Fourteen years of blog posts, seven languages, one laptop: an open-weight model did our hreflang backfill

GoodBarber, a no-code app platform, recently solved a complex SEO challenge by using open-weight AI models to backfill hreflang tags for 5,892 blog posts across seven languages. Facing a decade of fragmented content without translation metadata, the engineering team utilized a local MacBook Pro running Ollama with Gemma2 and BGE-M3 models. By implementing a strategy that combined a 90-day publication window, deterministic checks, and a two-model verification process, they successfully mapped 91.2% of their URLs. The project highlights the efficiency of running smaller, open-weight models locally for specific data-processing tasks, avoiding the costs and rate limits of hosted APIs. The team emphasized that task design—specifically narrowing the candidate pool—was more critical to accuracy than model size. This successful experiment underscores the accessibility of modern AI tools for practical, real-world engineering problems, which the company plans to discuss further at their upcoming Hacktoberfest event in Ajaccio.