AI Product Descriptions That Don't Sound Like AI

You can spot AI product copy in two seconds. It praises, it promises, and it says nothing you couldn't also say about the competing product.

The model is not the problem. The problem is that the AI knows nothing about your product, so it writes whatever the internet says about every product in that category. Type "write a product description for a toothbrush" and you get back the exact average of the internet.

At nano we have over 100 product texts and variants to write. By hand that is three to four hours per text if it is going to be good. With AI it is around 30 minutes, but only with the process described here. Four steps: gather raw material, define the brand voice, build a text template, then check and cut. Plus the list of things we never hand to the AI.

Why AI copy stands out

Three patterns give away an unchecked AI text immediately.

The first is the empty adjective. "High-quality", "thoughtfully designed", "perfectly balanced". Words that claim a judgement without giving a reason. A customer reads straight past them, because they appear in every description.

The second is the rule of three with nothing inside it. AI loves rhythm in threes: "light, durable and lovingly made". Sounds smooth, says nothing.

The third is the missing objection. A human eventually writes "for very sensitive gums this model is too firm". AI never writes that on its own, because its job is to sell. That one sentence is what separates advice from advertising.

All three patterns disappear once you give the AI enough real material. It invents filler only when it has nothing concrete to work with.

Step 1: Gather raw material

This is the step almost everyone skips, and it decides the outcome. Before a single prompt gets written, you collect the facts only you have.

At nano the material package for each product comes from six sources.

• Technical data. Dimensions, material, weight, bristle thickness, compatibility. Everything measurable, in numbers.
• Customer reviews. Especially the three and four star ones. That is where customers say what actually concerns them, without the enthusiasm and without the anger.
• Support requests about the product. Every recurring question is a gap in the text. We pull them from the categories our pre-sorting produces anyway.
• Return reasons. Why does this product come back? If "firmer than expected" is at the top, that belongs in the text, not around it.
• The manufacturer's notes. Why is the product built this way? That reasoning is missing from nearly every shop description.
• Two competitor texts. Treat them as a list of what to avoid. Whatever they say has to be said differently on our page.

This package is where the value sits. After that the AI is just the tool that turns it into sentences.

One side effect we did not expect: gathering the material surfaces product problems. It is how we noticed that an entire variant was described incorrectly. For getting structured return reasons, see our article on returns management.

Step 2: Define the brand voice

A brand voice is more than a list of adjectives. "Friendly, modern, approachable" helps no model, because every brand claims exactly that.

What works are rules and examples. Our voice document for nano has three parts and fits on one page.

Part one: hard rules. Informal address. No exclamation marks. No superlatives. Sentences under 20 words. No promises about medical effects. Every claim needs a reason in the same paragraph.

Part two: banned words. Ours holds a collection of terms we do not want to see in shop copy. Revolutionary, unique, game changer, premium experience. The list grows with every text that feels wrong.

Part three: three example texts. Two good ones, one of them short, plus one written deliberately badly. A model learns faster from one counterexample than from ten rules.

This document is attached to every prompt. It is the difference between copy that fits your brand and copy that fits everyone.

Step 3: The text template

We never let the AI write freely. It fills a template that is the same for every product.

The "who it doesn't fit" item is the most important one in the whole template. It is the only part that builds trust, and it lowers your return rate along the way.

The template has another practical advantage. Every text in your shop shares the same structure. Customers comparing two products find the same information in the same place.

And you can automate afterwards. A fixed template plus a fixed material package gives you a process a workflow can run. For what that looks like, see our article on marketing automation.

Step 4: Check and cut

The draft is never the text. We run every draft through four questions, and that takes about ten minutes.

• Is every fact correct? Check each number against the spec sheet. Models invent dimensions when they have none, and they do it convincingly.
• What can go? We cut a third on average. Usually the transition sentences that only summarise the paragraph above.
• Where is a claim without a reason? Either add the reason or delete the claim. There is no third option.
• Does a sentence sound like advertising? Read it out loud. Anything you would not say to a customer in a shop comes out.

The cutting is the step with the biggest effect. AI copy is almost always too long, because the model confuses completeness with quality.

An example from practice. At nano one draft contained the sentence "the extra soft bristles provide a gentle brushing sensation". After the check it read: "the bristles are 0.10 millimetres thin, roughly half the thickness of a standard brush." Same statement, verifiable, and the customer can decide for themselves whether that is enough.

What we never hand to the AI

There are places where we throw away every draft and write it ourselves.

• Health and efficacy claims. At nano that is legally delicate. Anything said about gums or effects we write ourselves and have reviewed.
• Every number in the final text. Prices, dimensions, delivery times, ingredients. We type them in by hand from the source file at the end.
• The headline. It decides clicks and search results. Five minutes from a human are worth more here than 50 variants from a model.
• Comparisons with competitors. Every statement about someone else's product is a legal risk. It has no place in a drafting process.
• Customer quotes. Quotes are never generated, not even "in the style of". Never.

This list looks different for every brand. But it should exist and be written down before the first text gets made. Otherwise someone decides it case by case under time pressure.

The limits

AI product copy is a tool with clear edges, and some of them hurt.

• Without raw material it will not be good. If you have no reviews, no support data and no manufacturer notes, you save no time. Then you are better off writing it yourself.
• The time saving is smaller than it sounds. The writing gets much faster, because all that is left is checking. Gathering and checking stay manual work. Anyone expecting full automation will be disappointed.
• At ten products the setup does not pay off. A voice document and a template cost you a day. That pays back from around 30 texts.
• Search engines no longer reward volume. 500 generated texts with no substance of your own are a risk, not an advantage.
• A model switch changes the tone. When you move to a new version, check the first five texts twice. We have seen an update treat our banned words more loosely.

And the most uncomfortable limit: a text can be clean, correct and still boring. The idea of why someone should buy this product comes from you. The AI only puts it into words.

Conclusion

The difference between a good and a generic AI product text happens before the prompt, not after it.

Gather material only you have. Write your voice down as rules and counterexamples. Give it a fixed template. And plan ten minutes of checking per text, in which you delete a third again.

If you notice along the way that your raw material is scattered across your inbox, your shop and a few spreadsheets: that is the real bottleneck, and the writing is not. Our overview of automating e-commerce processes shows how to pull that data together. And if you want to know what AI really carries in your shop and what it doesn't, read our assessment of AI in customer service or get in touch with Flowhouse.