Web Design
Generating Site Content with AI Without It Reading Like AI
Part 4 of the thread Building a site for someone
- My old setup asked a model for a full post as JSON and published it straight to WordPress. The publishing step is the part I'd cut.
- Keep the boring checklist: titles under 60 characters, meta descriptions of 155-160, test every code sample.
- For a client site, the facts come from the owner and the discovery answers. The model only helps shape them.
- A human reads every page before it goes live. That's the rule that makes the rest safe.
Somewhere in my WordPress notes is a PowerShell pipeline for writing posts without writing posts. It sent a topic, a tone and a word count to a language model, asked for the post back as JSON with a title, content, excerpt and meta description, and pushed it straight into WordPress through the API. A config file listed topics, and a scheduled task ran the whole thing. There was even a daily news roundup built from RSS feeds and a scoring formula for "engagement."
It worked, technically. Reading it again, I can see exactly why sites built that way all sound the same.
What I'd throw out
Publishing without reading. The pipeline's prompt asked for an "engaging introduction," a "strong conclusion" and something "engaging and valuable," then published whatever came back. That's how you get pages that open with a throat-clear and close with a summary of what you just read. Nothing in the loop could tell a good post from a bad one.
Content for the calendar's sake. Auto-posting a daily roundup of other people's headlines fills a blog without giving anyone a reason to visit it. A small business site with four real posts beats one with four hundred generated ones.
The engagement score. It weighted page views, time on page and shares into one number. None of that data existed yet when the posts were being written, so the score only ever judged work after the fact. It never made a post better.
What I'd keep
The other half of the notes was a plain checklist, and most of it still holds:
| Keep | Why |
|---|---|
| Titles under 60 characters | Search results cut off longer ones |
| One focus phrase per page | A page about everything ranks for nothing |
| A lowercase, hyphenated slug | Readable, and it survives being pasted anywhere |
| Meta description of 155-160 characters | It's the sales line under your link in search results |
| Test every code sample | Generated code often looks right and fails |
| Accessibility check | Headings in order, alt text on every image |
| Proofread before publishing | The step the pipeline skipped |
That's a checklist for a person, and it's a good one.
How I'd use AI on a client site now
The problem with generated site copy isn't that a model helped. It's that nobody put real facts in, so the model filled the gaps with the average of every other website.
So the facts come first, from the people who have them. The discovery checklist already collects most of what a page needs: what the business does, who it's for, what makes it different, what the site's one job is. The owner's own words, even rough ones from an email, are worth more than anything a model invents.
Then the model is useful for the dull parts: turning a page of notes into a first draft, suggesting three versions of a headline, trimming a meta description to length, or pointing out where a paragraph repeats itself.
Tip
Give the model the owner's actual phrases and tell it to keep them. "We've been fixing lawn mowers in the same garage since before the highway went in" is better than anything it'll write about "a trusted local partner."
And then a person reads every page out loud before it goes live. The tells are the same ones I listed for emails in keeping AI drafts from sounding like AI: stock openers, summary closers, triples on autopilot, words nobody says out loud. Site copy adds one more. If a sentence could sit on any business's homepage unchanged, it's not saying anything.
Heads up
Never let a model invent numbers, reviews or credentials for a client. "Hundreds of happy customers" is a claim somebody has to stand behind.
Where it fits in the build
Content lands in the middle of the project timeline, between the approved design and the staging siteA private copy of a website where changes get built and approved before they go anywhere near the real one.More: From Kickoff to Launch: A Realistic Website Timeline, and it's the milestone that waits on the client. AI can shorten that wait by turning scattered notes into drafts faster. It can't replace the notes. If the business hasn't said what makes it different, no model can find that out for them.