A supplier sheet says 400 ml. Your current listing says 450 ml. The photograph shows a mug, but proves neither figure. AI can turn that disagreement into a fluent, confident description with the wrong specification. Preparing Shopify product data before AI writes the copy means resolving those conflicts first. The payoff is a draft your reviewer can check against approved evidence, rather than another description that needs its own investigation.
Give the writer facts your team has checked
Shopify’s current documentation describes Sidekick writing product descriptions from details you supply. It also warns that generated text can include benefits or facts you haven’t provided. Merchants remain responsible for accuracy and must review the result before saving. On a published product, saving a new description displays it immediately.
Our recommendation at ZAGO is to put a small approved fact record between your source material and the writing tool. This is an editorial control, not a Shopify requirement. The record should identify the exact item, list verified facts and show where they came from. Keep its version and any excluded claims alongside it.
The record doesn’t need to be elaborate. It does need someone who can resolve disagreements. If the supplier sheet conflicts with the listing, the copywriter shouldn’t pick whichever value sounds more plausible. A product owner must confirm the correct value and record why it supersedes the other one.
Build a product fact record your team can use
Start with a stable product identifier and the exact variant. Add the plain product type, approved values with units, and a source reference precise enough for another colleague to find. Record the source version or date, the reviewer, and the approval date.
Keep allowed claims separate from unknown or disputed fields. A blank cell is ambiguous: it could mean nobody checked, the field doesn’t apply, or the supplier hasn’t answered. Those situations need different treatment. Zero is a value, not a substitute for unknown.
Separate facts from brand language
“Stone grey” can be an approved marketing colour without proving the material is stoneware. Keep the colour label in its own field and verify material independently. Photographs can help identify which item you’re discussing, but they cannot establish capacity, composition or certification. Another similar product is not evidence either.
Brand vocabulary belongs in a separate style brief. It can guide sentence length and preferred colour names. It cannot authorise a benefit claim. “Easy to clean” still needs support, even if it sounds less technical than “dishwasher safe”. Style may vary between channels; factual values must stay fixed.
Care and safety questions need a named fact owner. If an unresolved fact is required for publication, the product stays blocked until that owner resolves it. Omitting a claim from generated copy doesn’t automatically make the product ready to sell.
A worked example: the mug with two capacities
Consider a fictional product family with 250 ml and 400 ml stoneware mugs. The supplier record confirms the larger variant holds 400 ml, while its old description incorrectly says 450 ml. Care instructions remain unconfirmed. This is an illustrative example, not a ZAGO client record.
Some facts can belong to the shared product record, provided they apply to both variants. Capacity belongs to each variant. The 250 ml value must never become a specification for SKU MUG-STONE-400 merely because both mugs share a product page.
For a single family description, state the available capacities explicitly rather than attaching one capacity to every mug. Variant-specific copy should use only the selected variant’s approved value. Confirm shared material across both variants before using it as a family-wide claim.
An illustrative ready-to-draft record
- Identity: fictional Stone Mug family; SKU MUG-STONE-400; selected variant 400 ml.
- Plain type: mug. Verified capacity: 400 ml.
- Material: stoneware, approved only after supplier confirmation for this variant.
- Brand vocabulary: stone grey is the approved marketing colour label, separate from material.
- Evidence: hypothetical supplier specification, revision B, dated 1 October 2026, variant row MUG-STONE-400. This is a sample reference, not a real vendor document.
- Approval: designated product data owner reviewed record version 2 on 4 October 2026 in this hypothetical example.
- Allowed claims: mug; 400 ml capacity; stoneware; stone grey colour.
- Exceptions: old 450 ml description superseded; care unconfirmed. Do not claim dishwasher safety.
- Status: ready for partial copy with care excluded. Publication requires the owner’s decision on whether missing care information blocks release.
Keep the incorrect 450 ml value in the exception history, clearly marked as rejected. Don’t leave it beside 400 ml in an undifferentiated bundle of source text and expect the model to decide.
Give the writer a boundary, then check the draft
Once the record is approved, the writing instruction can be short:
Draft a description for the exact product or variant identified in the approved record. Use only approved facts and allowed claims. Follow the separate brand vocabulary brief without adding attributes or benefits. Do not infer anything from images, similar products or rejected values. Put missing and contradictory data in separate review notes, not customer-facing copy. Do not publish or save changes to the live product.
A hypothetical output for the record above could be: “A stoneware mug in stone grey, with a 400 ml capacity.” The review note would flag unconfirmed care instructions and the superseded 450 ml description.
That sentence is deliberately modest. Add richer copy when you have richer approved information, rather than asking the model to make sparse evidence sound substantial. The prompt is not a guarantee. A reviewer must compare every factual claim with the approved record, including implied benefits.
Store the facts in Shopify without assuming access
Shopify category metafields map to specific categories in its product taxonomy. Some values are added or suggested through AI predictions. Treat those suggestions as candidates for validation, not evidence that a product has an attribute. Category attributes are useful for standardised characteristics relevant to a product type.
Custom metafields can store specialised information such as care instructions or dimensions. Merchant-specific approval status and evidence references can live in suitable custom fields or an accompanying working record. Storing a value doesn’t automatically display it on the storefront: that needs theme support or a connection to the relevant field.
A working sheet is enough if ownership and version control stay clear. An existing PIM may be the better home where your team already manages supplier evidence. We wouldn’t buy a PIM solely to begin this process.
Check exactly what your writing tool or connector can read. Identify the available product fields and variant fields, then verify which approved values reach the drafting step. Don’t assume Sidekick or another AI tool automatically consumes every metafield or external record. Internal review notes should stay separate from customer-facing content.
Handle CSV as a data change, not a harmless shortcut
Shopify’s product CSV documentation supports product metafields but excludes variant metafields from product CSV import and export. Variant metafields can be edited through the variant bulk editor. CSV data also has dependencies, and imports can alter existing records.
Use a working copy for reconciliation. Don’t assume a product export contains all your data, or strip columns and import without checking the consequences. Test a small representative set and review the import result before wider changes. Include a multi-variant product and an unresolved record so your checks cover more than clean examples.
Make approval repeatable
Before drafting, use this preparation checklist:
- Name the approver and the owner of unresolved care or safety questions.
- Version the record and retain its evidence reference.
- Define required fields separately from optional copy details.
- Resolve duplicate records and confirm stable product and variant identifiers.
- Distinguish zero, unknown and not applicable; use consistent units.
- Keep published copy separate from approved source facts.
- Route exceptions back to the fact owner instead of asking AI to fill gaps.
- Review a small representative set before extending the process across the catalogue.
Measure factual corrections per draft, review minutes, and the reasons records remain blocked. Repeated capacity corrections point to a source or mapping problem. Repeated prompt edits won’t fix that reliably.
Assign an update trigger too. When a supplier changes a specification, the owner should reapprove the affected record and identify copy that needs review. Approval is only useful while the evidence remains current.
Takeaways
- Resolve conflicting facts before AI drafts a description.
- Keep shared product facts separate from exact variant specifications.
- Use approved evidence, explicit exclusions and a named reviewer.
- Check tool access and every draft claim before publication.
Prepare your product data with ZAGO
If supplier files and Shopify listings disagree, talk to our team about AI for ecommerce. We can help define the fact record, map the fields and put a review step between the AI draft and your product page.
Frequently asked questions
What product data should I approve before using AI to write Shopify descriptions?
Approve the product identifier, exact variant, plain product type and factual values with units. Include source references, a record version and reviewer, allowed claims, and explicit unknowns or disputes. Keep brand vocabulary separate from verified specifications.
Does Shopify Sidekick automatically use all product metafields?
Do not assume it does. Shopify describes Sidekick generating descriptions from supplied details. Check which fields your chosen tool or connector can access and verify that the approved product and variant values reach the drafting step.
Can Shopify product CSV files import variant metafields?
Shopify product CSV import and export supports product metafields, but not variant metafields. The variant bulk editor can edit variant metafields. Use a working copy and test representative records before broader imports.
Should missing care instructions block an AI-written product description?
Exclude unsupported care claims from the draft. The product owner must decide whether the missing information is required for publication. A partial draft can be ready for review while the product remains blocked from release.
