A catalogue of thousands of products: AI under control

On a catalogue of a hundred products, AI enrichment is a matter of review: one person reads every product page, compares it with the supplier’s text, corrects. On ten thousand, that is no longer possible. And the risk changes in nature: a badly written prompt does not produce one error, it produces the same one everywhere, overnight, with a dashboard showing green.

This guide is for teams managing a large catalogue who want to enrich it with AI without losing control over what it claims. The basic method, page by page, fits in four steps: shrink the catalogue, define the attributes, extract without guessing, review. This article shows what changes at scale, when full review gives way to control built into the process.

What changes between a hundred and ten thousand product pages

Three things change, and none depends on the choice of AI model.

• Review becomes sampling. You no longer review everything: you review what the checks flag, plus a random draw from each batch.
• One error becomes a series. A prompt that nudges towards a little “embellishment” does not distort one page, it distorts them all the same way.
• Cost becomes a topic. On a few pages, the price of AI calls is negligible; on a whole catalogue, rewritten several times, it needs managing.

The framework does not change. An invented feature remains false information about a product, and Google’s spam policies still target pages produced at scale without value for visitors, defining scaled content abuse as generating many pages mainly to manipulate rankings, “no matter how it’s created”.

The process, from prompt to publication

Sort first, because every enriched page costs money and must be reviewed: sales and Google Search Console tell you which references deserve the work. Test next on a few dozen pages chosen for their diversity (different families, long and short supplier texts, hard cases): that is where the prompt gets fixed, not on the whole catalogue.

Batch processing has two advantages. An error shows up on the first batch rather than on ten thousand pages. And processing interrupted by an outage or a provider limit resumes where it stopped instead of starting from scratch.

The checks a machine does better than a reviewer

A reviewer tires by the three-hundredth page. A rule does not. Here are the checks that are programmed once and run on every batch:

• Every value has its source. An extracted size, material or composition must appear in the original text; if not, the page goes to review.
• Forbidden words. “Eco-friendly”, “hypoallergenic”, “handmade”, “certified”: the list of claims that need proof, and that AI may not add on its own.
• Written forms. One unit, one form: “30 cm” everywhere, never “30cm” or “thirty centimetres”.
• Differences from the previous version. A page whose price, size or category changes during a text enrichment signals an error, not an improvement.
• Volume guards. A batch that comes back empty, or deletes more pages than it edits, blocks publication.

The last point is not theoretical. On Palmora Property, that is the scenario these rules block: without them, an empty response from the interface that feeds the site would publish a site with zero products, with a job shown as successful. A volume check is written once; its absence can cost a catalogue.

On the online plush toy store we are putting back on its feet, the coverage report showed exactly that: across 653 products reviewed and 26 fields defined, size was readable for only 41% of them. The rest is not invented. It is measured, or asked for.

What the human keeps

Review does not disappear, it concentrates where it is useful:

• The prompt and the sample, reviewed before any mass processing.
• Flagged pages, with the reason written down.
• A random draw from each batch, to catch what the rules did not anticipate.
• Statements that commit the store: regulatory, safety, age, warranty, never left to AI alone.

This split is what makes review sustainable: reviewers’ time goes to what matters instead of being spread across the whole catalogue.

Cost, managed rather than endured

On a large catalogue, three levers make the difference in cost, whichever provider you choose.

• Asynchronous batch processing, which the main providers offer at half price in exchange for a longer processing time: OpenAI, Anthropic, Google and Mistral all list it at 50% of the standard price. A catalogue enrichment does not need an answer within the second.
• Never pay twice for the same work. An unchanged page does not need to go back through AI: only new pages and those that changed go out again.
• The right model for the task. Extracting a size from a text does not call for the most expensive model; the choice is made by testing several models on your own sample, as our guide to evaluating an AI model before production explains.

We apply the second lever to our own translation pipeline: on Palmora Property, translation cost was divided by 60 by switching models, and 95% of segments now come out of a translation memory without calling the AI again. The full reasoning, from invoice to tokens, is in our article on model routing.

Publishing reversibly

A mass operation is prepared as if it had to be undone. On the plush toy store, every bulk change follows the same rule: dry run by default, backup before writing, rollback in one command, archiving rather than deletion. That is how 172 product unpublications were carried out, along with operations on 8 pages and 62 translations. One more rule was added: set up the redirect before removing a product page, so that no address that received visits answers into the void.

What an enriched catalogue makes possible

A catalogue with complete, verified fields serves the rest of the store. Site search finally finds “rabbit 30 cm”, filters rest on filled attributes, and the product page that converts answers the buyer’s questions. The feed sent to Google improves, as our Google Shopping and marketplaces guide and our article on what makes a Google Shopping campaign rise detail; the same data feeds marketplaces (see selling on Amazon alongside your store) and inventory synced across channels.

Accurate features also reduce avoidable returns (see the hidden cost of returns) and feed repeat-purchase emails (see welcome, post-purchase and repeat-purchase emails) as well as abandoned cart reminders. For export, they are translated into each language, a project described in our article on the multilingual, multicurrency store. Customer reviews point out what product pages say badly, and richer pages must not weigh the store down (see store speed). Before adding an enrichment app, run the calculation in our article on the cost of Shopify apps; the platform also decides where these attributes live, as our Shopify or Medusa comparison shows. The effects are tracked through the five numbers to track every week, and what payment really costs is checked with our article on real payment fees. The other decisions about a store are gathered in our ecommerce section.

If your catalogue is too large to review by hand, our ecommerce support starts with sorting and a sample: which product pages deserve the work, which fields are missing, and what a batch enrichment would cost, checks included.

Frequently asked questions

Sources

1. Google Search Central, spam policies for Google web search (new tab)developers.google.com
2. OpenAI, batch processing (new tab)developers.openai.com
3. Anthropic, batch processing (new tab)platform.claude.com
4. Google, Gemini API batch mode (new tab)ai.google.dev
5. Mistral AI, batch processing (new tab)docs.mistral.ai