In a retail chain of 20 to 500 people, the online catalogue stops where the person in charge of product listings has managed to get to. Suppliers send technical data sheets as PDFs, spreadsheets with entirely different columns, and photos of labels. Someone retypes it all. New references wait their turn, and older listings are never revisited when the supplier changes its product.
Generative AI seems made for this work: it reads a document and writes a listing in a few seconds. Yet it raises a specific problem. A language model produces plausible text without checking what it asserts: a composition, a dimension or a capacity can be well phrased and wrong.
What the retailer commits to by publishing a listing
The seller must provide the consumer with the essential characteristics of the goods before purchase (Code de la consommation, article L111-1). An incorrect characteristic on the listing binds the retailer, however the listing was written. It also costs money in returns and negative reviews, on products where the customer chooses based on the listing: the dimensions of a piece of furniture, the compatibility of a part, the composition of a textile.
To publish a listing written by AI, the retailer must therefore be able to say, for each attribute, which supplier document it comes from.
Separating extraction from writing
A reliable system splits the work into two distinct steps.
Extraction. The system reads the supplier documents in their original format and pulls out the attributes required by the retailer’s reference framework: material, dimensions, weight, origin, care. Each extracted value keeps its address: the document, the page, the table row. When two documents give two different values, both are kept and the discrepancy is flagged.
Writing. The listing is written solely from the extracted attributes, in the retailer’s tone. The text cannot assert anything that does not appear in the list.
This separation makes the result verifiable. A listing written in one go from a PDF would require a full review, document in hand, which is the original task all over again.
The check: every attribute has its source document
A second system, separate from the one that extracted and wrote, checks the listing before publication:
- each published attribute is present in a supplier document, at the stated address;
- units are consistent from one document to another;
- each characteristic mentioned in the listing text appears in the list of attributes;
- the mandatory statements specific to the category (composition of a textile, energy class of an appliance) are checked separately, with their own rules.
A listing that fails a check does not go live. It comes back with the attribute in question and the reason. The person who managed the listings rules on the cases the system presents: a discrepancy between two documents, a missing statement.
A record remains for each listing: the list of its attributes, with the source document and the date of the check for each. If a customer disputes a characteristic, the retailer can find where it comes from in a minute.
The digital product passport makes this record mandatory
The European Ecodesign for Sustainable Products Regulation (ESPR) provides for a digital passport for the products it covers: data on composition, repairability and end of life, attached to each product and accessible throughout its life. The working plan adopted by the Commission on 16 April 2025 names the first product families concerned: textiles, furniture, mattresses, tyres, steel and aluminium (European Commission).
For a retailer, this means keeping product data whose provenance is known, reference by reference. A retailer that already links each attribute to its supplier document has done most of this work.
Where to start
One product family is enough to start, preferably one where the listing decides the purchase and where returns are costly. You need:
- the raw supplier documents for that family, in the state in which they arrive;
- the retailer’s attribute framework, if there is one, or the best-maintained listings, which serve as a model;
- a person who knows the products, to rule on discrepancies.
The system works on these files, without connecting the PIM or the website to begin with. The result is measured by the number of references published with a complete listing, and the share of their attributes that point back to a source document.
Other uses of AI in retail, and a detailed scenario, are described on the page AI applied to retail.