are being implemented on a lot of sites. In my new post on the Elementor blog, I went through each one, how far along it is, and what we know about how agents use it.

llms.txt and markdown versions of pages help agents read, while structured data and MCP servers help them query. Discovery files tell them what's there, Content Signals covers part of the trust question, and WebMCP is aimed at taking action.

More on what's catching on, and what you can do now 👇

https://t.co/RE9ZRoAQmm

[Miriam Schwab @miriamschwab]

AIs need structure and direction. The structure comes from the platform it's working with, and WP has quite an ideal structure for these purposes.

Direction can be via skills, harnesses, or other documented instructions; or via the human in the loop, no matter what you're working on or with. WP isn't unique in that.

[Miriam Schwab @miriamschwab]

Ask an AI model to build a WordPress plugin and it usually knows what to do. That's because WordPress is open source, so its code, documentation, support forums, and roughly 70,000 plugins have been public for years, along with two decades of tutorials and Stack Exchange answers. The models trained on all of it.

Closed platforms don't get that advantage, since a model can't learn an API whose docs sit behind a login.

That's just one example of how WordPress was born ready for the agentic web. See more on that in my post 👇

https://t.co/DmBXFSj5Qt

[Miriam Schwab @miriamschwab]

@lozmatic I wish more people would write about WordPress' strengths, even in this AI era (or especially in this AI era). Oh well. In the meantime, I'll keep writing and hope it makes a dent.

[Miriam Schwab @miriamschwab]

@legare_dominic @natmiletic If you use our MCP you can cut many hours down to minimal hours. Have you tried it?

[Miriam Schwab @miriamschwab]

@andrewhoyer @elemntor Oh no I can’t believe you all didn’t know about this! I wonder why it was barely promoted 😞