Ad copy generation engine
It's not that you can't write it—it's that you don't know which line gets the click
Enter a product title and description. The system extracts the key facts and produces 5 ad-copy variants. Calibrated from 3,000 sets of actual click performance. API integration. About NT$1.6 per copy.
The bottleneck in copy isn't production—it's judgment
Three ad lines for 5,000 SKUs is 15,000 lines. People cannot finish that. Nobody disputes this.
The dispute is what comes next: once it's written, how do you know which line is worth the spend?
Most people now generate with generic AI. It can write, and it writes reasonably well. The problem is—
It has never seen delivery results for your category
It does not know whether “limited time” or “limited quantity” works on your audience
The generation logic is different every time, so you can't accumulate
Output length is uncontrolled; pasting into the ad console often exceeds character limits
The value of a copy tool isn't that it can write. It's knowing what works, and using the same standard every time.
Three steps
- 1
Extract, don't slap on labels
Parse the product title and description to extract category, specs, selling points, and usable differentiation. This step decides whether the copy hits the point—if the facts are wrong, polished writing doesn't help.
- 2
Generate variants from patterns that worked
Generation logic is trained and calibrated on 3,000 sets of actual click performance from ad copy, not inferred from a general language model. Output length is already aligned to ad-platform character limits; no extra trimming.
- 3
Five variants per call
A single API call returns 5 copy variants, ready for delivery or creative production.
Specifications
| Item | Specification |
|---|---|
| Input | Product title and description |
| Output | 5 ad-copy variants per call |
| Language | Traditional Chinese only |
| Length control | Already aligned to ad-platform character limits |
| Tone control | Not supported |
| Generation basis | Actual click performance from 3,000 sets of ad copy |
| Data use | Input is used for that generation only; it is not used to train the model |
| API | Token authentication |
| Billing | Billed by actual API calls, NT$8 / 5 copies |
Why not just use generic AI
| Item | Generic AI copy tools | AdCopy Engine |
|---|---|---|
| Judgment basis | General language knowledge | 3,000 sets of actual click performance |
| Consistency | Different logic each generation | The same standard every time |
| Length control | Uncontrolled; often exceeds platform limits | Already aligned to ad-platform character limits |
| Specifiable tone | Yes | Not supported |
| System integration | You handle it yourself | API, Token authentication |
| Billing | By token or subscription | By actual API calls, NT$8 / 5 copies |
The difference isn't how well it writes. It's the basis for judgment. Generic models infer from language knowledge. This system generates from results that were actually clicked.
A good fit
- Ecommerce and retail with many SKUs that need large volumes of copy variants
- Teams that need copy production embedded in listing or delivery workflows
- Already running ads but unsure which wording works
- Platforms that need merchants to generate copy themselves
Not a fit
- Need languages other than Traditional Chinese → not currently supported
- Need to specify tone, style, or particular selling points → specification not supported
- Need brand-campaign copy or long-form content → this system is for ad-copy variants
- Product data is not yet structured → the data source must be prepared first
Pricing
| Item | Price |
|---|---|
| Copy generation | NT$8 / 5 copies (about NT$1.6 per copy) |
- Billed by actual API calls. No monthly fee, no seat fee, no minimum usage.
- These are current prices, reviewed annually. After contract, the signed price is locked until the term ends; if list prices drop during the term, the lower price applies.
FAQ
How is this different from using ChatGPT directly?
Quality is not the gap—both write reasonably well. The difference is the basis for judgment. This system’s generation logic is trained and calibrated on 3,000 sets of actual click performance from ad copy, uses the same standard every time, and output length is already aligned to ad-platform limits. Generic models generate from general language knowledge, with different logic each time.
What do I input, and what do I get?
Input a product title and description. A single call returns 5 ad-copy variants.
Will the copy exceed ad-platform character limits?
No. Output length is already aligned to ad-platform character limits.
Which languages are supported?
Traditional Chinese only at this time.
Can I specify tone or selling points to emphasize?
No. The system generates from product information and patterns that worked. It does not accept tone or selling-point specification.
Will our product data be used for training?
No. Input product data is used for that generation only and is not used to train the model. We also do not receive your ad-performance data.
How does authentication work?
Token authentication.
How is pricing calculated?
Billed by actual API calls, NT$8 / 5 copies, about NT$1.6 per copy.
Run it on your products
Send us a batch of product titles and descriptions. We'll generate a set of copy for you to see. You'll see immediately whether it captures the point.
Book a system walkthrough