How Google Finds Every Restaurant in Japan — And Why Your Full-Text Search Can't

I recently scraped every unagi (eel) restaurant in Japan using the Google Places Text Search API. The results were absurdly precise: 1,914 restaurants across all 47 prefectures, with a noise rate under 1.6% . No ML pipeline. No fine-tuned classifier. Just API calls. Here's how that's possible — and what it reveals about the gap between "search" as most developers implement it and search as Google does it. The Task I wanted a comprehensive database of unagi specialty restaurants in Japan. The naive approach: query "うなぎ 鰻 {prefecture}" for each of the 47 prefectures and collect the results. That's exactly what I did. One query per prefecture. Save to JSONL. Done. The Surprising Precision When I audited the results, here's what I found: Category Count Restaurant name contains eel-related terms (うなぎ, 鰻, うな, etc.) 1,322 Name doesn't match, but reviews mention eel 562 Potential noise (no eel reference anywhere) 30 (1.6%) And even among those 30 "noise" candidates, most were legitimate eel re