How to use AI to enrich 10,000 product descriptions overnight

10,000 products. Each one needs a better description — factual, specification-rich, suitable for AI commerce channels. Manually: hundreds of hours and a large copywriting bill. With Trajekt's AI Rules and Claude: approximately 3 hours of setup, overnight processing, at roughly £10 for the initial run.

The caching advantage

Trajekt's AI Rules cache results against a hash of the input product data. A 10,000-product catalogue costs roughly £10 to enrich initially (at £0.001/product). After that, only changed products consume credits — typically 5–15% of a catalogue per day. Ongoing AI costs are therefore low.

Build the right prompt

A prompt that consistently works well:

Using ONLY the information provided in the product fields below, write a 100–200 word product description in plain English. State what the product is in the first sentence. List key specifications and measurements. Describe materials and construction where relevant. Mention use cases and who the product suits. No marketing language or superlatives. UK spelling throughout. Output only the description text with no introduction or formatting.

Critical constraints: "using ONLY the information provided" prevents hallucination. "No marketing language" prevents AI defaulting to promotional copy.

What a good enrichment prompt looks like

The difference between an AI description that passes Google Shopping review and one that doesn't is almost entirely in the prompt. Google Shopping rejects promotional language ("best", "unbeatable", "must-have"), HTML markup, and descriptions that are clearly generic. A prompt that generates this kind of output is more harmful than no enrichment at all — you're multiplying bad descriptions across your whole catalogue.

A production-ready enrichment prompt for Google Shopping should specify: write in plain text (no HTML, no markdown), 150–300 words, factual and specification-focused, include material, dimensions, or technical specs if present in the input data, no promotional language, describe what the product IS rather than how good it is. The prompt should also instruct the model to use only information present in the input — not to invent specifications that aren't there. Hallucinated specs in a Shopping description cause Merchant Center price and content mismatch flags.

Different prompts for different channels

Google Shopping and ChatGPT Shopping have meaningfully different description requirements. Google optimises for keyword matching and policy compliance — descriptions should include relevant search terms naturally. ChatGPT Shopping optimises for conversational relevance — descriptions should answer the questions a buyer would ask, in plain prose, without keyword stuffing.

In Trajekt, you configure separate enrichment rules per channel output. Your Google Shopping description rule uses a keyword-oriented prompt; your ChatGPT description rule uses a conversational-oriented prompt. Both generate from the same source product data but produce channel-optimised outputs. The caching layer means each product is only processed once per channel per change — if neither the product data nor the prompt changes, the cached result is reused.

What to do when enrichment produces wrong output

The most common failure mode is the model generating a description that's technically fine but factually wrong — it invented a feature, got the colour wrong, or described the wrong product variant. The fix is usually in the input data quality rather than the prompt. If your source description is "Red jacket" with no other context, even a well-prompted model has nothing to work with. Enrichment amplifies what's there — it doesn't substitute for it. Products with genuinely sparse source data need manual description writing before AI enrichment adds value.

Validate before full rollout

Before processing your full catalogue, process a 100-product test set. Review across different categories: Are specifications captured correctly? Is the AI staying within the source data? Is UK spelling consistent? Adjust your prompt based on what you see. One iteration typically resolves most issues.

Handle edge cases

• Thin source data: If you only have title + colour + size, add a rule: if enriched description word count < 50, fall back to original description
• HTML-only descriptions with no other data: Strip HTML first. If the result is empty, flag for human review
• Bundle products: Use a specific prompt variant: "This is a bundle containing the following items. Write a description that clearly lists what is included..."