Generative AI can rewrite thousands of product titles in minutes. That saves time but carries the risk of inventing features, repeating phrases and losing precision. Used with clear rules, review and testing, it improves feeds. Without them, it produces errors at scale.
The essentials in a definition
Optimising feed data with generative AI allows rewriting titles and descriptions in bulk, but requires control so as not to invent features. Combine the rules of Merchant Center, guidance on AI-generated content and structured data with a verification process: AI proposes, manufacturer and catalogue data confirm, and human review approves by sampling.
What AI does well in the feed
Normalising formats, standardising the order of information (brand, type, feature, colour, size), fixing spelling errors, translating and adapting short descriptions. In large catalogues, consistency is already a gain.
It also helps produce variations for testing and generate descriptions from existing attributes.
Where it fails
It invents features the product does not have, confuses similar models, swaps units, repeats identical phrases for many products and uses promotional tone in fields where it should not. In regulated categories it may make unacceptable claims.
An error in one rewrite rule can affect thousands of products at once.
Optimising titles with AI without inventing
- Use only real attributes as the source
- Define the format by category
- Review samples before bulk application
- Test on a small group and measure
Rules and protection
Give the AI only data that exists in the attributes. Forbid unsupported claims. Define the title format by category. Take a human sample before applying in bulk. Keep the original title and store versions so you can roll back.
Apply first to a small group and compare performance.
AI rewrites fast; only real data guarantees it writes the truth.
Measure
Compare impressions, clicks, conversion and disapprovals between groups with and without generated titles. Check returns and complaints, which may signal misleading descriptions. Review the results regularly and update the rules.
Always keep the source of truth in the real product data.
Decision table
| Step | Who | Control |
|---|---|---|
| 1. Define title rules | Team | Template with brand, type, attributes |
| 2. Generate proposals | AI | Use only the data supplied |
| 3. Validate | Automatic rules and human | Compare with real attributes |
| 4. Publish and measure | Team | Test on a sample before generalising |
Common mistakes to avoid
- Inventing attributes. AI may add colour, material or compatibility that do not exist.
- Repetitive titles. Differentiation and clarity are lost.
- Replacing everything at once. Test on samples and compare performance.
How to measure the result
Compare impressions, clicks and conversions of products with new and old titles, and count titles rejected in validation.
Limits and exceptions
Models can make subtle errors that are hard to detect at volume, and feed rules have requirements that change. The quality of source data limits the result: if the catalogue is incomplete, AI tends to fill gaps. Keep a record of versions, the ability to revert and human responsibility for what is published.
Illustrative example
Illustrative example: a shop generates new titles for 500 products with a model and a fixed structure rule. An automatic validator compares each title with the catalogue’s attributes and rejects those mentioning something non-existent. It tests on 50 products for four weeks and sees improved impressions. Only then does it apply to the rest of the range.
In practice: checkpoints
Generate titles from verified attributes only, test them on a small group of products and compare against current titles. Keep a human review step for brand-sensitive categories.
Frequently asked questions
Can I generate descriptions with AI alone?
With review and from real data. Without that, there is a risk of error and invention.
How do I avoid bulk errors?
With human sampling, gradual rollout and stored versions.
Does this improve results?
It can, mostly through consistency. Test with control groups.
Can I trust AI to write product titles?
With clear rules, supplied data and automatic and human validation; never without checking.
What error is most common?
Adding features the product does not have, or swapping sizes and models.
Generative AI speeds up feed optimisation, but must use only real catalogue data, with automatic validation and sample review before publishing.
Sources and recommended reading
- Merchant Center: product data specification
- Google Search Central: generative AI content guidance
- Google Search Central: product structured data
Last editorial review: October 2026. Informational content; always confirm current conditions in the official sources before deciding.