One e-commerce listing describes a food container with a frog on it, but the photograph shows a children’s dinner set. Another advertises an Arabic edition of Monopoly and the box in the picture is the standard game. A third offers a Sony headphone model in a colour that model has never been made in, which makes it either mislabelled or counterfeit.
These phenomena are not unusual. A marketplace hosting several million listings from thousands of independent sellers has no realistic way of checking them by hand and mostly does not try. The errors stay in the catalogue and that decides which products a shopper finds, which ones never surface and which ones are unsold because nobody can search for them.
Those three erroneous listings were caught by a software that compared the pictures of a listing against its text and flagged the mismatch. It was built by Youzu, a Berlin-based company that sells catalogue automation to retailers and one of a growing number of firms working on the same problem. Its founder and CEO, Nail Vali, mentioned that they are trying to solve an issue sitting underneath every online store in the region that determines what shoppers are shown and not.
The Invisible E-commerce Layer
J&T logistics service in Malaysia, representing the distribution part of the e-commerce layer. (J&T)
This layer is known as product discovery, everything that happens between a shopper arriving at a store and finding the thing they want, including search, filters, category navigation and recommendations. It is the half of e-commerce that is placed before the checkout page and it runs on the product catalogue underneath, the database of titles, descriptions, attributes and images describing what is for sale.
The scale of e-commerce in Southeast Asia is substantial. The sector was projected to reach 185 billion US dollar in gross merchandise value (GMV) in the region in 2025, with 60% of its population now shopping online.
Nail’s argument is that the product discovery layer has barely moved. “E-commerce in general hasn’t changed for the last 20, 25 years,” he said. “The major issue is actually on the product data layer, at the catalogue itself, because it’s not organised so well.”
Most retailers manage catalogues through product information management systems, which he describes as very manual processes. It has become a serious issue on marketplaces, because the data comes from thousands of sellers, each filling in fields their own way, against templates the platform cannot really enforce.
“Almost no one in the e-commerce space is automating the part where you introduce your content and then get the attributes, tags and classify the data in an automated way,” Nail said. The result is familiar to anyone who has shopped online in the region. Searches return nothing for products that exist, while filters exclude items that qualify and listings become invisible.
Nail is careful not to claim the idea is new. “I wouldn’t say we’re the first,” he said. Concept papers on automating catalogue data existed five to ten years before generative artificial intelligence (AI) arrived and retailers have wanted it for as long as they have had catalogues.
The first factor that held it back was system cost. Until recently, it was cheaper to pay people to do the work than automating it and generative AI was badly flawed. It used to create errors that invents a feature a product does not have, or misses one it does. That is why Nail keeps returning to what he calls product truth, where a description corresponds to the actual item. Generative models do hallucinate, which could be a problem in a catalogue of ten million products.
The second constraint is processing cost. Reading text is cheap, but not reading images and videos, and text alone gives an incomplete picture of what a product is. “With just text you won’t have the full picture of your catalogue,” Nail said. He notes that even the largest platforms have not solved this. “Even like the big tech companies like Amazon, they are still doing these processes, these processes are quite manual.”
Teaching the Catalogue to Read Itself
An example of how Youzu’s platform enhance the catalogue description on e-commerce platforms. (Youzu)
Youzu’s software allows a retailer to send its catalogue in whatever format and produces a structured version based on it. The first stage is moderation, checking each listing against the platform’s own policy rules before anything else happens. For Snoonu, a Gulf marketplace and one of Youzu’s customers, that includes categories the platform does not permit and it also catches the contradictions described above, where the image and the text describe different products.
The second stage is categorisation and enrichment. An iPhone would be placed under electronics, then mobiles, then smartphones. A chair goes to house and garden, then furniture, then living room or outdoor. Titles are normalised, attributes extracted and descriptions generated from the images where a seller supplied none.
The justification for this process is that richer, more consistent product data improves a retailer’s own search and recommendations and makes its products easier to find outside the store, both through conventional search engines and through AI assistants that increasingly sit between a shopper and a purchase.
On top of that foundation, Youzu also provides visual tools, including image search, shop-the-look for fashion and furniture and a room visualiser. This function was actually the company’s first product, before exploring the product discovery layer. “Once we started digging, we saw even the big guys have got these problems,” he said. The stated ambition now exists well past discovery and toward automating procurement, invoicing, logistics and warehouse management through one interface.
Who Decides Whose Goods Appear
An illustration of livestreaming sales, an increasingly popular e-commerce marketing method in Southeast Asia. (Alibaba Group)
Asked whether the region’s catalogues might be too disordered for the system to handle, Nail said the opposite. “The messier it is, the better,” he said. The larger complication is that discovery increasingly does not start in a search box at all. Video commerce reached roughly 25% of Southeast Asian e-commerce GMV in 2025, up from under 5% in 2022, with the number of sellers using video rising 80% year on year to more than three million. Nail does not see this as displacing the catalogue. “Video is just one of the funnels for discovery,” he said. “But the user will still go to the e-commerce page to buy this stuff.”
The shift arguably supports him. A recommendation feed draws on the same product data a search query does, but without a query to correct it, which makes the underlying record matter more rather than less. The case is strongest for small sellers. One filling in three fields where a brand fills in thirty is competing on data entry and losing. The alternative to automating that work is not manual review, which no marketplace does at this scale, but the catalogue as it stands.
Whichever surface the shopper arrives through, the classification underneath is not equally reliable across the region. The company will localise into any language the models support, but support and performance are not the same. “English is still better,” Nail said, “because there are more online resources, more vocabulary that’s available open source.” Consequently, less widely spoken languages further down again. The asymmetry is well documented in language technology research, where Southeast Asian languages are consistently described as underrepresented in model training data.
Pricing works in the opposite direction. Youzu prices enterprise packages by region as well as by catalogue size, on the basis that a customer in the US has a larger budget than one in Indonesia. Smaller markets consequently pay less and a marketplace in Cambodia or Laos can reach a system priced beyond it at US rates. Regional pricing is what puts this software within reach of anyone other than the largest platforms, though what a smaller buyer gets is the version that reads its language least well.
Which returns to the initial three rejected listings. Automated moderation does something no human team at that scale can and a catalogue left uncleaned is not fairer, only less usable. It also sets the boundary of what a shopper is permitted to see, under rules the platform writes and software applies, across millions of listings no person will review.
Automating catalogues is now cheaper than paying people to do it and platforms are buying. Southeast Asia has more to gain from that than most regions. What nobody has settled is who sets the rules the software follows, what a seller in Surabaya or Cebu can do when the software hides their listing and whether any regulator in the region treats this as a question of who gets to sell rather than one of what software to buy.
About The Author
