Your Shopping App Is Now Learning From Shoppers in Other Countries
A new AI trick boosted orders 2.64% on a major e-commerce site.
Big shopping sites usually learn about you using a separate system for each country. Your ID and the product IDs you see in one country don't connect to anything in another. That's a problem, because smaller markets simply don't have enough shopping data to train a smart AI. Think of it like a new store opening in a small town with only a few weeks of sales history — it has no idea what to recommend.
This paper, from a team of researchers in e-commerce, borrows an idea from language learning. Bilingual speakers often mix two languages in a single sentence, and AI language models trained on that mixed text get better at both languages. The team did the same with shopping: they created a shared "dictionary" of products based on what things look like and what gets bought together, then built pretend shopping histories that blend items from different countries. Real rules keep it believable — price, popularity and who usually buys what all have to make sense.
They also added a smart filter. Some mixed-up examples are realistic, others are nonsense. Their system gives more weight to the believable ones and mostly ignores the rest. This matters because bad training data makes bad recommendations.
The payoff was measured, not just promised. On two real multi-country datasets and a live A/B test (where real shoppers were split into two groups), the method clearly improved recommendations in countries with little data while not hurting countries with lots of data. The commercial results: a 1.77% lift in advertising revenue and a 2.64% lift in orders on a large e-commerce platform. That's a small percentage but a huge amount of money at scale.
- Shopping sites run separate AI systems per country, so small markets get worse recommendations — this fixes that by sharing knowledge across borders.
- The trick is mixing items from different countries into fake-but-realistic shopping histories, like bilingual speakers mixing languages in one sentence.
- In a live test on a big shopping platform, it increased ad revenue 1.77% and orders 2.64%, helping data-poor countries without hurting data-rich ones.
Why It Matters
Better product suggestions in smaller countries mean less scrolling, fewer bad buys, and more sales for online shops.