The same product, mismatched across German and English: how hard cross-language product matching is
A bilingual product benchmark shows machines need entity matching to recognise that two language versions describe the same item.
Key numbers
- Source: product data from the German price-comparison platform billiger.de, covering thirteen consumer categories including hard cases such as clothing and furniture.
- Language setup: every offer has an aligned English translation while all pairs, splits and labels stay fixed; cross-language test sets put German and English records inside the same pair.
- Validation: six supervised matchers plus zero-shot GPT-5.2 across the German version, the English version and cross-language test sets — showing how difficult the benchmark is.
- Key difference: most matchers score higher on average on the English version than the German one (largest gaps for RoBERTa and HierGAT), while the zero-shot LLM is largely language-insensitive; the benchmark is harder than existing English ones such as WDC Products and Abt-Buy.
Method and sample
arXiv paper “Billiger.de Products: A Bilingual Entity Matching Benchmark” (2026-09-29): a German-English product entity matching benchmark built from a German price comparison platform across thirteen consumer categories, with variants differing in corner-case and unseen-entity fractions, validated with six supervised matchers and zero-shot GPT-5.2 on both languages and cross-language tests.
What it means for brands
Cross-border sellers often assemble product data in several languages: inconsistent spec formats, different units, different title structures make machine matching hard. That affects price comparison and indexing, and it affects whether AI describes you consistently across languages.
How SourceSight handles it
SourceSight builds projects per market and asks in each market's language so mentions and descriptions can be compared across languages, while the content factory standardises specs, units and certifications to reduce cross-language mismatches.
Limits
The benchmark comes from one price-comparison platform and thirteen categories and serves entity-matching research; it measures matcher performance rather than brand visibility inside AI assistants.
References
Run it on your own project
What the same prompts return across 10 engines is only knowable by running them.
Related research
Cross-category recommendations: +36% add-to-cart on item pages, +15% on cart pages
Recommending general merchandise to grocery shoppers — a milk frother with the milk — has long been hard; LLM retrieval plus real-time cart context moved production A/B metrics.
Industries · 2026-10-07Home services: 22% of homeowners already use AI to find a provider, and 80% of businesses don't know how to be found
The local services funnel is moving upstream: before a homeowner opens a website or calls, AI has already synthesised reviews, reputation and service descriptions into one answer.