Written with AI assistance and reviewed by the NorwegianSpark SA editorial team.
Last updated: September 2026
Generating a few hundred product descriptions is now trivially easy, and that is exactly the problem. The technique that gets a catalogue written in an afternoon is very close to the technique Google's spam policies name explicitly, and the difference between them is not the tool — it is what goes into the prompt and what happens after the output.
This article sets out where that line actually sits, quoting the policy rather than paraphrasing it, and then describes a workflow that scales without walking across it.
What the Policy Actually Says
The relevant rule is scaled content abuse, and Google's Search Essentials states it plainly:
Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users. This abusive practice is typically focused on creating large amounts of unoriginal content that provides little to no value to users, no matter how it's created.
The listed examples include "Using generative AI tools or other similar tools to generate many pages without adding value for users" and "Stitching or combining content from different web pages without adding value". (developers.google.com, Search spam policies, checked 6 September 2026.)
Read the operative words carefully, because the whole distinction lives in them. The policy turns on primary purpose and on adding value, and it says explicitly "no matter how it's created". Volume alone is not the offence. Automation alone is not the offence. Producing many pages that add nothing, in order to rank, is the offence — and that description fits a hand-written catalogue of near-identical copy just as well as a generated one.
Google's separate guidance on AI content says the same thing from the other direction: "Appropriate use of AI or automation is not against our guidelines", and "Using AI doesn't give content any special gains. It's just content." We went through that guidance in more detail in AI content versus human content.
One thing we checked and did not find, because it is worth being precise rather than convenient: the spam policies page contains no rule specific to thin or duplicated product descriptions. Articles that cite a product-page duplication rule to that page are attributing something to it that is not there. The applicable standard is the general one quoted above, and the burden it places on you is a burden of value rather than of word count.
The Real Problem Is Duplication, and It Predates AI
The characteristic failure in ecommerce copy is not machine writing. It is that thousands of retailers publish the manufacturer's supplied description, unedited, for the same product. Every one of those pages is unoriginal, and the search engine has no reason to prefer any particular one of them.
Generation makes this worse in a specific and avoidable way. Asked to write a description from a product name alone, a model produces fluent copy assembled from what such products generally are — which is, functionally, a paraphrase of the same manufacturer text everyone else is publishing, plus a risk of invented specifics. You have replaced identical duplicate copy with near-identical copy that may also be wrong.
The thing that makes a product page worth existing is information that is yours: how it fits compared to the other things you sell, what it is actually like in the hand, what customers ask about it before buying, what they send it back for, what it works with, who it is wrong for. None of that is in the manufacturer's feed and none of it can be generated from a product name. It has to be supplied.
Prose and Structured Data Do Different Jobs
A frequent mistake is writing a description that tries to serve both machines and people, and serving neither well — the paragraph that lists dimensions, materials and SKU in a sentence.
The machine-readable job belongs to structured data. Schema.org defines a Product type — "Any offered product or service. For example: a pair of shoes; a concert ticket; the rental of a car; a haircut; or an episode of a TV show streamed online." — with properties for exactly the facts that do not belong in a paragraph: name, description, image, brand, sku, gtin, material, color, model, aggregateRating and offers.
One detail worth getting right, because it is commonly got wrong: price is not a property of Product. It belongs to Offer, reached through the offers property — schema.org describes that as "An offer to provide this item—for example, an offer to sell a product, rent the DVD of a movie, perform a service, or give away tickets to an event." Putting a price directly on a Product node is invalid markup, and the validator will tell you so if you run it. (schema.org, checked 6 September 2026.)
Once the facts are in the markup, the prose is free to do the job prose is for: helping a person decide. That is a much easier brief to write to, and a much easier one to generate against.
A Workflow That Scales Without Becoming the Thing the Policy Names
The principle throughout: the model transforms information you supplied; it does not invent information you did not.
- Fix the attribute data first. Materials, dimensions, compatibility, care, fit, country of origin, what is in the box. This is the unglamorous part and it is where the value comes from. If a third of your products are missing an attribute, a third of your descriptions will be padding.
- Add the knowledge that only you have. Two or three lines per product: the question customers always ask, the reason it gets returned, who it suits and who it does not. Your support inbox and your returns data are the source, and this step is what makes the page original in the sense the policy actually cares about.
- Generate from those inputs only, with an explicit instruction to use no fact not supplied. This is the single prompt decision that separates a useful description from a hallucinated one.
- Vary the structure, not just the words. Two hundred descriptions in an identical shape read as generated even when each is individually accurate. Different products deserve different emphases — that is a signal of real editorial attention, and it is cheap to instruct.
- Check every specific. Numbers, compatibility claims, materials, care instructions, anything safety-related. A wrong dimension is a return; a wrong care instruction is a ruined item and a complaint; a wrong safety or allergen claim is considerably more serious than an SEO problem.
- Have a person review before publication, and record who. This is good practice on its own, and it is also the arrangement that keeps accountability with a named human rather than with a pipeline.
- Publish in batches and watch what happens. Which pages get listed, which get impressions, which convert. A staged rollout gives you a signal; publishing the whole catalogue at once gives you nothing to compare against.
Does the EU AI Act Require You to Label These?
A question worth answering precisely, because the over-claiming in both directions is unhelpful.
The AI Act's Article 50(4) text-disclosure duty is narrow. Its second subparagraph applies to AI-generated text "published with the purpose of informing the public on matters of public interest", and it does not apply "where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content". Commercial product copy is not, on its face, text informing the public on a matter of public interest — and even where a disclosure duty did bite, the human-review exemption describes exactly the workflow in step six above.
Separately, Article 50(2) places a marking obligation on the providers of the generating systems rather than on you, and the amended Article 111(4) gives providers of systems already on the market until 2 December 2026 to comply with it. All quoted from the Official Journal text as amended and checked on 6 September 2026. None of this is legal advice, obligations differ by jurisdiction, and consumer-protection and advertising law — which absolutely does govern what you may claim about a product — is a separate body of rules and a question for your own advisers.
Eight Checks for a Description That Earns Its Page
- Does it contain at least one fact that is not in the manufacturer's feed?
- Would it help someone choose between this and the next item in your own catalogue?
- Is every specific in it verifiable against your own product data?
- Does it say who the product is wrong for? Almost nothing does, and it is the most trusted sentence on most product pages.
- Does it answer the question your support team hears most about this item?
- Are the machine-readable facts in structured data rather than buried in the prose?
- Would you be comfortable if a customer quoted it back to you in a complaint?
- Is it different in shape from the description above and below it?
The Counter-Argument: Fewer Pages, Better
The honest objection to this entire exercise is that for many catalogues the right move is not better descriptions at scale but fewer product pages.
If you sell the same shirt in twelve colours as twelve URLs, you have created eleven near-duplicate pages and a canonicalisation problem, and no amount of description quality fixes the underlying structure. One well-built page with variants usually outperforms twelve thin ones, and it is less work. Similarly, a long tail of products nobody searches for does not become valuable because it now has three paragraphs each; it becomes a larger surface of low-value pages, which is closer to the thing the policy describes, not further from it.
Spend the effort where demand exists. For the rest, accurate structured data and a short honest description are enough, and adding words to them is the opposite of the fix.
Tools
Shopify is the platform most of this work happens on and it carries its own generative features — the company describes Shopify Magic as "a suite of free AI-powered features that are integrated across Shopify's products and workflows" covering "store building, marketing, customer support, and back office management" (Shopify Help Centre, checked 6 September 2026). Zebao works the ecommerce AI side, and for catalogues sourced from suppliers rather than made in house, Doba is a supplier network — worth naming here because supplier-supplied copy is precisely the duplicate text this article is about replacing.
Disclosure: this article contains affiliate links. If you sign up through them we may earn a commission at no extra cost to you. It does not change what we recommend. Nothing here is legal advice.
Related Articles
