AI Product Enrichment and PIMRestructuring for a DIY Retailer

Context

Since February 2026 I have been working as PIM advisor for a European DIY retailer, on the Akeneo catalog behind its e-commerce brand. That catalog feeds six channels, from the website to print and AI-generated content, and it shares its data model at brand level with the group's other master data systems.

The first request sounded simple: fill the empty attributes with AI. But before asking a model to write anything, we had to answer a more basic question, one that most AI projects skip. Does the catalog describe products the way a customer actually looks for them?

The challenge

The catalog had been modeled on ETIM, the classification suppliers use to exchange technical data. ETIM is a good language between companies, but it is a poor map for a shopper, a bit like using a warehouse inventory list as a shop window. The result was 1,324 flat families, each carrying between 196 and 385 attributes, and 5,181 attributes in total.

With that many attributes per family, completeness meant nothing: a product could look almost complete while missing the few facts a customer needs to choose, and 150 attributes stayed empty across the catalog. Adding AI on top of that model would only have produced more text in the wrong places.

AI amplifies the data quality you already have, good or bad. So the model came first.

What we did

1. Audit what is really there

A complete audit of the existing families, the 5,181 attributes, the attribute groups and the 25,818 options, presented to management as an interactive dashboard. It surfaced 320 duplicate options and 37 attributes carrying more than 100 options each.

2. Design for the shopper, not the supplier

We restructured the flat families into a two-level model of parent families and sub-families, and moved from ETIM codes to durable label-based codes. Out of 5,181 attributes, 555 were kept for the web, after reviewing 1,633 family-by-attribute recommendations with the business teams. A classification standard is not a data model: ETIM stays useful with suppliers, the customer simply needs a different structure.

3. Put AI inside guardrails

Only then did AI come in. I designed 200 AI configurations and 950 Akeneo rules, with prompts forced to pick from existing attribute options and never invent a value, and with specific handling for units and dimensions. We started with a 260-product pilot before scaling to 2,600+ product texts across 13 categories.

4. A pragmatic stopgap for supplier files

While the supplier portal was still being deployed, simple Python scripts checked incoming supplier Excel files for the GTIN key and mandatory fields. Not elegant, but effective, and far better than waiting for the perfect tool.

Result

• 80% coverage on 150 previously empty attributes, with an error rate below 5%.
• 2,600+ product texts generated across 13 categories, starting from a 260-product pilot.
• A 10-phase staging migration plan toward the new model, built around the 62 attributes the e-commerce platform needs.
• Training workshops and documentation so the teams own the model after the project.

The lesson I keep from this one: AI works when it is forced to respect the model. Give it a messy catalog and it will write you a messy catalog, only faster. More on the modeling side in PIM vs spreadsheets, and on the AI side in why products don't show up in AI search.