Generating 10,000 AI product descriptions is the easy part
Generating product copy with AI is easy. The work starts after: checking 10,000 descriptions for accuracy, holding one voice, and keeping every channel on the same version.
On this page
Picture an AI content run. About 10,000 product descriptions in English, German and French, 3 locales, for an electronics catalog that runs from simple accessories to networking equipment with dozens of technical specifications. Add titles and bullet points and that is tens of thousands of pieces of content, which today’s models write quickly and cheaply.
That is not the interesting part. Any team with an API key can generate 10,000 descriptions in a day or so. The hard part is finding the invented specification hidden somewhere across 10,000 products, holding one brand voice from product #1 to product #10,000, and keeping the same facts straight across 3 locales, 5 channels and 2 compliance frameworks. Then making all of it repeatable, as a system rather than a one-time project.
Generation is the solved half.
The AI content conversation is still mostly about generation. “How do we write descriptions more quickly?” “Which model produces better copy?” “How do we prompt for better output?” Those questions are answered. Any modern LLM can write a plausible product description. What is left is operational.
Generation takes hours. Designing and running the review, the validation and the coordination around it can take weeks.
What goes wrong at 10,000 products.
At 50-200 products, AI generation feels magical. One person can read every output, catch the errors and fix them. At 10,000 products, the same failures repeat across the catalog and nobody meets them one at a time:
- Invented specifications. The model fills gaps in the source data with attributes that sound right. A networking switch with no documented port count comes back as “8-port,” because that is the likely answer in the training data. Nobody checks every technical claim by hand across 10,000 products.
- Voice drift. Brand voice holds less well as the batch grows. Early batches sound distinctive. Later batches sound like anyone. The model aims for plausible, and plausible trends toward the mean.
- Uneven quality by language. Expect problems such as English loanwords for technical terms, inconsistent formality, or European and Canadian French conventions mixed into the same description.
- SEO cannibalization. 500 products in one category come back with near-identical descriptions and the same keyword clusters. Instead of 500 pages that can rank, you have 500 pages competing with each other.
- Formatting the marketplace rejects. Copy over the character limit, HTML that isn’t allowed, a missing bullet point count or keyword placement. You find out when the listing is rejected.
- Claims with nothing behind them. The model writes “hypoallergenic,” “medical-grade” and “energy-efficient” without checking a certification. At 10,000 products, some of those reach the page, and they carry legal liability.
- Content that contradicts the taxonomy. The description says “outdoor use” and the product is classified as “indoor electronics.” The same happens with features and use cases that belong to another category.
- Near-duplicate variants. Color variants of one product get descriptions that barely differ. The variant gains nothing, and search engines read the set as thin content.
- Nothing links the copy to its source. The attributes change next month, from a supplier update or a spec correction, and the description written from them is now stale. No record ties one to the other.
- Channels drift apart. Amazon has the V1 description. Shopify has V2, after someone edited it. Google Shopping has V1 again, because the feed was cached. Three channels, three versions of the truth. The AI wrote every one of them, and nothing kept a record of which one is current.
Note
The amplification problem
AI amplifies whatever quality is already in your product data. Products with rich, structured attributes tend to get more of their drafts accepted; thin records tend to see more rejected or rewritten. Good product data produces AI content you can use. Thin product data produces confidently wrong content, at a volume no team can review by hand. And because generation runs that quickly, you can build a massive review backlog in a day.
Product context often matters more than the model.
Among today’s capable models, which one you pick usually matters less than the product context you hand it.
What decides the quality of AI content:
Scroll sideways to see the whole table.
| What you hand the model | What it buys in the output | What goes wrong without it |
|---|---|---|
| Typed, structured attributes rather than free text | The specifications are right. The features exist. The claims have something behind them. | The model fills the gaps with whatever is statistically likely, and invented specs repeat across the catalog. |
| A settled taxonomy and a category per product | The copy uses the terms that category uses, and the use cases suit it. | A B2B industrial product gets written up in consumer lifestyle language, with claims that belong to another category. |
| What each channel requires | The copy fits the format, the length and the required elements of each channel. | Descriptions run over the limit, miss a required element or use formatting that is not allowed, and the marketplace rejects the listing. |
| Examples of the brand voice | Tone, register and style hold across the catalog. | The output slides toward generic, and what made the brand sound like itself fades over a long run. |
| The conventions of each locale | Terminology, formality and local convention are respected. | The German copy takes a casual register for professional products. The French mixes European and Canadian conventions. |
| What the rest of the category already says | The copy says something of its own instead of repeating the language every competitor uses. | Every product in the category sounds the same, and nothing on the page says why to pick this one. |
| The compliance limits | Claims stay inside the certifications you can show, and regulatory terms are used the way the regulation uses them. | The model writes “waterproof” for IP44 products, which are splash-resistant and not waterproof, and the claim carries legal liability. |
Completing the attributes, settling the taxonomy and fixing the context you send often does more than switching models. The product record matters more than the model.
What you have to build around the run.
AI content at this volume needs checks and records that many teams haven’t built. This is a build list; check how much of it any tool you evaluate covers.
1. Check every draft.
Check each description before it enters the catalog. Compare its specifications against the source attributes, and flag any claim that isn’t in them. Then check the character limit for each channel and the required structural elements. Check the language of each locale, how closely the voice follows the brand, and every compliance term against the certifications you hold.
2. Suggestions wait for a person.
Make AI suggestions wait for review by default. Let the checks flag problems, have a person review the suggestions, and check products against each channel’s rules before sending.
In CataZenta, agents draft descriptions, SEO titles and translations from the product record, and by default each draft waits in one approval queue, where your team can edit it, or approve and reject drafts in bulk. Product edits keep a version history, and Amazon and Shopify pushes are checked against channel rules. There isn’t a brand voice setting, acceptance tracking or a stale-copy flag; you write your tone into your own agent in Agent Studio.
3. Watch the acceptance rate.
Track the acceptance rate per product family, per locale and per channel, batch after batch. Set a threshold for each; pause generation when a rate falls below it, adjust the prompts and run the batch again. That way the drop is caught and corrected before the content is published.
4. Keep the lineage.
Mark every AI-written description as AI-written. Record which model and prompt template wrote it and which product attributes it read, when it ran, who approved it, and which version of the source data it was based on. When the source data changes, flag the descriptions that depend on it as possibly stale.
5. One record, different copy per channel.
One product needs different copy per channel, and the facts in each version have to match. Generate each channel version from the same source record. When a specification changes, regenerate the channel descriptions built on it and send them back through review.
Generation is cheap. The rest is not.
Here is what each approach costs, line by line:
Scroll sideways to see the whole table.
| What it costs | Generation on its own | Generation plus the work around it |
|---|---|---|
| Writing the content | Modest compute cost | The same modest compute cost |
| Automated checks | Nothing, because there is no pipeline | Building the pipeline, plus the compute to run it |
| Human review | None (ship and pray), or reading all of it by hand | A sample, weighted toward what the checks are least sure about |
| The checking and approval machinery | Nothing, because it was never built | A one-time pipeline build, or a platform that covers part of it (ask which parts) |
| Fixing the errors | Unknown until a marketplace rejects a listing or a buyer complains, and often more than preventing them would have cost | Should be lower, because more errors are caught before publishing |
| Time to publishable | Hours to generate, plus an unknown stretch of weeks or months finding and correcting errors | Hours to generate, plus a planned review and approval window |
| Doing it a second time | Nothing carries over: each batch is a new manual project | The pipeline carries over and should improve with each batch. Batch #10 should take less work than batch #1. |
Building the pipeline costs more upfront than “generate and ship.” The total cost of ownership can come out lower. Count error correction, marketplace rejections, returns from incorrect specifications, and the operational cost of manual review with no pipeline behind it. And you build it once. Batch #1 pays for the pipeline, and every batch after that uses it.
Tip
The test
For any description on the page, can you say it was checked, approved and still matches the record? If not, the content is a liability that grows with the catalog.