Which AI turns a recipe page into a clean recipe best?
Parsely-bench runs real recipe pages through combinations of AI models and web scrapers, then scores each extracted recipe against the page's own recipe data: ingredients, steps, title, times and servings, and whether anything was made up.
hourglass_empty
Loading results…
calculateHow scoring worksexpand_more
Every extraction gets a score out of 100, compared with the recipe data the page itself publishes (schema.org JSON-LD).
- Ingredients 40 · how closely the ingredient lines match
- Steps 25 · word-level overlap across all instructions
- Title 10 · word overlap with the recipe title
- Times & servings 10 · within 10% of the page's values
- Grounding 15 · every extracted line must appear on the page (catches made-up content)
Failed searches, scrapes or extractions score 0. "JSON-LD on" means the model was also given the page's structured data, as Parsely does in production; "off" makes it read the page text alone.