AI for product descriptions in catalogues of thousands of items, without junk content

A store with three hundred products can afford to have someone write every page by hand. One with eight thousand cannot. What you usually find in large catalogues is a mix of supplier descriptions, empty pages and a handful written with real care. Artificial intelligence changes that equation: what used to take months of writing now takes days. The problem is that it can also produce eight thousand texts that say the same thing in different words, and that neither sells nor ranks.

This article is not about whether AI writes well, but about how to organise the work so the result is useful when the catalogue is large.

The problem is not the AI, it is what you give it

A language model writes from whatever it receives. If you only give it the product name, it will fill the gaps with generalities: “modern design”, “high quality”, “ideal for everyday use”. These are phrases that could go on any product page, which is why they offer nothing to the customer or to a search engine.

The quality of the text depends almost entirely on the input data:

• Real features: dimensions, materials, compatibility, composition, power, capacity.
• Context of use: who it is for, in what situation it is used, what it goes with.
• Differences within the product family: what sets this model apart from the one next to it.
• Questions customers already ask you by phone, email or chat.
• Restrictions: what must not be said (health claims, promised results, comparisons with brands).

Before generating anything, look at what you have. If the features are empty, the first task is to fill them in, not to write. A correctly entered EAN lets you query public product databases; a supplier’s specification table can be turned into PrestaShop features. With that, the AI has something concrete to say.

Prioritise: not every product page deserves the same effort

With thousands of products, treating everything the same is the quickest way to spend time and API money for nothing. Split the catalogue into three groups:

1. Products that sell or get visits. They are usually a small part of the catalogue. Here a full description is worth it, reviewed by someone who knows the product, with FAQs and clear benefits.
2. Products that get searches but do not convert. They appear in Google with impressions and few clicks, or get visits that leave. These gain the most from a better page.
3. The long tail. Spare parts, sizes, very technical references. A correct description generated from good features, with spot-check reviews, is usually enough.

You already have the data to make this split: sales in PrestaShop, visits in your analytics, and impressions and clicks in Search Console.

Templates per family, not one prompt for everything

A single prompt (“write a 200-word description”) applied to the whole catalogue produces pages with the same structure and the same tone, whatever the product. A template per family or category works better:

• For spare parts, what matters is compatibility, the original part number and how to check it is the right part.
• For fashion, the fabric, the cut, how it sizes and how to wash it.
• For food, origin, format, storage and allergens, which also come with their own legal obligations.
• For technical equipment, specifications, recommended use and what is in the box.

Each template should set out the structure, your brand’s tone, the approximate length and, above all, which data to use and what not to make up. A sentence such as “if a detail is not in the information provided, do not mention it” prevents a good share of the errors.

Work in batches, with drafts

Generating thousands of texts and publishing them straight away is a recipe for disaster. The workflow that works is this:

1. Test batch of twenty or thirty products from the same family.
2. Full review of that batch: correct data, tone, repetition, empty phrases. Adjust the template based on what you find.
3. Large batch as drafts: the texts are saved without replacing the current ones yet.
4. Spot-check review, comparing before and after. If a sample shows errors, you fix the template and regenerate the batch, not page by page.
5. Publication, noting the date so you can measure.

Working in the background, with a queue and a scheduled task, avoids locking up the back office and the execution-time cut-offs that appear when processing hundreds of products in a single request. And keeping a history of every change lets you roll back if a batch goes wrong.

What to check in each sample

• Data: do dimensions, materials and compatibility match reality?
• Differentiation: if you read three pages from the same family in a row, can you tell them apart?
• Filler phrases: “exceptional quality”, “you won’t be disappointed”. Out.
• Prohibited or risky claims: health, safety, guaranteed results.
• Consistency with the product page: the text must not contradict the features or the price.
• Languages: if you translate, units, sizes and expressions must be adapted to the country.

What Google penalises is not AI

Google has said publicly that it does not care how content was produced, only whether it is useful. What it does go after is mass-generating pages with the main aim of manipulating results, with no value for the reader. For a store, the difference is practical: an AI-generated page built from real data and reviewed helps the buyer; a thousand pages with the same text shuffled around help nobody, whether a machine or a person wrote them.

Do not expect rewriting descriptions to move traffic overnight either. Measure by product group and over periods of several weeks: impressions, clicks, product page conversion rate and returns for “not what I expected”.

A technical detail: keep the store up to date

Processing the catalogue in bulk touches many tables at once and depends on scheduled tasks, connections to the AI provider’s API and recent PHP versions. On stores running very old PrestaShop or PHP, current tools often cannot be installed, and what you save on writing is lost on patches. Keeping the store updated is part of the project.

How to solve it in PrestaShop

Easy AI Magic Catalog Suite works inside the back office with your own OpenAI, Gemini, Claude, DeepSeek, Mistral or Cohere key. Its AI Mass Processor puts products, categories or CMS pages into a queue that is processed in the background by a scheduled task, and generates the texts as drafts: you compare before and after and only then apply them, with a history to restore from. AI Product Enrichment looks up data by EAN, UPC or ISBN in public databases, AI Features extracts specifications from descriptions and turns them into features, and AI Premium Bullets and AI FAQs complete the page. If the problem is that new products take too long to go live, AI Catalog Flow creates the product from your phone from a photo, with its categories, images and combinations, and applies the default values you set.