What makes good source data

Enrichment writes from your existing product data and will not invent facts beyond it. Input quality sets a ceiling on output quality, so this is the first thing to check.

Enrichment amplifies, it does not fabricate

If a record says 100% cotton, enrichment can write about breathability and care. If the record never mentions material, the output won’t either — the gap stays a gap rather than being filled with a plausible guess.

That is the intended behaviour. Invented specifications are worse than missing ones.

What a workable record contains

The more of these you have, the more there is to write from.

Present in the record	What it enables
A real product name, not a SKU code	Everything. A name like `SKU-4471-BLK` gives the tool nothing to work with
Materials, dimensions, capacity	Concrete description copy and factual Q&A answers
Use cases or intended audience	Intent Q&A, and copy that speaks to a buyer
Distinguishing features	Content that separates this product from the next one
Existing description, even a rough one	A starting point and a signal of what matters

What to fix first

Products with no usable name. These fail or produce generic text. Fix them or exclude them.
Placeholder text. “Description coming soon” and lorem ipsum produce exactly what you would expect.
Your thinnest records. Sort by description length. The bottom of that list either needs a data pass first or should be left out of early batches.
Inconsistent units and terminology. Standardising these reduces contradictions between generated fields.

Errors are usually a data problem

Products that fail to generate appear on the Errors tab, and retrying them is free. But a retry on an unchanged record produces the same result. Read a few failures before retrying — the common cause is a record too sparse to work from.

What consumes an entitlement

Entitlements are consumed when products are sent for enrichment, not when you apply results. A product counts once whether or not it has variants. Retrying errored products is free; re-running a product that generated successfully consumes another entitlement.

A quick audit before your first batch

Pick 20 products at random. For each, ask whether a copywriter with no product knowledge could write a useful paragraph from the record alone. If the answer is no for most of them, fix data before spending entitlements.