A human review step for AI product copy is a handoff inside the same app: one person generates, another person checks the claims, and only then does a field go live. Squadexa AI can save a draft and compare it with the original before publish. Use those two controls as the review. This is a workflow, not a second app, and the catalog does not document a formal approval status.
The failure mode is a 50-product generation that is published the same hour. Pro allows up to 50 products in one CSV job, at $49 a month. That number is a generation cap. It is not a suggestion to skip reading. Basic is $9 a month and 10 products per job. Free is $0 and single-generation only. On every plan, the library can hold the draft until a person says the text matches the product.
Roles, without a permission system
Think of two jobs, even when one person does both on different hours. The operator prepares the input and runs the generation. The writer reads the output and decides what may be published. The catalog does not assign those jobs to Shopify staff permissions, and it does not list an approver role. Anyone who can open the app and press publish can change the live product. The protection is the agreement to use Save for later until the read is finished.
- Operator: product name, verified keywords, CSV within the plan cap, or a regenerate of a named product.
- Writer: measurements, certifications, claims, tags, and the decision to publish a field.
- Shared rule: a record stays saved in the library until that decision is explicit.
Save for later
Save for later persists edits without touching Shopify. It is the review inbox. The product library lists the records, with search, pagination, and status badges: created, updated, not created, and not updated. “Not updated” and “not created” mean the live catalog has not received this version. That is the column a writer should work from.
The edit view holds the title, the rich-text description, the price, the SEO title and description, the status, the product type, the vendor, and the tag chips. The writer can fix a bullet, delete a tag, and shorten a meta description while the record is still local. Price and status are on that screen because a later full publish can send them. The writer should look at them even when the project is “copy only,” so a full update is not clicked by habit.
Single-product generation stores a new record from a name and primary keywords. CSV bulk inserts many records when the job finishes. The progress display polls every few seconds and survives a refresh, so the operator can leave the job running and tell the writer when the library has the rows. The writer starts at the top of that list, not at the publish button.
Compare
When the product already exists, regenerate-existing is the review surface. The Squadexa page lists it on Basic. The app snapshots the current SEO fields, drafts a replacement, and shows the original beside the new version. The writer reads both. A handmade warning, a size, or a material on the original stays unless the draft copies it correctly. Save for later still applies if the writer needs another pass.
Compare does not publish. The publish choices are a single field or a fuller update. The single-field push covers title, description, meta title, meta description, or tags. Use it when only one of those survived the read. A fuller update can send title, description, tags, SEO fields, and status, and it can update the variant price. Reserve it for a record where every one of those values was checked.
Free single-generation does not include this compare flow in the plan card. A new record has no live original inside the app until you publish it. The writer still reads it. They simply are not looking at a side-by-side snapshot. Confirm the plan before you tell a teammate that every draft will have an original beside it.
The claims the writer checks
The writer’s list is short on purpose. Ordinary sentences about what the product is, who it is for, and one use can stay after a read. Specs and promises cannot stay unless they were supplied. The generation does not document a fact check or a filter that blocks a claim. Keyword-aware means the words you provided can appear. It does not mean an unsupported sentence was rejected.
- Every measurement matches a sheet or a sample, or it is deleted.
- Every certification is one you can show for this product, or it is deleted.
- Health, safety, and “best” lines are sourced or removed.
- Competitor names are removed unless someone wrote that comparison on purpose.
- Tags match the labels your collections already use.
- The SEO title and the meta description do not add a claim the body just had removed.
On a CSV, the writer checks the first rows completely, then scans the rest for the same error. A repeated invented seal is a source-sheet problem. Sending the other forty rows live spreads it. Pro’s 50-product job makes that scan more important, not less. Basic’s 10-product job is small enough to read in full, and it should be read in full.
Publish
Publish is the writer’s action, or the operator’s action after the writer has marked the record in whatever shared note you use. The app itself will not show an “approved” badge. Agree on a signal you can see: the description has been edited, the bad tags are gone, and the status badge is still not updated until the chosen person presses the update.
After publish, open one product in the admin and on the storefront. Confirm the field you meant to change is the field that changed. If it is wrong and this product was regenerated, revert restores the pre-regeneration snapshot onto the live product. The page does not state how long that snapshot is kept, and it does not promise a snapshot for every CSV row. For a CSV publish, your own export of the previous descriptions is the backup to hold before the run.
A review that generates 50 products and publishes them in the same sitting has skipped the middle step.
Questions about reviewing AI product copy
Can two people use the app?
The catalog does not publish a staff-permission model or an approver role. Two people who can both open the app need a working agreement: one generates and saves, the other compares and publishes. The software will not block the operator from publishing early.
Is there an approval status?
No formal approval status is documented. Use Save for later, the not-updated badge, and the compare view. Treat those as the review step.
What if the writer and the operator are the same person?
Split the hour. Generate and save. Do the claims list in a later pass. Publish one field when that pass is done. The same screens support that pause.
Save the draft, compare it when the product is already live, and publish only after a person has checked the claims. That review uses the tools on Squadexa AI, not a second product.