AI Recommended Practice

This document provides guidance on using ProductGen AI effectively, including when to use it, how to prepare data, and how to avoid common pitfalls. The recommendations reflect patterns observed across customer implementations and are intended as starting points rather than rigid rules.

When to Use ProductGen AI

Ideal Scenarios

ProductGen AI delivers the most consistent value in scenarios that are high-volume, repetitive, or structurally well suited to automation. The table below lists the scenarios where customers most commonly have strong returns.

Scenario Why ProductGen AI Helps
High-volume content creation

Generate descriptions for hundreds or thousands of products faster than manual authoring
Consistent catalog enrichment Maintain uniform tone, style, and quality across product content
Multi-language expansion Translate content to new markets without a proportional increase in localization resources
Image-to-data automation Extract structured data from unstructured image content
Data quality gates Automated consistency checking before publication

When Manual Approaches May Be Better

AI is not always the right tool. Recognizing situations in which a human-led approach works better up front avoids a great deal of rework later:

  • Highly regulated content — Legal disclaimers, safety warnings, or claims that require human review.

  • Brand-critical messaging — Flagship product positioning that requires marketing team involvement.

  • Complex technical specifications — Content requiring domain expertise verification.

  • Low-volume, high-value products — Where individual attention provides more value than automation.

Data Quality Considerations

Input Quality Affects Output Quality

ProductGen AI generates content from the data you provide, which means input quality has a direct effect on output quality. The most effective thing a team can do before launching AI-driven enrichment is invest some time focused on source-data hygiene. Also, before using ProductGen AI, users should:

  • Ensure core attributes are populated — Product name, category, and key features should be complete.

  • Standardize attribute values — Consistent formatting (for example, “10 kg” rather than a mix of “10 kg”, “10kg”, and “ten kilograms”) improves results.

  • Remove outdated or incorrect data — AI will faithfully incorporate errors in source data.

  • Verify image quality — For image analysis, ensure images are clear and representative.

Recommended Data Preparation

Different kinds of data benefit from different preparation. The table below summarizes recommended steps by data type, useful as a checklist during a pre-launch data-quality sprint.

Data Type Preparation Steps
Text attributes

Trim whitespace, standardize casing, remove unnecessary special characters
Numeric values Ensure consistent units, remove outliers or placeholders
Images Use high-resolution product images, ensure good lighting, remove watermarks where possible
Categories Verify products are correctly classified before generating category-specific content

Prompt and Input Guidance

Writing Effective Prompts

A small amount of upfront investment in prompt clarity produces large improvements in output consistency. The following principles apply across a wide range of use cases:

  • Be specific about the output format — Specify length, style, and structure.

  • Provide context — Include target audience, channel requirements, or brand voice guidelines.

  • Use examples when helpful — Show the AI what good output looks like; this is often more effective than describing it in the abstract.

  • Test iteratively — Try prompts on sample products before bulk application.

Example Prompt Structure

The following template illustrates the kind of structure that generally produces dependable outputs. Placeholders in braces would be replaced with attribute values at runtime:

Generate a product description for an e-commerce channel.

Requirements:

  • Length: 100-150 words

  • Tone: Professional, benefit-focused

  • Include: Key features, target use case, differentiators

  • Avoid: Superlatives without support, competitor mentions

Product attributes:

  • {product_name}

  • {brand}

  • {category}

  • {features}

  • {specifications}

Input Mapping Best Practices

When deciding which attributes to send to the model, restraint is usually a virtue. Sending only relevant information tends to produce cleaner output than a kitchen-sink approach:

Map only relevant attributes — Including unnecessary data can dilute output quality.

Order inputs logically — Place the most important attributes first.

Use descriptive labels — Help the AI understand what each input represents.

Common Pitfalls

The following pitfalls account for most disappointing early results. Watching for them from the start is the most efficient way to keep an implementation on track.

Pitfall 1: Expecting Perfect Output

Problem: Treating AI-generated content as final without review.

Solution: Implement a review step for generated content, especially for customer-facing material. Use AI to create drafts that humans refine.

Pitfall 2: Ignoring Output Variations

Problem: AI may produce slightly different outputs for similar inputs.

Solution: For consistency-critical content, generate once and review rather than regenerating.

Pitfall 3: Overloading Prompts

Problem: Trying to accomplish too much in a single prompt.

Solution: Break complex requirements into multiple prompts. Generate title and description separately rather than everything at once.

Pitfall 4: Not Testing on Representative Data

Problem: Prompts work on test products but fail on edge cases.

Solution: Test prompts on a representative sample including products with minimal data, extensive data, different categories, and different languages where applicable.

Pitfall 5: Ignoring Usage Costs

Problem: Running AI on all products when only some need it.

Solution: Use filters and conditions to target AI execution. Prioritize products that need content over those that already have quality data.

Governance Recommendations

Review Workflows

Consider implementing review gates for AI-generated content. Options, which can be combined:

  • Automated quality checks — Basic validation such as length and required elements.

  • Sample review — Randomly review a percentage of generated content to catch systemic issues early.

  • Exception handling — Flag unusual outputs for human review.

Version Control

When modifying prompts, document the change and rationale, test before deploying to production, and consider a phased rollout for significant changes.

Performance Optimization

Batch Processing

For large-scale operations, a batch-oriented approach almost always outperforms one-at-a-time execution. Process products in batches rather than individually, schedule intensive operations during off-peak hours, and monitor usage against your credit allocation so that a runaway job is noticed early rather than at month-end.

Targeting the Right Products

Conditions let you apply AI selectively, which keeps both cost and noise under control. Useful targeting patterns include:

  • Products missing descriptions.

  • Products in specific categories.

  • Products entering specific workflow states.

  • Recently updated products.