Human-review checklist for AI-generated product content

A compact, repeatable review workflow to catch hallucinations, privacy leaks and factual mismatches in AI-drafted product titles, descriptions and specs.

Editorial illustration: pale mint background with a browser-style product card showing a SKU line and a separate paper checklist and pencil—two clear groups showing product data and a reviewer checklist.

If you use AI to draft product titles, descriptions or specs, make one change first: force every generation to start from a short authoritative snippet copied from the product record, and record that snippet plus the exact prompt and the reviewer’s checklist on the product. That single habit reduces hallucinations and gives you the evidence you need if something looks wrong.

Contents

Authoritative snippet — fields to copy before generation

Before you generate text, copy a short snippet from the canonical product record. Keep it deliberately small and factual: the model must rewrite from those values, not invent them. At minimum include:

  • SKU (exact)
  • GTIN / MPN if available
  • Price and currency
  • Weight and units
  • Dimensions (width × height × depth) and units
  • Stock status (manage_stock, stock_quantity where used)
  • Availability text or shipping note that affects buying

These fields match the canonical product properties used by WooCommerce and the schema.org product vocabulary; preserve exact values when you later update JSON‑LD or the product API. See WooCommerce product fields for the authoritative field names and schema.org for product property names: WooCommerce REST API product fields, schema.org Product.

Safe prompt template (inject snippet and constraints)

Always inject the snippet into the prompt so the model rewrites from verified values. Record the snippet and the exact prompt alongside the draft for auditability.

{SNIPPET}
SKU: 12345-XL
GTIN: 0123456789012
Price: 69.00 GBP
Weight: 0.45 kg
Dimensions: 30 × 20 × 5 cm
Stock: in_stock (qty 12)
Shipping: Standard UK 2–4 days

PROMPT:
Rewrite a short product title (max 80 chars) and a 50–120 word description for the product above. Do not change or infer numeric values (price, weight, dimensions, SKU, GTIN). Do not add medical or safety claims. Keep language factual and suitable for a small ecommerce shop.

OUTPUT: Title:\nDescription:

Use that template (or a similar constrained prompt) for every generation and save both the snippet and the prompt to the product record so a later reviewer can reproduce the input.

Human-review checklist — must-run checks for every AI draft

Run these checks before publishing. Treat a single failed check as a blocker until it’s resolved.

  1. SKU match — does the draft show the SKU exactly as the canonical value? If the SKU appears in title or description, it must be byte-for-byte identical to the product record.
  2. GTIN / MPN match — any global identifiers must match the source record.
  3. Numeric specs identical — weight, width/height/depth and units in the draft must equal the source snippet. If the draft omits these specs, confirm whether omission is allowed; do not rely on the model to invent them.
  4. Price and offer text — ensure price and any shipping or availability language reflect the authoritative offer. If price wording (sale vs list) changed, verify with the pricing admin or the offers API.
  5. Brand and trademark use — confirm brand names are allowed and spelled correctly. Flag ambiguous trademark references for brand/legal review.
  6. Regulatory/medical/safety claims — any claim that could be regulated must be escalated for legal/compliance review. Don’t treat the model as an authority on certification or compliance.
  7. Structured-data parity — check JSON‑LD Product fields (sku, gtin, weight, dimensions, offers, availability) match the canonical snippet exactly after the new copy is placed on staging.
  8. Style & clarity pass — edit for tone, readability and shop style; don’t add facts during this pass.
  9. Approval record — save reviewer name, date/time and checklist pass/fail on the product record (or a review log) before publishing.

Quick copyable verification checklist (one line per check you can paste into your review tool):

Checks performed: SKU match? | GTIN/MPN match? | Numeric specs match? | Price/offer accurate? | Brand/trademark OK? | Regulated claims flagged? | JSON‑LD parity? | Editor initials: ______ | Date: YYYY-MM-DD

Where to save the prompt, snippet and reviewer approval

Add reviewer metadata to the product record so checks are retrievable via your product API. On WordPress/WooCommerce you can add or register product meta/REST fields to store the snippet, prompt and approval result — for example with register_post_meta / register_rest_field. See the WordPress REST API guidance and WooCommerce product fields for implementation details: Extending REST responses (WordPress), WooCommerce product API.

If your CMS cannot store reviewer fields, keep a simple spreadsheet or internal report with product ID, snippet, prompt and reviewer checklist rows. The important part is reproducibility: someone else can re-run the same prompt against the same snippet and see the same inputs.

Policy & privacy — pre-checks before sending product data

Check your chosen AI provider’s current data-handling and retention documentation before sending any product data. Some fields might be sensitive (for example, custom notes with personal data) and should be redacted. Do not send customer personal data to third-party models unless your contract and internal approvals explicitly allow it. For provider guidance, read the platform documentation such as OpenAI’s data controls: OpenAI data handling guidance.

Staging QA — verify JSON‑LD and structured data

Always publish the revised copy to a staging site first. On staging:

  • View the page source and locate the JSON‑LD Product block; verify schema.org properties (sku, gtin, weight, dimensions, offers.price, offers.availability) match your canonical snippet. See schema.org Product and Google’s product structured-data guidance for expected properties: Google product structured-data.
  • Run a structured-data validator or your usual QA tool and resolve any mismatches before publishing. (See our guide on validating product markup for practical test steps and tools: Validate product structured data.)

Only publish when the staging JSON‑LD values equal the canonical snippet and the reviewer checklist is signed off on the product record.

Fictional example — labelled and hypothetical

Fictional example: A small seller used an AI assistant to generate 50 product descriptions. After a random spot-check revealed a mismatched weight on one listing, the seller required that every AI-generated description include a saved prompt/snippet and a reviewer initials field. The reviewer checklist now blocks publishing until SKU, dimensions and price are identical to the product record.

Two-minute checklist — get this working now

  1. Pick one product and copy its canonical snippet into the prompt template above. Generate a draft and save the snippet + prompt on the product record (or spreadsheet).
  2. Run the core human-review checklist above against that draft; if your CMS supports it, add a reviewer meta field using the WordPress REST guidance so the result is visible via the API (WordPress REST).
  3. Publish to staging, check the page JSON‑LD values match the snippet and only then publish live (Google product structured-data and structured-data validation guide).

Related practical reads: Writing product descriptions with AI: keep the facts under control and Validate product structured data with Google’s Rich Results Test and URL Inspection.