Improve with AI

← Back to documentation

What the suggestions cover

Improve with AI on the analysis page opens the improve page and starts generating. The button is only enabled while at least one of the things below can still earn points; when the listing already scores full marks on them, there is nothing for it to do.

• Description. One rewritten description, at least 150 words, in the language of your current text.
• Title. One alternative (occasionally more), under 70 characters and different from the current title.
• Category fields. Values for the category's attributes that are still empty, chosen only from the values Shopify's taxonomy allows for that attribute, and only where your text supports the choice. When the category audit flags a filled value as wrong, a replacement is suggested too.
• Category. When the product has no category, a leaf category from Shopify's taxonomy that fits the text; a category is only assigned when one truly fits, because none is better than a wrong one. When the current category looks wrong, the best replacement the audit found, which you should check before applying.

The page also lists what AI cannot write for you: images (with a link to the product's media), the store policies (with a link to Shopify's policy settings), and the remaining gaps from the breakdown under More opportunities.

The rules the suggestions follow

• No invented facts. A suggestion keeps every factual claim in your listing and adds no materials, measurements, certifications or other facts that are not already in your own text. Category values are picked only when the title or description clearly says so; otherwise nothing is picked.
• Same language. Suggestions are written in the language of your existing listing.
• Written for the rater. The model is given the same rubric that scores your product, then each draft is scored by the fixed rating model and rewritten against the critique, up to two rounds, stopping early when a round does not improve. What you see is the best draft, with the points it is expected to earn.
• Nothing is applied without you. A suggestion only fills the form on the improve page. Your product does not change until you press Save, Apply category, or Apply all suggestions and confirm.

Reviewing and editing

• Title: each suggestion has a Use this button that puts it in the title field, and a dismiss button. Edit the field freely afterwards.
• Description: a plain text editor with a word count, the suggested description below it with a Use this button, and a row of shopper asks the description should answer.
• Category fields: chips per attribute with the suggested value marked; Accept the suggestion (or Accept all N suggestions) fills them, and you can pick other allowed values or add your own. Attributes that come from a product option (Color, when it is a variant option) are set by the option and cannot be changed here.
• Category: Assign a category or Change the category shows the suggested one and a search box for any other. Apply category (or Change category) writes it immediately, then the fields for the new category load so you can fill them.

Some attributes are saved together (Color with Pattern, for example): a suggestion carries both, and a save with only half of such a pair is stopped with a note rather than written.

Apply all suggestions, the page's primary action, takes every suggestion that is expected to earn points at once: it fills the title and the description, accepts every category field suggestion, saves, then assigns a pending category and fills that category's fields too. It first shows exactly what will be written and asks you to confirm; you can tick Don't ask again on this device.

The badges (+N pts, up to +N pts) are the points a check is expected to earn from that suggestion, and the line at the top adds them up into an expected score. The real score after a rescan can differ slightly, because the catalog checks are live.

Saving: what is written to your store

The title, description and picked category fields are saved together from the save bar (Save and Discard), as one all-or-nothing write: if any part cannot be written, for example a category field Shopify does not accept on that category, the title and description are put back and the product is left exactly as it was, and the message says so. Leaving the page with unsaved edits asks you to confirm. Only the fields you changed are written, and nothing else on the product is touched:

Field	What is written
Title	The product title.
Description	The product description as paragraphs. The editor is plain text, so a description that had custom HTML formatting (headings, tables, embedded images) loses that formatting when you replace it. Paragraph breaks are kept.
Category	The product's category in Shopify's taxonomy. Written when you press Apply category.
Category fields	The values you picked, as the category metafields Shopify defines for that category. Octolift creates the field definitions and the taxonomy value entries a store does not have yet.

After a save the Back to analysis button reads Rescan, which scores the product again.

History and undo

Every change you apply is recorded together with the previous value, the suggestion it came from, and whether you applied it as suggested, edited it first, or wrote your own. On the Products page, press View changes on the product's row to see them, newest first.

• Title and description changes can be undone there with the Revert button on each (Revert all appears when several can be undone), as long as the product still holds what Octolift wrote. If you have edited that field since, the undo is refused rather than overwriting your work.
• Category and category field changes are shown but not undoable from the app, because there was usually nothing to go back to; change them on the product in Shopify.
• An undo is recorded as a change of its own, pointing at what it undid, so the history stays a complete account of what happened.
• History is kept for 30 days, after which it is deleted; a change can only be undone within that window. The product's score history stays.

What it costs

Each generation uses one product improvement from your monthly allowance on the default model (a stronger model uses more per product; see Plans, allowance and models). Regenerating suggestions for the same product uses an improvement each time. Reviewing, editing, saving and rescanning use nothing. When the allowance is used up, the improve page says so and shows when it renews; scoring keeps working in the meantime.

Related guides