Product Data

Product Data uses AI to predict product attributes your feed doesn’t already provide cleanly — most commonly color and category — and gives you a review workflow to confirm or correct each prediction before it’s used anywhere else (search, filters, boosting).

How it works

flowchart TD
    A[Product imported] --> B{Attribute present\nand reliable?}
    B -->|Yes| C[Feed value used]
    B -->|No| D[AI predicts a value]
    D --> E{You review it}
    E -->|Confirm| F[Confirmed value used\neverywhere]
    E -->|Reject| D

What it does

🤖 Predicts missing or messy attributes

Color is predicted from product images; category is predicted from title and other text; other attributes can be predicted depending on your catalog.
🛡️ Never overwrites your feed silently

A predicted value is staged separately from your feed’s original value. It only takes priority once you (or a bulk-confirm run) explicitly confirm it — confirmed values always win, your original feed value is the fallback, and a prediction only surfaces when neither exists.
🗂️ Works per product or in bulk

Predict and review one product’s attributes at a time, or run a bulk prediction job across the whole catalog for a given attribute.

Configuring in Panel

Product Data has two parts in Panel: a campaign that controls what gets predicted, and a per-product review screen where predictions are confirmed.

Product Data campaign

Open Site → Campaigns, and open or create a Product Data campaign.

General settings — Name and an Active toggle (there’s no “default campaign” concept here).

Attribute Prediction — turn on which Product Attributes (e.g. title, category, color, material) and Image Attributes (e.g. image type) you want predicted. The available attributes are read from your catalog schema.

Advanced settings

Reference catalog — optionally ground predictions (e.g. category classification) against another site’s catalog instead of this one.
Color — once color is a selected attribute, tune Color Family Threshold (minimum share of an image a color family needs to be reported), Maximum Color Families (cap on how many distinct families are returned), and Color Threshold (how many individual shades are kept within a family).

Reviewing predictions

Open Site → Products, click into a product, and go to its Predictions tab (the tab shows a badge with how many predictions are pending). From there:

Run Data Prediction (or Refresh Data Prediction if it’s already run once) triggers prediction for that product.
Each attribute shows as confirmed (accepted), predicted (awaiting your review — click to accept it), or empty.

For catalog-wide changes, a bulk prediction run can predict an attribute across every product at once, optionally auto-confirming the results instead of requiring per-product review.