Photo: free stock photography (Unsplash licence) — see imprint
Google never said a machine may not write it
The question we get asked, roughly weekly since the spring, is whether Google punishes AI-written text. The answer that has emerged from Google's own guidance is narrower and more useful than the panic suggests: the criterion is whether content is helpful and made for people, not which tool produced it. Automation aimed primarily at manipulating rankings is spam — that has been the position for years and it applies to a text mill run by humans exactly as much as to one run by a model.
This is worth internalising, because it moves the conversation off a question nobody can enforce and onto one that decides your ranking anyway. Google cannot reliably detect authorship, and it does not need to. It can measure whether a page satisfies the person who landed on it. A model-written text that answers the question well is fine. A human-written text that answers nothing has always been in trouble. Authorship was never the variable.
So why do 30,000 generated product texts still fail?
Because of what they contain, not who typed them. Look at what actually comes out of a bulk run against a thin attribute set. 'The XY-400 impresses with high quality and reliable workmanship. Thanks to its modern design it is suitable for a wide range of applications.' That sentence is true of every product ever made. It answers no question, distinguishes nothing, and would be equally at home on a shovel or a laptop. Thirty thousand of those is not thirty thousand pages of content. It is one empty page, printed thirty thousand times with the model number swapped.
This failure mode predates the current tools by a decade. Agencies were selling spun near-duplicate product copy long before any of this, and it did not rank then either. What changed is the cost. Producing 30,000 hollow pages used to require a budget and an offshore team, which acted as a natural brake. Now it takes an afternoon, so shops that would never have made this mistake are making it at scale.
The test that matters: could a competitor publish the same sentence?
Take any paragraph off your product page and ask whether your closest competitor could paste it onto their page for a different product without changing a word. If yes, it carries no information, and no amount of keyword placement rescues it. If no — because it names a specific tolerance, a specific incompatibility, the reason the cheaper variant exists — then it is doing work, and it does not matter whether a person or a model assembled the sentence.
This is also why the input matters more than the prompt. A model given a product name and told to write 200 words will fabricate adjectives, because that is all it has. The same model given the actual dimensions, the compatible machines, the two things the supplier's datasheet says and one line from your support team about who returns it and why can produce a text that passes the competitor test. The bottleneck was never the writing. It was that nobody in your company had written down what makes the product different.
- Could a competitor publish this sentence unchanged? Then delete it.
- Does the page answer the question the searcher actually typed?
- Is there one fact here nobody else has? A tolerance, a fit, a caveat.
- Would you send this text to a customer who phoned to ask?
The real risk is not a penalty. It is being ignored.
Shop owners picture a manual action arriving by email. That is not usually what happens. What happens is much quieter: the pages get crawled, get judged unremarkable, and settle onto page four where nobody sees them. There is no notification, no message in Search Console, nothing to appeal. You simply spent a budget on 30,000 URLs that produce no traffic and dilute the ones that would have.
That dilution is the part people miss. A catalogue of 500 genuinely useful pages competes better than the same 500 buried inside 30,000 hollow ones, because crawl attention and internal link equity get spread across all of it. Generating text for products nobody searches for is not neutral. It is a cost with an ongoing tax.
What we would actually do with the budget
Sort the catalogue by search demand and revenue, take the top slice — often a few hundred products out of thousands — and give those real texts. Use a model to draft them if you like; it will be faster and nobody will be able to tell. Leave the long tail with clean structured attributes and no prose at all. A product page with accurate specs and no marketing paragraph is not thin content; it is an honest page. A product page with 200 words of generated adjectives is thin content wearing a suit.
And if the honest answer is that your 30,000 products are genuinely interchangeable commodities that nobody searches for by name, then no text strategy fixes that. Your traffic will come from category pages, filters and comparison content, and you should spend the budget there. It costs us money to say so, but writing 30,000 texts you do not need is a worse deal than admitting the catalogue does not need them.
| Page type | Generated text? | Reason |
|---|---|---|
| Top 200 products by demand | Draft with AI, edit hard | Real search demand rewards a real answer |
| Long-tail commodity SKUs | No prose, clean attributes | Nobody searches them; text adds dilution |
| Category and filter pages | Yes, with human structure | This is where commodity traffic actually lands |
| Guides and comparisons | Draft only; needs expertise | The value is judgement the model does not have |
- Google's criterion is helpfulness, not authorship — the tool was never the issue.
- If a competitor could publish your sentence unchanged, it carries no information.
- The punishment is not a penalty email — it is 30,000 URLs nobody ever sees.
- Better input beats a better prompt; the bottleneck is unwritten product knowledge.
Frequently asked questions
Not for being AI-generated. Google's stated position is about whether content is helpful and made for people; automation whose main purpose is manipulating rankings has counted as spam for years, regardless of who or what produced it. In practice you are not judged on authorship — you are judged on whether the page answers the question someone typed.
Reliably, no — and it is the wrong thing to worry about. Detection tools available today produce false positives on ordinary human writing often enough to be useless as evidence. What Google can measure is whether a page is substantially the same as thousands of others and whether visitors get what they came for. Optimise for that, not for evading a detector.
Almost certainly not. Sort by search demand and revenue, write real texts for the top slice, and leave the long tail with accurate structured attributes and no prose. A page with correct specs and no marketing paragraph is honest. A page with 200 generated adjectives is thin content that also dilutes crawl attention across your catalogue.
Feed it facts, not a product name. Real dimensions, what it fits and does not fit, why the cheaper variant exists, and one line from support about who returns it and why. Then apply the competitor test: if a rival could paste your paragraph onto a different product unchanged, it is worthless. The editing pass is not optional and it is where the ranking comes from.
We do this for a living — Shopware, Node.js, React, ERP integration and automation for B2B.
Talk to an engineer