Titles and attributes at scale with AI: generate, validate and measure without breaking the feed

Rewriting 3,000 titles from a template is doable with rules; rewriting 3,000 titles well —with the attribute each product needs and the vocabulary each category uses— is a job for a language model. But a model that writes titles with no limits makes things up: attributes the product does not have, promotions, adjectives. This lesson is the method for using AI on the feed with a safety net.

What AI can do well in a feed

• Titles: assemble, for each product, the title from its category's template with the correct attributes and in the order that market uses.
• Structured attributes: pull material, color, pattern, product_detail (capacity, power, dimensions) and product_highlight out of the description or the spec sheet.
• Descriptions: rewrite the manufacturer's copy in your own words, with synonyms and use cases.
• Classification: map products to google_product_category and product_type when the platform does not.
• Translation with market context (lesson 1).

What it cannot do: decide prices, availability, GTIN, brand. Those are facts, not text.

The context it needs

Per product: the feed's current attributes, the description, the spec sheet if there is one, the category, the category's title template and —the piece that makes the difference— the search terms that converted for products in that category (the real vocabulary). Plus the rules (below) as explicit instructions.

The validation rules

Every AI output goes through a validator before it reaches the feed:

Whatever is rejected goes back with the reason, or keeps the previous title. Typical rejection rate at the start: 10-20%; it falls once the prompt is tuned.

The batch workflow

1. Pilot category (a few hundred products, enough data volume).
2. Generate in batches (dozens per call, with the category's context).
3. Validate automatically.
4. Human review by sampling: 30-50 random titles per category; if more than 5% have problems the validator missed, adjust the prompt or the rules and repeat.
5. Deliver as a supplemental feed (id + title + new attributes): the original is never touched.
6. Test per category (split A/B by label, 3-4 weeks; module 6 of Intermediate).
7. Extend to the next category; keep a log of title versions per product.

Cost: pennies per product with today's models; a category of 500 products costs less than an hour of a copywriter's time and is reviewed in half an hour.

Measuring the effect

Per category and title version: impressions, CTR, conversions, ROAS before/after and against the control group. Signs of success: more impressions (more searches covered) with equal or better CTR; warning signs: impressions going up and CTR falling (a title promising something the listing is not) or new irrelevant search terms (an attribute wrongly applied).

Governance: who approves what

• Templates and rules: marketing, with the validator as the law.
• AI output: never straight into the feed; always validator + sampling + a reversible supplemental feed.
• Versions: every title with a date and a version; batch rollback if the test fails.

💡 Ninja trick: Shopping Ninja generates AI titles with exactly this context (attributes, template, terms that converted) and this validator, and delivers them as a supplemental sheet with an approval checkbox per product. The AI writes; the validator watches; you approve; the original feed does not change. It is the same "proposal with memory + human veto" pattern as RSA Optimizer, applied to the catalogue.

What you should remember

• AI is for titles, attributes, descriptions, classification and translation; never for facts (price, stock, GTIN).
• Context = attributes + template + terms that converted.
• A validator is compulsory: length, existing attributes only, no promos or adjectives, formatting, brand/model, no repetition.
• Batches by category, human sampling, a reversible supplemental feed, an A/B test, a version log.