Contents7
Picture a product description written by an AI: clean, well phrased, and stating “machine washable”. Nobody checked. The supplier’s text said nothing about it, but the sentence sounded true, so it went live with all the others. The first customer who believes it sends the product back, and they are right to.
AI writes fast and fairly well. The risk is not that it writes badly: it is that it writes what nobody knows. This guide is for teams of 20 to 500 people who sell online and want to hand their product pages to AI without losing control of what those pages claim. It complements Google Shopping and marketplaces: getting your store ready, which details what Google expects from a product feed. This work is also where our ready-to-sell store starts, with a catalogue imported, cleaned, enriched and then reviewed before publication.
In short
AI files into fields what sits in free text, writes the same information the same way everywhere, drafts and translates; it must never guess a product attribute.
The order of work matters more than the model chosen: trim the catalogue, define the attributes, extract, review in batches, measure what is missing.
An invented attribute is false information about a product, which consumer law classes as a misleading practice.
Merchant Center requires descriptions or titles written with generative AI to be declared in attributes designed for that purpose.
What can AI do on a product page?
It helps when the information exists but sits in free text: the supplier’s description, an old product page, a spec sheet. It files it into fields, writes it the same way across the whole catalogue (“30 cm” rather than a mix of “30cm” and “thirty centimetres”), then drafts a description from those validated fields. It translates too: on Palmora Property, a property listings site in six languages that we run for its agency, the translation cost was cut by a factor of 60 by switching to an AI model.
The line is easy to state and hard to hold: AI works on what you know, never on what you don’t.
What AI does well
Extracting size, material or colour from a supplier’s text
Writing the same information the same way across the catalogue
Drafting a description from attributes already validated
Translating product pages into your buyers’ languages
What it must never do
Guess a size, a material or care instructions missing from the text
Add “eco-friendly”, “hypoallergenic” or “handmade” without proof
Invent a barcode or a certification
Write regulatory notices or terms of sale
For barcodes, Google says so itself on its GTIN page: when in doubt, leave it out, and do not try to guess one or make one up. We quote it in our guide to Google Shopping.
Why is an invented attribute so costly?
Because it is false, and false information about a product is not a matter of style. European consumer law counts among unfair commercial practices “misleading practices, either through action (giving false information) or omission (leaving out important information)”, as the European Union’s page on unfair commercial practices points out. The same page states that it is illegal for a business to advertise “false, inaccurate, or exaggerated claims about its environmental achievements or commitments”. An “eco-friendly” added by an AI to make the page look nice falls squarely there.
On the search engine side, Google is not targeting AI, it is targeting pages with no value. Its spam policies describe the abuse that lies in wait for a catalogue generated in one go:
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.
Among the examples, the same text lists “using generative AI tools or other similar tools to generate many pages without adding value for users”. An accurate, complete and reviewed product page has nothing to fear from this policy. A thousand pages that repeat the product name without teaching the visitor anything do.
In what order should you work?
The order matters more than the AI model chosen. We learned it on an online store we are putting back on its feet, and we now follow it for every catalogue: narrow the scope, define the attributes, extract. Never the other way round.
- Trim the catalogue.
First remove what no longer deserves a page: products discontinued for good, duplicates, items that do not sell and bring nobody in. Sales figures and Google Search Console settle this better than intuition.
- Define the attributes.
Write the list of fields every product page must carry (size, material, colour, care instructions…), with how each is written, before touching a single product.
- Extract without guessing.
Ask the AI to fill these fields from the existing text, and to leave a field empty when the text says nothing.
- Review in batches.
A person reviews each batch before publication, field by field, comparing against the original text.
- Measure what is missing.
For each field, count the share of products where it is filled: that figure tells you what to ask the supplier for or measure yourself.
The third step carries the whole method. An empty field is information: it tells you what is left to find. A guessed field is a mistake waiting for its customer.
Working in batches has another advantage: a badly written instruction shows up on the first batch, not across the whole catalogue. On Palmora Property, 1,726 property listings were rewritten with the help of AI, in 36 batches, with a single failure.
What does Google ask of a description?
Merchant Center’s page on the description attribute sets out rules that apply well beyond the Google feed, because they describe a good product page:
- put the most important details in the first 160 to 500 characters;
- describe only the product itself: size, material, intended age range, special features, technical specifications, shape, pattern, texture, design, variants;
- mention neither the price, nor the sale price, nor promotion dates;
- add no links to your store or to other sites;
- make no comparisons with other products, and do not use capital letters to highlight information.
This is exactly what a well-guided AI can follow, provided you give it these rules in its instructions rather than hoping it will guess them.
Watch out
A description written by AI must be declared in the feed
According to the same page, “All descriptions created using generative AI must be provided using the structured description” attribute (structured_description), whose sub-attribute digital_source_type takes the value trained_algorithmic_media. The page on the title attribute asks the same for a title created with generative AI, with the structured_title attribute. If your feed still goes out with only the description attribute, fix that before enriching the catalogue.
What coverage taught us
The most useful figure from an enrichment project is not the number of pages rewritten. It is the share of products for which each field is actually known, because that is what decides which filters and collection pages you can build.
From the field
Filters come from coverage, not intuition
On the store we are putting back on its feet, we reviewed 653 products and defined 26 fields before extracting anything. The AI enriched 481 product pages, all reviewed. But size, the first filter anyone thinks of, was only readable for 41% of products, and it lived at variant level. The plan called for four pages by size range, for six products in total: two of them would have been almost empty. Since then, a collection page is only created from five products upwards; below that, the script refuses.
The rest cannot be invented. It gets measured, or requested from the supplier.
Where to start?
Take twenty or so of your best sellers and do the exercise by hand: the list of fields, what the supplier’s text says, what is missing. You will quickly know whether the AI will have material to extract, or whether you first need to chase your suppliers for the information. The choice of platform, which decides where these fields live, is covered in Shopify or Medusa: which should a small business choose?, and what can be automated around the catalogue, from emails to reports, in our guide to marketing automation.
If you are rebuilding your store, this work opens our ecommerce offer: the catalogue is imported, cleaned, enriched in batches and reviewed before publication, and information missing from your data stays missing from the page. If you do not yet know which project to start with, our digital and AI maturity audit ends with a written GO or NO-GO, project by project.
Frequently asked questions
Can AI write all my product descriptions?
It can write the text of all your product pages, provided it starts from verified attributes. What it cannot do is find information missing from your data: a size, a material or care instructions that neither the supplier nor your team has provided. For those fields, ask it to answer “unknown”, then measure the product or ask the supplier.
Does Google penalise product descriptions written by AI?
Not for the way they are written. Its spam policies target content mass-produced to manipulate rankings, of no value to visitors, “no matter how it’s created”. An accurate, complete and reviewed product page is not affected. Hundreds of pages generated in one go, which teach the visitor nothing, are.
What should you do when information is missing?
Ask for it, do not guess it. The supplier’s text, a spec sheet or the packaging often give the answer; otherwise, measure the product yourself. In the meantime, publish the page without that field rather than with an invented value: an empty field can be filled later, a published error comes back as a return, a negative review or a complaint.
Should you disclose that a product page was written with AI?
In the Google Merchant Center feed, yes: a description or title created with generative AI goes through the attributes designed for that, structured_description and structured_title, with the value trained_algorithmic_media. On your store’s own page, the question falls under other rules, which we detail in our article on text written by AI.
Sources
- The European Union’s page on unfair commercial practices (new tab)europa.eu
- Merchant Center’s page on the description attribute (new tab)support.google.com
- The page on the title attribute (new tab)support.google.com
Jim · founder of Brain Plus AI
Also in e-commerce
Practical guide8 min
Google Shopping and marketplaces: preparing your store
Before listing a catalogue on Google Shopping or a marketplace, you need clean product data: identifiers, titles, prices, availability, shipping and returns. What Google Merchant Center asks for, and why it is the same foundation for every channel.
Choose and price6 min
Shopify or Medusa: which to choose for an SME?
Shopify charges a subscription and sometimes a commission; an open source platform like Medusa costs nothing in licence fees but needs a team to develop and run it. Here is how to compare the two over three years, and when each one wins.
Practical guide11 min The topic guide
Marketing automation for SMEs: with or without AI
Emails, social media, SEO in several languages, support, reporting, catalogue: most of an SME's marketing tasks can be automated with simple rules. AI is justified at a few precise steps, and a person reviews whatever commits the brand.