AI Product Copy at Scale: Why Quality Is a Data Problem

“Let’s just do the product copy with AI” sounds like a prompting task. It isn’t. Anyone trying to generate hundreds or thousands of product descriptions at scale doesn’t hit the limits of the language model – they hit the limits of their own data. A model can only write about a product as well as the information it’s given – and that’s exactly where most projects fail, long before the first prompt gets written.

Key takeaways

• Scaled product copy almost never fails because of the language model – it fails because of incomplete or unstructured product data.
• Gold-standard examples per category secure consistency more reliably than endless prompt fine-tuning.
• Colour and size variants sharing an identical base text are the most common duplicate-content mistake in scaled copy.
• A four-step workflow – attributes, template, sampling, indexing decision – delivers speed and quality at the same time.
• Below roughly 1,000 products, manual writing with AI assistance usually beats full automation.

The real limit: structured attributes

A prompt like “write a compelling product description for this hiking boot” produces the same effect every time when the underlying data is thin: generic filler sentences, interchangeable adjectives, hallucinated dimensions or materials. Why that happens is straightforward: a language model fills gaps when it isn’t given facts – and the result reads plausibly at first glance, but is likely to be wrong. For more on the basic mechanism behind such fabrications, see our article on understanding AI hallucinations.

The solution isn’t a better prompt – it’s a clean attribute foundation: material, dimensions, country of manufacture, target group, use case, differences from the previous version. The more structured this data is in your PIM or product table, the less the AI has to invent. This investment pays off several times over: a clean attribute schema doesn’t just help with text generation, it also helps with filtering, product comparisons and visibility in marketplace listings, as with Amazon listing optimisation.

One example makes the difference clear: without an attribute foundation, a model working from the prompt above often produces a sentence like “this premium hiking boot impresses with first-class craftsmanship and optimal wearing comfort” – interchangeable, without a single verifiable fact. With a complete attribute foundation (nubuck leather upper, waterproof membrane, weight roughly 420 grams per shoe, 12-centimetre shaft height), you instead get a sentence like “this boot pairs a waterproof membrane with a nubuck leather upper and, at roughly 420 grams per shoe, weighs noticeably less than many comparable models with a similar shaft height” – concrete, verifiable, and exactly the kind of fact that’s also citable in search engines and AI answers.

Gold standards instead of endless prompting

The second lever is example copy that serves as a reference: two or three human-written or human-edited “gold-standard” descriptions per category that set the tone, structure and level of detail. A template plus these examples in the prompt keep things consistent across hundreds of variants – far more reliable than renegotiating the wording for every new product.

The reason this works so well: language models are strong at continuing a pattern, weak at freely inventing new style rules from a pure text description. A concrete example gives the model exactly the pattern it can reliably continue – an abstract tone instruction like “write casually but professionally” doesn’t.

Worked example: what the workflow actually saves

A rough worked example shows the order of magnitude. For 500 products written entirely by hand, the effort typically runs to about 30 to 40 minutes per text – roughly 300 hours for the whole range. At an internal hourly rate of €35, that’s a good €10,000 in pure writing time, before any correction rounds.

With a clean attribute foundation, a template and gold standards, the time per text drops to a pure review of roughly 5 to 8 minutes – about 50 hours for the same 500 products, around €1,750. The one-off investment in building the attribute database and templates comes on top of that, but typically pays for itself by the first large range run-through. These figures are rough orientation values, not a guarantee – the actual saving depends heavily on the starting quality of your product data.

Facing the duplicate-content risk honestly

Fitting colour variants or sizes with the same base text produces exactly what search engines warn against: thin, interchangeable content with no value of its own. The risk doesn’t just hit search engine visibility – interchangeable variant copy also weakens citability in AI answers, as our article on LLM optimisation for stores describes.

The countermeasure is technical, not rhetorical: keep variant text short, bundle the substantive content on the main product page, and set canonical URLs for genuine duplication. The order matters here – get the technical structure right first (canonical, short, bundled), and only then scale up the volume of copy. Get the order backwards and you’ll have to clean up thousands of already-generated texts after the fact.

The most common mistakes with AI product copy

Four patterns cause most quality problems in practice:

1. Prompting without attribute data: the model fills gaps with plausible-sounding fabrications. Fix: don’t start text generation before the attribute foundation is complete.
2. One prompt for all categories: a hiking boot needs a different level of detail than a T-shirt. Fix: define gold standards and templates per category.
3. No sample QA: errors go unnoticed until customers report them. Fix: check a fixed percentage per batch before going live.
4. Variant copy duplicated word for word: creates a broad duplicate-content risk. Fix: keep variant copy short, bundle the substance, set canonicals.

Governance: who owns the copy long term?

Scaled product copy needs a permanent owner, not just a one-off project team – without clear responsibility, quality control ends up orphaned by the second or third range refresh at the latest. The role doesn’t need to be big or expensive, but it has to exist and be built into regular day-to-day work, not run alongside as a special project.

In practice, a lean split works well: one person owns attribute data quality (usually sitting in category management or purchasing), a second owns template and gold-standard maintenance (usually in the content or marketing team), and a third runs sample QA before new batches go live. What matters less is exactly who holds which role, and more that all three roles are actually filled – in smaller teams they can be split between one or two people, but they shouldn’t disappear entirely.

From experience: a recurring mistake is letting responsibility informally lapse after the first successful rollout. New product categories, new suppliers with different data quality, or a model switch in the AI system you’re using regularly change the starting conditions – without a defined owner, this drift often only gets noticed once customers start complaining about incorrect product information.

The workflow: attributes, template, sampling, index

In practice, a four-step chain works well:

1. Complete the attribute foundation before any text gets written at all – from manufacturer data, PIM, or structured capture.
2. Define a template and gold-standard examples per category so tone and structure stay consistent across the whole range.
3. Sample QA per batch – not every text individually, but a fixed percentage gets reviewed by a human before the next batch starts.
4. Adopt a deliberate indexing strategy: not every automatically generated page needs to go into the search index if it has no standalone search value.

Follow these four steps in this order and you get speed and quality at the same time – rather than a thousand texts nobody wants to read.

The bottom line

The biggest misconception on this topic is thinking that a single prompting run marks a project’s completion. In reality, text creation is only the visible last step of a chain that starts with clean data maintenance and continues with ongoing quality control. Treat that chain as a continuous process rather than a one-off project, and you get product copy that still holds up after the tenth range refresh. The right first step is almost never the prompt – it’s an honest look at your own product database.