Home / Insights / Generative AI and LLMs
Generative AI and LLMs
Generative AI for Ecommerce Product Content: Building a Catalogue Workflow with Brand Control and Approval
Writing and maintaining product content across thousands of SKUs is slow, repetitive work, which makes it a natural fit for generative AI. The hard part is not producing text; it is producing accurate, on-brand text that a merchandiser can approve quickly. This guide explains how to design that workflow.
TopicsGenerative AI and LLMsBusiness Operations and Integration
Key takeaways
- Generate from structured product attributes, not from a model’s general knowledge, so facts come from your data.
- Encode brand voice and banned claims as tested rules and examples, not a single paragraph in a prompt.
- Check every draft automatically for missing attributes, invented specifications, length and prohibited wording before a human sees it.
- Design the approval screen for speed: side-by-side source data, inline editing and bulk approve for low-risk categories.
- Measure editing effort and rejection reasons, and feed them back into prompts and rules.
On this page
Most retailers know the problem well. New ranges arrive with a spreadsheet of attributes from the supplier and very little usable copy. Marketplaces and the brand’s own site need titles, bullet points and descriptions in specific formats, sometimes in several languages. Existing descriptions drift out of date as products change. Content teams spend their days on repetitive writing instead of the hero products and campaigns that need real creativity.
Generative AI can take on much of the first draft. But a naïve approach, pasting a product name into a chat tool, produces confident text that may invent materials, dimensions or certifications the product does not have. In retail that is not a style problem; it creates returns, complaints and, in some categories, regulatory exposure. This article describes how we approach catalogue content as a generative AI development workflow: grounded in your data, constrained by your rules and approved by your team.
The workflow at a glance
- Collect inputs. Product attributes from the PIM or ERP, category, target channel and any approved marketing notes.
- Select a template. Each category and channel has its own structure, length limits and required fields.
- Generate a draft. The model writes using only the supplied attributes plus brand rules and examples.
- Run automated checks. Format, length, required attributes mentioned, no unsupported claims, no banned words.
- Human review. A merchandiser approves, edits or rejects with a reason.
- Publish. Approved content flows back to the PIM or storefront through an integration, with a record of who approved it.
- Learn. Edits and rejection reasons are analysed and used to improve templates and rules.
Each step is a separate, testable component. That matters when the model is changed, a new channel is added or a category needs special handling.
Grounding: making the product data the source of truth
The single most important design decision is that factual content comes from your structured data, not from the model. In practice that means:
- Pass attributes as clearly labelled fields (material, dimensions, care instructions, compatibility) and instruct the model to use only these for factual statements.
- Tell the model explicitly what to do when an attribute is missing: omit the point rather than guess.
- Include approved marketing notes as a separate, clearly marked field, so persuasive language is based on something the brand actually claims.
- Ask for structured output, for example title, bullets and description as separate fields, so each part can be checked and mapped to the right channel field.
Data quality becomes visible quickly. If attributes are incomplete or inconsistent, drafts will be thin or wrong. Many retailers find that the first phase of a content project includes a round of attribute clean-up, which also benefits search and filtering on the storefront. Our article on a data strategy for AI covers this preparation more broadly.
Brand voice and claim rules
Brand voice is often described in a paragraph of adjectives that a model interprets loosely. It works better as concrete, testable material.
| Element | How to encode it |
|---|---|
| Tone | Several approved example descriptions per category, plus a few rejected ones with reasons |
| Vocabulary | Preferred terms and banned terms lists (for example avoid superlatives such as “best” or “perfect” if your brand policy does) |
| Claims | Rules for regulated wording: health, environmental, safety and certification claims allowed only when a specific attribute supports them |
| Structure | Templates per channel with field lengths and bullet counts |
| Localisation | Glossaries for product terms and units per market |
Claim rules deserve particular care in categories such as cosmetics, supplements, children’s products and electrical goods, where wording can carry legal obligations. Here the automated checks should block drafts that mention a claim without the supporting attribute, and those categories should always go to a human reviewer.
Automated checks before human review
Reviewers should spend their time on judgement, not on catching obvious errors. Automated checks run on every draft before it reaches the queue.
- Format and length: title and bullet limits per channel.
- Attribute coverage: required attributes (size, material, compatibility) are mentioned.
- Unsupported facts: numbers, materials or certifications in the text that do not appear in the input attributes are flagged.
- Banned and regulated terms: blocked or flagged according to category rules.
- Duplicate content: drafts that are too similar to other products in the catalogue are flagged, since near-identical descriptions help nobody.
Drafts that fail checks can be regenerated automatically once, then routed to a reviewer with the failure reason shown. This is where generative AI differs from ordinary content tooling: the guardrails are as much part of the product as the generation itself.
Designing the approval interface
The approval screen decides whether the workflow actually saves time. A merchandiser approving dozens of products an hour needs:
- The source attributes side by side with the draft, so facts can be verified at a glance.
- Highlighted check results, such as a flagged claim or a missing attribute.
- Inline editing, with the edited version saved alongside the original draft for learning.
- Quick reject reasons (inaccurate, off-brand, too generic, wrong format) rather than free text only.
- Regenerate with a short instruction (“shorter”, “mention the warranty”).
- Bulk approval for low-risk categories once quality has been demonstrated, with spot checks.
For more on designing review steps that people actually use, see designing human review for generative AI output.
Measuring results and understanding cost drivers
What to measure
| Measure | Why it matters |
|---|---|
| Approval rate without edits | Shows whether drafts are usable as generated |
| Edit distance on approved drafts | Indicates how much reviewer effort remains |
| Rejection reasons by category | Points to which rules or templates need work |
| Time from product creation to live content | The business outcome most teams care about |
| Post-publication corrections | Catches errors the review missed |
Cost drivers
Model usage per draft is usually a small part of the total. The larger effort sits in attribute preparation, PIM and storefront integration, the number of distinct category and channel templates, the review interface, and evaluation. Volume matters for model choice: very high volumes may justify a smaller or self-hosted model for simple categories, with larger models reserved for complex ones. See reducing LLM token and cloud costs for the optimisation levers.
Content generation also connects naturally with other retail AI work; for a wider view, see AI agents in retail.
Frequently asked questions
Will AI-generated product descriptions contain mistakes?
They can, which is why the workflow grounds generation in your product attributes, runs automated checks for unsupported facts and keeps a human approval step, especially for regulated categories.
Can the system write in our brand voice?
Yes, when voice is encoded as approved examples, vocabulary lists and claim rules, and refined using reviewer edits. A single paragraph describing the tone is rarely enough.
Does this work for multiple languages and marketplaces?
Yes. Each channel and market gets its own template, length limits and glossary. Localised output should still be reviewed by someone fluent in the target language, at least until quality is established.
Where do approved descriptions go?
Typically back into your PIM, ecommerce platform or marketplace feeds through an integration, with a record of who approved each item and when.
Do we need clean product data before starting?
Reasonably complete attributes are needed for good drafts. Many projects begin with a short data review and clean-up for the categories in the first phase.
Related reading
Continue your research
Could generative AI clear your product content backlog without risking accuracy?
We can review your product data and channels and design a drafting and approval workflow that fits your catalogue.
