Four thousand descriptions that do not read like a machine wrote them

Generating four thousand descriptions with AI is the easy part: an afternoon's work. The problem arrives afterwards, when you read them and realise they are all the same description with the words changed.

The fault is not the machine. It is that you asked for them without giving it anything of yours, and with nothing in hand it writes the only thing it can: the average of everything it has read.

Why you can spot them

Text generated on empty always has the same tics, and they are easy to list.

They open with a vague promise. "Discover the timeless elegance of". No human being starts a sentence that way about a product they know.

They describe feelings instead of things. An adjective does not tell you whether the table fits your kitchen. Buyers want the measurement, not the atmosphere.

They are all the same length. Three paragraphs for an 8-euro product and three for an 800-euro one, because length was a setting rather than a decision.

They never name a constraint. No product has drawbacks, in any description, ever. More than anything else, that is what makes you suspect nobody wrote them.

What to give before asking for text

Three things, and all three are already in the company.

The product's data, not the product's name. Measurements, materials, compatibility, what is in the box. With those, text becomes specific by construction, because it contains facts no other product has.

How your brand talks, written down. Not "friendly but professional", which describes everyone. The terms you use, the ones you avoid, what you call things, whether you address people informally. Three pages of glossary beat any prompt.

What customers ask. The questions arriving at support are the best content you own, and it is already written: they are the things people actually want to know, in the words they use.

Review is not reading four thousand texts

This is where projects die. Nobody re-reads four thousand texts, so you either publish blind or you do not publish. Both are wrong, and the way out is not willpower.

You sample, but intelligently: not twenty random products, twenty products chosen where the text is most likely to be wrong. Categories with thin data, products with complicated variants, items where an error costs a return.

And you check the verifiable things across the whole catalogue automatically: that measurements in the text match the ones in the fields, that no promises appear about shipping or warranties you do not offer, that no two products share a paragraph.

Three rules worth more than a long prompt

Ban empty adjectives in writing. A list of forbidden words makes more difference than ten lines of instruction about tone.

Require at least one verifiable fact per paragraph. If a paragraph contains no number, measurement or fact, it probably serves nobody.

Always ask who it is not right for. A line about limits makes the text credible, reduces returns, and not incidentally is the part an answer engine quotes most willingly.

Frequently asked questions

Does Google penalise AI-written text?

Not for how it was written, but for how useful it is. Generic, repetitive text does badly because it is generic, not because it is generated: the same text written by hand by a bored intern would do just as badly. What gets penalised is content that adds nothing.

How long should a description be?

As long as it takes to answer the questions about that product, and no longer. A simple accessory can be done in three lines. A technical product bought after consideration needs far more. A fixed length across the catalogue is the first sign it was not a decision.

Rewrite everything, or only the products that sell?

The products that sell already have a description that works, so the gain there is small. The value is in the tail: the items nobody ever looked at are also the ones no competitor has worked on.

How do you stop all the texts resembling each other?

By giving different data, not by asking for variety. If the only difference between two products in your data is the name, any tool will produce two similar texts, and it will be right to. Variety comes from attributes, not instructions.

Do you still need a person reviewing?

You need a person who decides the rules and checks the samples, not one who reads everything. The human work moves: less writing, more judgement about what is acceptable. It is a different job, and a more useful one.