How I Use AI to Generate 10,000 Product Descriptions (Without Them Sounding Like AI)
Two years ago, a client called me. He ran an outdoor shop with 8,000 products – and 6,000 of them had exactly the description the manufacturer had supplied.
“Tobias, Google is ignoring half my products. And the copy we do have is word-for-word the same as my three biggest competitors.”
He was right. Duplicate content, zero added value, zero rankings. The solution was clear: every product needs a unique, SEO-optimized description. The problem: having 6,000 texts written manually? At €15–20 per text, we’re talking €90,000–120,000. The budget wasn’t there.
So I started building a solution. Today, 60,000+ generated product descriptions later, I can tell you exactly what works, what doesn’t – and why most people use AI copy the wrong way.
The Problem: Why Product Descriptions Are So Painful
Let me briefly explain why this is such a big deal.
If you run an online shop with 500+ products, you almost certainly have one of these problems:
- Manufacturer copy everywhere: You’re using the same descriptions as every other retailer. Google recognizes this as duplicate content and rewards nobody for it.
- No descriptions at all: Some products have only a title and an image. For Google, they barely exist.
- Descriptions that don’t sell: Technical spec sheets instead of copy that explains to a customer why they need this exact product.
Manual writing doesn’t scale. Even with a good copywriter, you’ll manage maybe 20–30 texts a day. With 5,000 products, you’re busy for months – and by the time you’re done, the catalog has already changed.
”Just Let ChatGPT Write the Copy”
I hear this constantly. And I understand the logic: take ChatGPT, give it the product name, press Enter. Done.
Here’s what you get:
Manufacturer text: “Trekking backpack, 45L, Nylon 210D, waterproof, back ventilation, adjustable hip belt, weight 1.8 kg, dimensions 65x32x22 cm”
Naive ChatGPT prompt (“Write a product description for this backpack”): “Discover the perfect companion for your next outdoor adventure! This high-quality trekking backpack offers a generous 45 liters of storage space for all your gear. Thanks to the robust Nylon 210D material, it withstands wind and weather…”
You spot it immediately, right? That AI tone. “Discover the perfect companion” – that’s on thousands of shops by now. It’s the equivalent of manufacturer copy, just with more adjectives.
The problem isn’t the AI. The problem is what you feed it – and what you do with the output.
What Actually Works: The Three Pillars of Good AI Copy
After 60,000+ product descriptions, I’ve learned that the difference between “sounds like AI” and “sounds like a good copywriter” comes down to three things.
Pillar 1: Context Is Everything
The secret isn’t the prompt – it’s the data you give the AI.
A bare product name isn’t enough. But when you feed the AI the following, something completely different happens:
- Product specifications – Technical data, material, dimensions, weight
- Category context – Which category is the product in? Who’s the target audience?
- Shop context – What kind of shop is this? Premium, discount, specialized?
- Existing content – What did the manufacturer supply? What’s currently on the page?
The more context, the better the result. The difference is like between “Write a text about a backpack” and “Write a text for a 45L trekking backpack in a premium outdoor shop targeting experienced hikers looking for lightweight gear for multi-day tours.”
Pillar 2: Prompt Engineering Is Real Work
I spent months refining prompts. Not days – months.
A good prompt for product descriptions defines:
- Tone and voice – Does the shop address customers formally or informally? Is the tone factual or emotional?
- Structure – Should the text lead with a benefit statement? Should technical details come at the end?
- SEO signals – Which keywords should be woven in naturally? How long should the text be?
- What to avoid – No superlatives without evidence. No generic phrases like “high-quality” or “perfect companion.” No made-up features.
That last point is critical. The biggest danger with AI copy isn’t that it sounds bad – it’s that it makes things up. An AI model happily “hallucinates” features the product doesn’t have. If you don’t control for that, you have a legal problem.
Pillar 3: Post-Processing Makes the Difference
The raw AI text is never the final product. Between generation and publication, a lot still happens:
- HTML structuring – Clean semantic structure with correct headings (h2–h6), lists, and paragraphs. Google loves this.
- SEO optimization – Keyword placement, meta descriptions, internal linking opportunities.
- Quality checks – Do the mentioned features match the actual product data? Are there contradictions?
- Deduplication – If you’re describing 500 backpacks, you can’t write “the perfect companion” 500 times. The texts need to be unique within your shop too.
Before and After
Here’s a realistic example. Same backpack, but this time with the full pipeline approach:
Manufacturer text: “Trekking backpack, 45L, Nylon 210D, waterproof, back ventilation, adjustable hip belt, weight 1.8 kg, dimensions 65x32x22 cm”
After the full pipeline:
Trekking Backpack 45L – Light Enough for Long Tours, Big Enough for Everything That Matters
1.8 kg on your back – you carry the rest in the 45 liters of storage. This trekking backpack is designed for multi-day tours where every gram counts.
The Nylon 210D handles rain without going overboard on weight. The back ventilation means you won’t be stuck to your pack on day three. And the adjustable hip belt shifts the load where it belongs – onto your hips instead of your shoulders.
Great for: Multi-day tours, hut-to-hut hikes, long-distance trails
Notice the difference? Same facts, but:
- Concrete benefits instead of buzzwords
- Speaks directly to the hiker
- No made-up features
- Clear structure that Google understands
The Numbers: What Happens When You Swap 6,000 Texts at Once
Back to my outdoor client. We generated and swapped his 6,000 product descriptions over several weeks. Here’s what happened:
| Timeframe | Organic Visibility | Indexed Pages | Conversion Rate |
|---|---|---|---|
| Before the swap | Baseline | ~2,800 of 8,000 | 2.1% |
| After 4 weeks | +120% | ~5,400 of 8,000 | 2.3% |
| After 3 months | +500% | ~7,200 of 8,000 | 2.4% |
The +500% sounds dramatic, but the reason is simple: Google suddenly found 4,400 pages that had been practically invisible before – because they were either duplicate content or had too little content.
The conversion rate rose by about 15% – from 2.1% to 2.4%. Not because the copy was magic, but because customers finally understood what they were buying.
The Most Common Mistakes I See
When shop owners try AI copy on their own, they usually fail at one of these points:
1. No Context, No Results
You give ChatGPT a product name and expect good copy. That works about as well as emailing a copywriter just the product name and saying: “Write something.”
2. Copy-Paste Without Checking
The AI text goes straight into the shop – without verifying whether the mentioned features are accurate. I’ve seen shops where AI-generated copy described features the product didn’t even have. That’s not just embarrassing, it’s legally actionable.
3. Every Text Sounds the Same
If you use the same prompt for 1,000 products, 1,000 texts sound the same. Variance in tone, structure, and opening is crucial – otherwise you’re just replacing one type of duplicate content with another.
4. Generate Once, Never Touch Again
Your catalog changes. Prices change. Seasonal relevance changes. AI copy isn’t “fire and forget” – you need a process to keep it current.
Who Is This For?
Not everyone. Here’s my honest assessment:
Yes, if:
- You have 500+ products with generic or missing descriptions
- Your organic visibility is stagnating or declining
- You’re using manufacturer copy that your competitors also have
- You’re in a market where product content can be a differentiator
No, if:
- You have 20 products – a good copywriter is the better investment
- Your products are so specialized that only a domain expert can write about them (e.g., medical devices)
- You don’t have a strategy for keeping the texts up to date
What I Built from This
I poured the experience from this and many subsequent projects into a platform: mitKai. Over 60,000 product descriptions for dozens of shops later, it takes care of exactly the work I’ve described here – from context building to generation to SEO-optimized HTML output.
But whether you use mitKai or build your own approach – the three pillars remain the same: good context, thoughtful prompts, clean post-processing.
If you have questions or want to know whether AI copy makes sense for your shop – reach out. I’m honest, even when the answer is “no.”