How I write category page texts with AI (without the AI slop)

Author: Katarina Dahlin · Senior Growth Hacker & SEO Consultant
Updated: 2026-11-17 · Reading time: ~5 min

You can generate category page texts with AI in minutes. The hard part is getting texts that are true, useful to customers, and quotable by AI search, and getting them without checking every sentence by hand.

So I didn’t write a better prompt. I built a system. This is how it works, and the principles behind it apply to any e-commerce site.

Start with one rule

The most important rule turned out to be the simplest: a sentence without a fact gets deleted.

Not rewritten. Not “made more engaging”. Deleted.

“A wide selection for every need” is deleted. “A clear polycarbonate sheet lets through about 88–89 % of light, a white one about 30 %” stays. The second one is the only sentence a customer can actually use, and it’s the kind of sentence ChatGPT or Google’s AI Overviews can quote.

A simple test: remove the key word from the sentence. If the sentence loses information, keep it. If not, it was just praise.

Everything else in the system exists to enforce this rule at scale.

Separate the roles

How do you know AI-written text is true? You don’t, so I built the system around that.

  • A writer agent writes, but only from the client’s own product data. No general knowledge, no “industry facts”, no competitor content.
  • A checker agent checks every sentence against fresh product data and marks it as supported, needs a human, or wrong, with the source quote next to it.
  • Scripts reject text that breaks the rules automatically.
  • A human makes the decisions that need judgment.

In the first test run, 966 sentences were supported, 73 were flagged for a human, and 1 was wrong. The fixing rule mattered as much as the check: if in doubt, make the claim narrower or delete it. Never add.

What the writer agent is told

Answer first. AI search doesn’t read your page the way a person does. It grabs one section, sometimes just one sentence, and answers with it. So every section starts with the direct answer to its heading, followed by proof, a fact or number from the product data.

Every sentence stands alone. No “as mentioned above”, and no “it” or “these” pointing back. Product names are written in full, every time. It makes the text a bit repetitive for a human reader, but it’s the price of being quotable.

Don’t infer. No synonyms the data doesn’t use, and no generalising one product’s feature to the whole category. If it’s not in the data, it’s not in the text.

Don’t write what changes. Before writing a sentence, the question is: will this still be true in three months? Prices, availability and “our most popular model” change, and the product listing already shows them.

Make it the retailer’s own. If you ask AI to write a category text, you get the same text your competitor gets. The fix is to include information only this retailer has, such as its own brands and what they offer in that category.

Turn every lesson into a rule

Every time I spotted a problem, I asked myself how to make sure it never happens again. Writing it in a prompt isn’t enough, because prompts drift and get forgotten.

So every new rule went into the checking script, the writing guide and every agent the same day. A text that breaks a rule gets rejected automatically, no matter who or what wrote it. Feedback that becomes code compounds. Feedback that stays in your head repeats.

Use humans for judgment

I didn’t want to read every warning. The clear cases were fixed automatically. The rest came to me as a short numbered list of one-line questions, which I answered in minutes.

That’s the human in the loop I believe in: not a person reviewing everything, but a person making the decisions that actually need judgment.

The system finds the data problems too

The biggest surprise wasn’t in the AI’s text. It was in the client’s own product data, where names and descriptions contradicted each other. AI didn’t create those errors. It just finally read everything, line by line, the way no human team has time to.

Now the text doesn’t pick a side. The claim stays out, and the contradiction goes on a list for the client. Better data means better pages and better answers from AI search.

Why it scales

A system like this doesn’t depend on how many pages you give it. The rules are shared by every agent and script, and a person only sees the questions that need judgment. Adding pages doesn’t add the same amount of manual checking.

And the texts can be short. If the data doesn’t support a sentence, it doesn’t get written, even if the page ends up at 200 words instead of 600. Length was never the goal. Usefulness was.

If your AI content could appear on a competitor’s site unchanged, it’s not your content.

0