Enriching an incomplete supplier catalog: attributes, descriptions and categories for an online retailer

Implementation example. This scenario describes a request we receive in this form. It is not a project delivered to a named client, and no results are claimed.

The context

An online retailer sells home equipment: small and large appliances, kitchenware, linens, storage and lighting. Its catalog has several thousand SKUs and changes every season. It works with dozens of suppliers and brands, and part of its sales go through marketplaces.

The e-commerce team is small: a catalog manager, one or two product listing specialists, and sometimes temp staff during peak season. The tools in place:

an e-commerce platform and, depending on size, a PIM;
feeds to marketplaces and comparison shopping sites;
supplier files in Excel, CSV or PDF, each with its own structure.

The day-to-day problem

Every supplier sends its product sheets in its own way. Dimensions are in centimeters for one, in millimeters for another. Wattage appears in the title, not in a field. Descriptions are copied from the brand's catalog, identical to those of every competitor, or missing altogether.

Products go live with empty filters: there is no way to sort vacuum cleaners by battery life or saucepans by induction compatibility. Some SKUs are miscategorized. Marketplaces reject listings for missing attributes.

Launching a new collection online takes weeks of data entry, and the slowest-selling product pages are never completed.

What the retailer wants

“I want every product to go live with its attributes filled in, a useful description and the right category, without my team spending the whole season retyping supplier files.”

The solution, step by step

1. Define the data model for each product family. For each product family, we set the list of attributes that matter to customers and to filters: dimensions, wattage, capacity, material, compatibility, care, warranty. Each attribute has a unit and a list of allowed values.

2. Read the supplier files. The tool reads files as they arrive (spreadsheets, CSV, PDF sheets) and maps each column or piece of information to the right attribute. It converts units and harmonizes values (“inox,” “acier inoxydable” and “stainless” become a single value).

3. Extract what is hidden in the text. When a piece of information exists only in a supplier title or description (“2,000 W,” “dishwasher safe”), the AI spots it and suggests filling in the matching attribute, stating where it came from.

4. Categorize each SKU. The AI suggests the site category and the category for each marketplace, with a confidence level. Uncertain cases are set aside for a human decision.

5. Write the descriptions. Using only validated attributes, the AI writes a standardized title, a description and key points in the site's tone of voice. It invents no features: a missing attribute does not appear in the text.

6. Flag what is missing. For each SKU, the tool lists missing or mandatory information: attributes required by marketplaces, product safety information and manufacturer contact details, energy label and product information sheet for the relevant appliances. These requests are grouped by supplier, ready to be sent to them.

7. Validate and publish. The catalog team validates in batches, makes corrections if needed, then publishes to the site and the feeds.

The rules to follow

For products sold online in the European Union, the General Product Safety Regulation (EU 2023/988) requires displaying, among other things, the product identification, the manufacturer's contact details and safety warnings. Appliances subject to energy labeling must show their label and product information sheet. The tool checks that this information is present; it does not create it.

What the AI does, what the team does

The AI reads, standardizes, extracts, categorizes and writes, always citing its source. The team defines the data model, settles uncertain cases, validates and publishes. Suppliers fill in what is missing.

The data and access required

an export of the current catalog and the category tree;
recent, representative supplier files;
the requirements of the main marketplaces;
a few product pages considered successful, to set the tone;
access to the PIM or platform for importing.

The project timeline

Weeks 1 and 2: assessment. Analysis of a sample of the catalog and supplier files, measurement of missing attributes. This is the format of our audit (8 to 10 days).
Weeks 3 and 4: model and reading. Data model for two or three priority families, reading of the main suppliers' files.
Weeks 5 and 6: production. Enrichment, categorization and writing for these families, batch validation, import.
Then: expansion. Other families and new suppliers, collection by collection.

The deliverables

a data model for each product family;
a tool for reading and standardizing supplier files;
enriched product pages for the priority families, ready to import;
a report of missing information by supplier;
a batch validation procedure;
a short training session for the catalog team.

The metrics to track

the attribute completion rate by family;
the time between receiving a supplier file and going live;
the number of listings rejected by marketplaces;
filter usage on the site and conversion on category pages;
the time the team spends on data entry.

These metrics are recorded at the start, then tracked with each collection. No results are promised in advance.

Limitations and prerequisites

Information missing from supplier files cannot be guessed: it is requested from the supplier.
The data model requires decisions from the team: it is the foundation for everything else.
For a few dozen SKUs, manual entry is still simpler.

Indicative budget

The assessment corresponds to our audit, from €1,000 excl. VAT. The next phases are quoted on request, depending on the number of families, suppliers and channels.

Frequently asked questions

Do we need a PIM? Not to get started. The tool can produce import files for the e-commerce platform. Beyond a certain volume, a PIM becomes useful; we discuss it during the audit.

Will the descriptions be different from our competitors'? Yes: they are written from the attributes, in the site's tone of voice, instead of being copied from the brand's catalog.

What happens if the AI gets an attribute wrong? Each suggested value states its source, and uncertain cases are set aside. Batch validation makes it possible to correct before publishing.

Can supplier PDF sheets be processed? Yes, the tool also reads technical data sheets in PDF.

What next?

If your new collections go live with empty filters and copied descriptions, the audit lets you measure in 8 to 10 days what is missing and where to start.

Client challenge