Shopify writes product descriptions for free. That was never the problem.

Field notes on the distance between a good AI feature and the system a catalog actually needs.

There is a button in the Shopify admin that writes product descriptions. It is free, it lives inside the product editor, and it is genuinely good at its job.

We mean that. We build on Shopify and we want its AI to win. This piece is about the day the button meets your catalog.

First, the nice things, because they're true

The facts, from Shopify's own pages. Shopify Magic drafts a description from a prompt: the title, a few keywords, a tone. It ships with every Shopify plan at no extra cost, nothing to install. And Shopify keeps investing in this layer; its Magic landing page now leads with Sidekick, whose Writer persona pitches product copy directly.

Two details in the documentation matter for what comes next. Text generation wants a desktop browser, phones are not supported. And Shopify states plainly that you are responsible for the accuracy of everything the AI writes, because generated text can sound plausible and still be wrong. Their own blog goes further: review every generated description before it goes live, since the model can invent details.

For one product, that loop is honest work. Prompt, read, fix the invented spec, publish.

Every catalog we open

Now put the loop inside a store doing real volume. In almost every catalog we open, someone has already found the button, and you can tell exactly where. The tone shifts mid-collection. A shirt is brushed organic cotton in one variant's copy and a poly blend in its sibling's. The care note contradicts the size chart it sits next to.

For a while we read this as a prompting problem. Write a better prompt, save it in a shared doc, make the team use it. We built that doc (it was very tidy, and nobody opened it twice).

The drift kept coming back. Because the drift was never about the words.

The question the button never asks

Most catalog content problems are not writing problems. The button answers one question, what should this product say, and answers it well. A catalog runs on different questions. Is this claim true against the spec sheet? Does this variant contradict its siblings? Is the material called the same thing here as everywhere else in the collection? Who approved this, and if it is wrong, what exactly do we roll back to?

The generator has no position on any of that, and no memory either: it starts fresh on every product, blind to the catalog around it. Even Shopify's own admin splits the work this way, quite reasonably. The bulk editor applies changes to existing values across many products, prices, channel visibility; it edits, it does not write. Clicking the Magic button once per product is a plan the way a teaspoon is a dredging strategy.

That gap has a name. A feature solves the moment. Infrastructure solves the repetition, with a record.

What running it looks like from our side

This is the layer we work in, so read this as our practice, not a law of nature. Inside ContentScope, our catalog content system, no text reaches a store on a single model's word. A writer agent drafts against the product's actual data. A critic agent then attacks the draft with the catalog's own rules: claims against specs, terminology against the controlled list, variants against their siblings. A human approves before anything publishes. Autonomous with human oversight, because Shopify's accuracy warning does not shrink at scale, it compounds.

That pipeline moves roughly 5,000 products a day and has put 642,000+ SKUs through the same discipline. The volume is not the brag. The brag is boring: the last SKU reads like the first one.

Where the free button wins

Honest disqualification, because we would rather lose you now than waste your quarter. Below our own floor of $250K a month in store revenue, Shopify Magic plus a careful reading habit is genuinely enough. Use it, check what it writes like Shopify tells you to, and keep your money.

Above that line the math flips. Consistency stops being taste and starts being returns, support tickets, and a brand that reads like it was written by committee, one product at a time.

Taped to the wall

• A feature answers a question once. Infrastructure answers it every time, the same way, with a record of who approved what.
• If AI writes into your catalog, verification and rollback are part of the writing, not accessories.
• Shopify's accuracy disclaimer is an operations requirement wearing a legal footnote's clothes.
• Below the consistency line, the free button wins. Above it, the discipline does.

If you want to know which side of that line your catalog sits on, OKART's Infrastructure Stress-Test maps it, read-only, in a few days.

The button is fine. The catalog is the job.

Sources

1. Automatically generating product descriptions: per-product generation workflow in the admin, desktop-only text generation, merchant responsibility for accuracy.
2. Shopify Magic: the Magic feature suite, plan and language availability, store-level data isolation.
3. Shopify Magic for product descriptions: included with Shopify plans, keyword-and-tone workflow, scope of the built-in feature.
4. AI Product Descriptions With Shopify Magic: recommended inputs and the review-every-description-before-publishing guidance.
5. Sidekick, Shopify's AI assistant: current landing page for Shopify's merchant AI, Writer persona, included with Shopify plans.
6. Bulk editing products: the bulk editor applies changes to existing values across many products at once.