How we test recipe imports

1. Build a balanced source set

Use the same source set for every product or release.

• A standard recipe website with structured recipe data
• A social post with a complete caption
• A captionless video with spoken quantities
• A silent video with on-screen ingredients
• A video that splits details across speech and frames
• A clear recipe photo or screenshot
• A difficult image with glare, handwriting, or a cropped edge

2. Create a source truth record

A reviewer records each visible and spoken ingredient, quantity, step, time, temperature, serving count, creator, and source link. The reviewer must not use an importer output to create this record.

3. Score fields separately

Use field-level results instead of one subjective quality score.

• Ingredient name: exact, acceptable variation, incorrect, or missing.
• Quantity and unit: exact, convertible equivalent, incorrect, or missing.
• Step order: correct, partly correct, or incorrect.
• Time and temperature: correct, incorrect, or missing.
• Attribution: creator and source both present, partly present, or missing.

4. Record corrections

Count every manual field correction. Also record the time between import start and a recipe that is ready to cook. These two measures expose errors that a completion rate can hide.

5. Publish limits with results

A result must state the source count, source mix, test date, product version, reviewer count, and excluded failures. Small tests support product diagnosis, but they do not prove universal accuracy.

Current status

Recivy publishes this method before it publishes benchmark results. No measured comparison result appears on this page yet. This prevents an untested claim from looking like evidence.

Use this method

Test a source with the recipe importer. Then use the recipe saving app checklist for the wider product workflow.

Read the Recivy editorial and corrections policy.