Catalogue and content automation
Product listings for 40,000 SKUs written, translated and kept compliant by AI
A seller with 40,000 SKUs on Amazon, eBay, Otto and its own store could not keep listings complete, consistent or translated. New products took weeks to go live. We built a content pipeline that turns supplier data into optimised, marketplace-compliant listings in six languages, with humans reviewing exceptions rather than writing from scratch.
- Client
- Home and garden marketplace seller
- Region
- Germany, selling across the EU
- Scale
- 60 to 90 employees
- Timeline
- 10 weeks
Representative engagement. Client details are anonymised and figures reflect typical outcomes from comparable deployments.
What the client got
0 days
From supplier feed to live in six languages
Down from four to six weeks.
0%
Fewer suppressed listings
Compliance checks run before upload rather than after a warning.
0%
Uplift in conversion on regenerated listings
Measured against the previous listing on the same SKUs.
The bottlenecks
Bottleneck 1
Supplier data was incomplete
Product feeds arrived with missing attributes, inconsistent units and descriptions written for trade buyers. Every listing needed manual enrichment.
Bottleneck 2
Each marketplace wanted something different
Title length, attribute names, image rules and prohibited claims varied by channel and country. Listings were suppressed for violations nobody noticed.
Bottleneck 3
Translation was a bottleneck
Six languages across five marketplaces meant an agency queue. New products launched in German first and elsewhere months later.
What we built
Step 1
Attribute enrichment
Supplier feeds are normalised and an LLM fills missing attributes from the product name, category, images and manufacturer documentation, with a confidence score and source for each value.
Step 2
Channel-specific generation
Titles, bullets, descriptions and backend keywords are generated per marketplace against each channel’s rules and the category’s best-performing patterns, then checked for prohibited claims and length limits before upload.
Step 3
Translation with glossary control
Listings are translated into five more languages using a maintained glossary of brand and category terms. Native-speaking reviewers see only flagged segments.
Step 4
Monitoring and repair
Suppressed or underperforming listings are detected daily. The workflow diagnoses the cause, regenerates the failing element and resubmits, logging the change.
The automation workflow
Supplier feed to live listings
Per SKU, per channel, per language
Questions clients ask about this engagement
Does AI invent product attributes?
Every generated attribute carries a source and a confidence score. Low-confidence values are reviewed by a person before publication, and safety-relevant attributes are never inferred.
How do you avoid duplicate content penalties?
Each channel gets content generated against its own rules and patterns rather than a copy, and the store receives a distinct long-form description.
More in E-Commerce & D2C
Industry overview
AI email marketing automation
An AI email marketing engine that ships campaigns in hours, not weeks
An agency running email for a dozen consumer brands was limited by how fast its team could brief, write, build and QA campaigns. We built an AI engine that turns a product feed and a campaign brief into segmented, on-brand emails ready for review, and automates the lifecycle flows every brand needs.
5x
More campaigns per brand per month
Customer support automation
Order status, returns and exchanges handled without a support agent
Three quarters of support tickets were about order status, returns or exchanges, and the queue doubled every sale. We automated those tickets end to end, connecting the help desk, Shopify and the carriers, and left agents with the conversations that need a person.
72%
Tickets resolved with no agent
Wholesale and retail lead generation
Taking a D2C beverage brand into 300 retail doors with outbound
A direct-to-consumer beverage brand had strong online reviews and rising ad costs. Growth needed wholesale: independent grocers, gyms, cafés and regional chains. We built an outbound engine to buyers and owners, with samples and sell-sheets handled by workflow, and the brand opened hundreds of retail doors without a field sales team.
312
Retail doors opened
Conversational commerce automation
A WhatsApp shopping agent that recovers carts and answers sizing at 2am
A fashion brand selling across India and the Gulf did most of its customer conversation on WhatsApp, with a team of eight answering sizing, delivery and exchange questions by hand. Abandoned carts were followed up with a generic blast. We built a WhatsApp agent that knows the catalogue, the customer and the order, recovers carts with a conversation, and hands off to a person when it should.
23%
Abandoned cart recovery rate
Have a similar bottleneck?
Tell us how the work happens today. We will show you the workflow we would build and prototype it on your own process.