Skip to main content
Quality review reads the copy enrichment produced and records findings — named, self-explanatory notes about what looks wrong. The checks are advisory: they never block publishing and never change your data. They only report. You need settings permission to view the page and write permission to change it. There is deliberately no 0–100 quality score. A number can’t tell you what is wrong with a description; the findings can, and the count of them is what sorts the worst products to the top.

​
Turn it on

Settings → Quality Review has a master switch, Run quality checks, off by default. While it is off nothing in your organization is scored, whatever the individual checks say. Turn it on and every enrichment run is scored when it finishes. Scoring is separate from enrichment. You can change anything on this page and re-score an existing run from its job page to compare — without re-enriching the products.

​
The checks

Six checks, each with its own switch, severity, and per-property exclusions. A few details worth knowing before you rely on one:
  • Near-duplicate strips your boilerplate patterns before comparing, so your own standard sentences don’t register as duplicates.
  • Length outlier needs at least five scored products for the run’s median to mean anything.
  • Banned phrase matching is exact string matching, not a model call. It can never invent a phrase you didn’t configure.
  • Field contradiction is a deterministic string comparison. It only checks bullets whose own prompt instruction asks for the value exactly; the wiring report lists the ones it skips.
  • Prompt rule adherence only runs when the product declares a gender.

​
Billable checks

Prompt rule adherence and Fabrication risk ask a model, so each adds a model call for every generated property on every scored product. That’s why they are off by default and badged Billable LLM call in the app. The model is chosen for you and is a cheap one by design. Re-scoring a run doesn’t re-enrich it, but it does re-run whatever checks are enabled — including the billable ones.

​
Severity and exclusions

​
Warn counts as a defect worth review. Info is recorded but reads as a note rather than a problem. Use Info to keep a check running while you decide whether to trust it.
​
Comma-separated property keys the check must not run for. Matched exactly and case-sensitively, so spell the key the way the property does. Leave it empty to check every scored property.
A check you have never saved shows a Default badge — it has no stored setting and is following the built-in behavior shown. Saving pins the values you see.

​
Your house copy

Two lists describe your own writing, so the checks judge the generated copy rather than your standards.
​
Sentences your copy repeats on purpose. They’re stripped out before the near-duplicate check compares two products — without that, a standard line you add to every description makes every product look like a duplicate of every other. One pattern per line; * matches any run of text, so one pattern can cover the same sentence with different product names in it.
​
Wording that must not appear in generated copy. One phrase per line.
Each list is read by exactly one check. Boilerplate patterns do nothing while near-duplicate copy is off, and banned phrases do nothing while banned phrase matching is off. The page warns you when that happens.

​
Where findings show up

​
The run’s own findings, worst products first, with counts for how many were scored, how many have findings, how many hit a check error, and how many couldn’t be judged. The list itself is capped on large runs, but the counts always cover the whole run. This is also where you re-score a run after changing settings.
​
A Quality column per product: a dash for never scored, Clean for scored with nothing found, and a count of findings otherwise. Never-scored and clean read differently on purpose. You can filter the list down to products that have findings.
​
The same findings summary for the batch’s most recent enrichment run, on the batch page — so you can look at quality before approving. The batch header’s Enrich N Products button starts that run, and re-scoring is available from the summary.
A check that couldn’t run is recorded as an error, never as a pass. A check that had no evidence to judge — typically a product enriched before the generator started recording what it showed the model — is reported separately and doesn’t count toward the finding total.

​
Enrichment wiring report

The same page carries an Enrichment wiring report, which audits your enrichment configuration rather than any run: per category and property, which declared context fields are actually populated and at what rate, which spec-sheet bullets bind to an empty or ambiguous field, and which placeholders a prompt references but never declares. It is computed against your current catalog and config, not against a job, and you generate it on demand. Use it when findings point at a configuration problem rather than a copy problem.

How enrichment works

What produces the copy these checks read.

Content settings

Tone, brand guidelines, and the grounding thresholds behind generation.

Reviewing enriched products

Enrichment status and completeness alongside quality findings.

Jobs

Where a run’s findings and the re-score action live.