> This page is for Developer.

> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.bigcommerce.com/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.bigcommerce.com/_mcp/server.

# Deciding what to change

> How to diagnose whether weak output comes from your inputs or your source data, and what to change before re-running.

Re-running consumes an entitlement per product, so it is worth knowing what to change first. Re-running the same products with the same inputs and the same data produces much the same output.

## Let the scores narrow it first

Which factor scored low points at which input to fix: accuracy at your source data, consistency at the record or the batch mix, brand adherence at your guidelines. See [Understanding quality scores](/developer/docs/ai-tools/catalog-enrichment/review/quality-scores).

That narrows the question. The next step is deciding which of the two levers to pull.

## Which lever is short?

Two things determine output. Read a weak result and decide which one failed.

**Source data** sets what there is to say. If the output is accurate but empty — correct sentences carrying no information — the record had nothing in it. No amount of input tuning fixes that.

**Inputs** set how it is said. If the output is substantive but sounds wrong — off-tone, generic, using words you avoid — the facts were there and the instructions were not specific enough.

| What you see                          | Change                                                                                                                                                       |
| ------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| Accurate but says nothing             | Source data. Add materials, dimensions, use cases                                                                                                            |
| Sounds generic or off-brand           | Brand guidelines. Add do's, don'ts, vocabulary                                                                                                               |
| Right facts, wrong emphasis           | Additional context. State what to lead with                                                                                                                  |
| Terms shoppers use are missing        | Search keywords                                                                                                                                              |
| Copy reads as keyword-stuffed         | Shorten the keyword list                                                                                                                                     |
| Good on some products, poor on others | Batch composition — split into narrower batches. See [Planning your batches](/developer/docs/ai-tools/catalog-enrichment/before-you-enrich/planning-batches) |

## What costs what

| Action                             | Cost                        |
| ---------------------------------- | --------------------------- |
| Editing a field during review      | Free                        |
| Retrying products that errored     | Free                        |
| Re-running products that generated | One entitlement per product |

This shapes the order of operations: edit what is nearly right, and re-run only after changing something that will produce a different result.

## Re-running with new inputs

You can repeat an enrichment on the same products with revised brand voice and context. Change one thing at a time where you can — revising guidelines, keywords, and context simultaneously tells you the output improved without telling you why.

## Test on a smaller batch

Before re-running 250 products, re-run 25 with the revised inputs. If the change worked, apply it to the rest. If not, you have spent 25 entitlements learning that rather than 250.