Reverse-engineering Google's image labels: 100 tags per photo, product attributes reuse Shopping filters

gaganghotra_ · x · 2026-11-10

RESONEO's latest "Understanding Google" installment dissects what Google's vision models actually read from an image:

• A photo is first a list of 100 labels with 0-1 confidence scores; top ranks are Knowledge Graph entities (/m/, /g/ MIDs like Footwear 0.99, Snow 0.98)
• Lower down sits a second vocabulary: /j/ dimensions and /a/ values describing object attributes and scenes (Shoe style → Lace up, Round toe), which reuse Google Shopping filter names — found across 67,499 Gemini-generated images and cross-checked against Shopping filters and the Vision API
• The piece also maps Google Maps' photo-to-menu pipeline: one bowl of ramen exists at three levels — the generic label, the dish entity (Tonkotsu ramen), and the restaurant listing's specific dish ID
• Includes testable methodology and downloadable test photos for image SEO/AEO practitioners

Original post →

More from Apps

• VC predicts EvenUp's AI transformation of personal injury law will become an HBS case study — aaref · 2026-11-10
• Other Words lets you click any token and see what GPT-6 Luna almost said — csvoss · 2026-11-10
• Tavus stress-tests its PAL digital human at a coffee kiosk against hijacking attempts — ycombinator · 2026-11-10
• Dev builds a wine-making learning game with Claude Fable 5.2 — 'make a game to learn' — kieranklaassen · 2026-11-10
• Creator makes 5-minute sci-fi film 'I WAS HERE' using ChatGPT's new ImageGen — VraserX · 2026-11-10
• Leak: XChat is building a Calendly-style "Book a call" card — nima_owji · 2026-11-10

Apps channel →