Case Study: thousands of product pages rebuilt by AI agents
A neglected e-commerce catalogue, brought back to life from a name and a barcode
Situation
This toy and games retailer, anonymised here, sells online through a WooCommerce site that had gone without maintenance for a long time. The catalogue held thousands of empty product pages: a name, a barcode, and nothing else. No photo, no description, no category. Pages a customer cannot assess and a search engine has no reason to show. Filling them in by hand would have taken months of data entry, and the site's developer had never managed to get the job moving.
Objective
Rebuild the content of every page automatically without giving up on reliability: a photo that does not match the product or the wrong age range costs more than an empty page. The rule set from day one: an agent only publishes what it is certain of, and everything else goes through a human.
Pages filled automatically, 100% reliability
80%
Pages filled, sent for human review
15%
Pages missing at least one piece of information
5%
Starting point for each page
Name + barcode
Process
The barcode as the anchor
For each page, AI agents start from the only two reliable pieces of data, the name and the barcode, to identify the exact product: images, specifications, description. They then write the page and file it in the right shop categories. Mission led by Meidy Baffou.
A confidence level for every page
Every rebuilt page gets a confidence level. At 100%, it is published directly. Below that, it is filled in but goes into a human review queue. When a piece of information cannot be found, the page is flagged as incomplete rather than filled in by guesswork.
SEO, page by page
The rebuild was also used to optimise every page for organic search: an empty catalogue becomes one that search engines can read and rank.
The team takes back control
On the consultant's advice, the team started using Codex, a coding agent, to improve the site on its own. A training session is planned to anchor that autonomy.
Results
80% of pages were filled automatically, with 100% reliability.
15% of pages are filled but awaiting human review, because their confidence level did not reach 100%.
For the remaining 5%, at least one piece of information could not be found automatically: these pages are identified and set aside, not made up.
Every page was optimised for SEO as it was rebuilt.
What makes a project like this possible is not an agent's ability to write a description: it is the sorting. Knowing which pages to publish with confidence, which to show a human and which to leave aside is what turns thousands of hours of data entry into a few hours of review. Do you have a catalogue, a document base or a CRM to put back in order? That is the kind of project we scope.
Testimonials
- Thank you very much for the start of this engagement. On your advice, we have already started using Codex and I have to admit the results are pretty impressive. Things we had been asking our developer for years, which he never managed to do, were done in 2 minutes. The efficiency is almost scary! (the client, a few days after the engagement)
Contact us
- Want to upskill your teams?
Let's schedule a meeting to assess your training needs
- Want to identify high-ROI use cases for generative AI?
Let's schedule a strategic call