Product descriptions are the most obvious job to hand to a language model: repetitive, text-heavy and never quite finished. Used well, AI can clear a backlog of thin listings in days. Used carelessly, it produces confident copy about features a product does not have, environmental claims that became unlawful in the EU this autumn, and hundreds of near-identical pages that search engines have little reason to rank.
What AI does well
The strongest case for AI-generated copy is consistency at volume. A homewares retailer adding 400 SKUs from a new supplier typically receives a spreadsheet of attributes (dimensions, materials, colours, care instructions) and little usable prose. A model given those attributes, a style guide and two or three examples of the shop’s best existing descriptions can produce readable drafts with the same structure, tone and level of detail across the whole range.
Other genuine gains:
- Variants and channels. Shorter versions for marketplaces, longer ones for your own store, bullet summaries for comparison tables, all from one source.
- Translation and localisation. Drafts in French or German for a shop expanding cross-border, which a fluent reviewer then corrects for idiom and local terminology.
- Gap-filling of the right kind. Turning “100% cotton, 180gsm” into an explanation of what that weight means for the wearer is useful, provided the explanation is accurate in general.
- Accessibility. Drafting image alt text at scale, which many small shops simply never get round to.
The common thread is that the model is transforming information you have verified, not inventing information you do not have.
Failure modes
Invented specifications
Language models generate plausible text, not checked facts. Ask for a description of a stainless-steel water bottle without stating its capacity, and the model may supply one. Ask for a description of a lamp and it may describe a dimmer the product does not have. Google’s own guidance on generative AI content makes the same point from a search perspective: these tools predict likely word sequences, so output must be fact-checked before publication.
For a seller, invented specifications are not a stylistic problem. A description is a commercial communication; if it misstates the main characteristics of a product it can be a misleading practice under consumer law, and it is a straightforward route to “not as described” returns and disputes. The seller is responsible, whoever or whatever wrote the words.
Unsupported environmental claims
Models trained on years of marketing copy reproduce its habits, including vague sustainability language. That is now a specific legal risk in the EU. Directive (EU) 2024/825, which amends the Unfair Commercial Practices Directive, adds to the list of practices banned in all circumstances:
- making a generic environmental claim (“environmentally friendly”, “eco-friendly”, “green”, “climate friendly”, “biodegradable” and similar) where the trader cannot demonstrate recognised excellent environmental performance relevant to the claim;
- making an environmental claim about the entire product when it concerns only one aspect of it;
- claiming a product has a neutral, reduced or positive greenhouse-gas impact on the basis of emissions offsetting;
- displaying a sustainability label that is not based on a certification scheme or established by public authorities.
Member States had to adopt the national measures by 27 March 2026 and apply them from 27 September 2026. As of November 2026, then, these rules should be live across the EU, though the exact national provisions and enforcement bodies vary, so check how your country has implemented them. UK sellers face a parallel framework: the Competition and Markets Authority’s Green Claims Code sets out how existing consumer law applies to environmental claims.
The practical rule: an AI tool should never be the origin of an environmental claim. If a claim is true and substantiated, it belongs in your locked fact sheet with its evidence; if it is not there, the model must not write it.
Health and safety claims
The same applies to health. Food and food supplements sold in the EU may only carry authorised nutrition and health claims under Regulation (EC) No 1924/2006; a model writing “supports immune health” for a tea blend has created a compliance problem. Cosmetics, toys, medical devices and electrical goods each have their own rules on claims and warnings. Safety information such as age restrictions or choking-hazard warnings should be copied verbatim from the compliance documentation, never paraphrased by a model.
Sameness
Feed a model the same prompt for 200 similar products and you get 200 descriptions with the same rhythm and the same handful of adjectives. Shoppers notice, and so, in effect, do search systems that look for content that adds something.
SEO: what Google actually says
Google’s position, set out in its 2023 Search Central blog post and expanded in its current guidance, is that it rewards helpful, original content regardless of how it is produced. AI content is not penalised as such.
The line it does draw is in its spam policies. “Scaled content abuse” is generating many pages primarily to manipulate rankings rather than help users, and the listed examples include using generative AI tools to produce many pages that add no value for users. The policy is about intent and value, not tooling, but bulk AI generation is precisely the situation in which a catalogue drifts towards it.
Two further points from Google’s guidance are easy to miss:
- Metadata counts. Titles, meta descriptions, structured data and alt text generated by AI need the same accuracy check, because they can appear directly in search results.
- Shopping feeds have labelling rules. Google Merchant Center requires AI-generated product titles and descriptions to be submitted using the
structured_title and structured_description attributes, with the digital source type set to indicate trained algorithmic media. AI-generated images must carry the corresponding IPTC metadata, and sellers should not strip it.
Our assessment: for a small shop, the SEO risk of AI copy is less about penalties and more about wasted opportunity. The descriptions that earn visibility are those with information competitors lack: real measurements, fit notes from your own returns data, honest comparisons with neighbouring products. A model cannot supply that; it can only help you present it.
A workflow with human review
This is the process we would suggest for a small team. It is designed so that the reviewer checks against a source rather than re-researching each product.
- Build a locked fact sheet per product. Every attribute that may appear in copy, taken from the supplier specification or your own measurement: dimensions, materials, capacity, compatibility, certifications (with certificate references), safety warnings. If a fact is not on the sheet, it does not go in the description.
- Write a brief once. Audience, tone, length, structure (for example: one-line summary, two short paragraphs, care bullets), banned words (including generic environmental terms), and an instruction to use only the supplied facts and to flag anything missing rather than guess.
- Generate in small batches. Twenty products at a time is reviewable; four hundred is not.
- Review against the sheet. The reviewer checks every number, material and claim against the fact sheet, deletes anything not traceable to it and reads the result as a customer would. A short checklist (specs, claims, safety text, tone) keeps reviews consistent.
- Add what only you know. Fit, feel, how it compares with your other products, the questions customers actually ask. This is where most of the value is added.
- Record provenance. Note which descriptions were AI-assisted and who approved them. That makes Merchant Center labelling straightforward and helps if a claim is ever challenged.
- Sample after publishing. Re-read a random handful each month and watch returns reasons and customer questions for signs that copy is misleading.
Keeping a brand voice
A brand voice is a set of choices: sentence length, how technical to be, whether to use humour, what you never say. Models can follow such choices reasonably well if they are written down. A useful style guide for this purpose is short and concrete: five words you use, five you never use, two exemplary descriptions, and a note on how you address the customer. Update it when reviewers keep making the same edit.
What models cannot supply is the judgement behind the voice. A small seller’s descriptions often build trust precisely because they sound like a person who has handled the product. If AI drafts flatten that, the efficiency gain may cost more in conversion than it saves in time. Our guide to building customer trust as a small online retailer covers why that matters, and our analysis of how small retailers compete with larger platforms argues that specificity is one of the few advantages a small shop has.
Bottom line
AI-generated product descriptions are worth using where they transform verified data into consistent copy, and risky wherever they are allowed to fill gaps. The legal exposure has grown in 2026, with the EU’s ban on generic environmental claims now applying, and the search-quality bar has not moved: pages need to help shoppers. Lock the facts, brief the model properly, keep a named reviewer, and spend the time saved on the details only you can add. For the wider picture of AI in a small shop, see our practical guide to AI for small ecommerce businesses; for the broader regulatory context, our analysis of EU digital-platform rules for online sellers.
This article is general information, not legal advice. Check how Directive (EU) 2024/825 has been implemented in the countries you sell to.