How I Connected AI to WordPress (Starting With a Safe Sandbox)

I have been thinking for a while about how useful it would be to manage website content with AI without handing over the keys to everything at once. This week, I finally tested the first practical step: connecting AI to a WordPress website safely.

Rather than trying it on every site I own, I used Lee Read as a sandbox. That gave me room to test the connection, see what the AI could actually read, and make sure any new content stayed under my control.

Why I started with a sandbox site

AI can be genuinely useful for content work, but it should not be allowed to publish, delete, or rewrite things without a clear process. A test site is the right place to build that process.

For this first experiment, the rules were simple:

• AI can read posts, pages, categories and tags.
• AI can search for related articles before suggesting something new.
• New articles are created as drafts only.
• Nothing is published without my approval.
• Nothing is deleted without my approval.

That sounds obvious, but it matters. The aim is not to replace judgement. The aim is to remove the repetitive parts of content management while keeping the final decision human.

How the connection works

WordPress has something called a REST API. In plain English, it is a safe doorway that lets another tool ask WordPress for information or, when authorised, create content.

Instead of using my normal WordPress password, I created a separate Application Password just for this connection. It can be revoked on its own if I ever want to stop using the system. That is far safer than giving an outside tool my main login details.

Once connected, the AI was able to confirm basic site information, look through existing posts, and create a clearly labelled test draft. The important part is that the draft stayed a draft. Nothing appeared publicly on the site.

What I want AI to help with next

The connection itself is only the foundation. The more interesting part is what it can support once the workflow is reliable.

• Checking whether I have already covered a topic before writing another article.
• Creating a useful first draft with headings, a sensible structure and an excerpt.
• Suggesting categories, tags, a URL slug and an SEO title.
• Finding relevant internal links to articles already on the site.
• Helping me refresh older content without losing my own voice.

For example, I have already written about the future of SEO in a post-search-engine world. A better workflow does not simply produce another article on the same point. It should understand that existing piece, find a useful angle that adds something new, and link the two where it helps the reader.

The line I do not want to cross

There is a big difference between AI helping with a draft and AI taking control of a website. I want the first, not the second.

A good system should make the boring work easier: searching old articles, structuring ideas, spotting duplicate topics and preparing drafts. But it should still stop before publication and ask for a review.

That gives me the benefit of speed without turning the website into an unattended content machine.

What this could mean for the future

If this works well on Lee Read, the same safe setup can eventually be reused across my other websites. The goal is not to build a completely separate system for every site. It is to have one sensible way to connect each site, keep its credentials separate, and work with content in a consistent way.

For now, though, this is exactly where it should be: a small, controlled test. One site. Drafts only. A clear record of what the AI has done. That feels like the right way to start.