The product catalogue is the sales engine of any online store. When a company manages 50 references, writing each listing by hand is reasonable. When the catalogue grows to 500, 2,000 or 5,000 SKUs, that model breaks down: the team cannot keep up, listings stay incomplete or are copied verbatim from the supplier — with the risk of being penalised for duplicate content that this entails. Generating product listings with artificial intelligence has moved from being a promise to a productive workflow that several Spanish SMEs have already implemented with measurable results.
This article explains how that workflow operates, which tools are involved, what the most common mistakes are, and when it makes sense to rely on a specialist consultant to prevent automation from producing mediocre content at industrial scale.
Why the product listing problem is bigger than it looks
A product listing that converts is not just a descriptive paragraph. It includes a search-optimised title, a benefit-oriented description, the correct technical attributes (size, material, weight, compatibility…), the long-tail keywords the real buyer uses, the brand tone, and in many cases variants adapted for different markets or channels. Multiplying that work across thousands of references is unviable manually.
According to Semrush data from 2025, 34% of product pages in Spanish ecommerce have duplicate or supplier-copied descriptions, which drastically reduces their organic visibility. The solution is not to hire ten copywriters: it is to implement an AI-assisted workflow that generates unique, brand-consistent drafts that a human editor reviews and approves in a fraction of the time it would take to write them from scratch.
How the real AI product listing generation workflow works
The process does not start with a language model: it starts with data. Without structured input data, AI produces generic text that does not differentiate your product from the competition. The typical workflow has five phases:
1. Catalogue cleaning and structuring
The starting point is usually an Excel file or an ERP export with heterogeneous columns, inconsistent field names and empty values. Before feeding any model, that file must be normalised: unify units, fill in mandatory fields, remove reference duplicates and map the supplier's attributes to the ecommerce data schema. This phase is the most underestimated and the one that most determines the quality of the final result.
2. Listing template definition
A master prompt is created that defines the brand tone, the text structure (title, lead, benefits paragraph, technical specifications, call to action), the target length and the constraints — what cannot be claimed, which terms to avoid, whether or not competitive comparisons are included. This prompt acts as a mould: all generated drafts will follow that structure.
3. Batch generation
With a clean catalogue and a defined prompt, generation runs in batch: the system takes each catalogue row as context and produces a listing. Depending on the volume and the tool chosen, this can be done with a language model accessible via API (OpenAI GPT-4o, Anthropic Claude, Google Gemini) integrated in n8n or Make, or through specialist platforms such as Describely, Akeneo with an AI module or Jasper Commerce. For catalogues of more than 1,000 references, direct API integration is more flexible and cost-effective than closed platforms.
4. Human review and validation
AI generates drafts; an editor approves, adjusts or rejects them. In a well-calibrated workflow, an editor can review between 80 and 150 listings per day — compared with the 15–25 they could write from scratch. The review focuses on verifying that the technical attributes are correct (AI can hallucinate values not present in the input data), that the tone is consistent and that there are no claims that breach advertising regulations or the selling channel's policies.
5. Publication and iteration
Approved listings are imported into the ecommerce platform (WooCommerce, Shopify, PrestaShop, Magento) via API or CSV. The workflow does not end here: conversion and organic ranking data feed periodic improvements to the master prompt, closing the continuous improvement cycle.
To learn in detail how we design and implement this type of workflow, see our AI content generation service, where we accompany you from data architecture through to pipeline deployment.
Tool comparison for generating product listings with AI
| Tool | Type | Best for | Ecommerce integration | Indicative cost (2025–2026) |
|---|---|---|---|---|
| OpenAI API (GPT-4o) | Base model API | Custom pipelines, large catalogues | Via n8n / Make / own code | ~€2–5 per 1,000 listings (estimated by tokens) |
| Describely | Specialist SaaS | Marketing teams without development | Shopify, WooCommerce, BigCommerce | From ~$28/month (Starter plan, 2025–2026, indicative) |
| Akeneo + AI | PIM with AI module | Complex multi-channel catalogues | Native with major ecommerce platforms | PIM licence + API token cost |
| Jasper Commerce | AI copywriting SaaS | Brands with a defined style guide | CSV export / API | From ~$49/month (Creator plan, 2025–2026, indicative) |
| n8n + model API | Open-source automation | Maximum flexibility and data control | Any ecommerce platform with API/webhook | Server cost + API tokens |
Token costs are indicative and vary according to the model chosen, listing length and volume. For an SME with 2,000 references generating listings of around 300 words each, the generation cost with API models typically falls between €50 and €200 in total — well below the cost of a full-time copywriter for a month.
Which factors determine the quality of the result
The quality of the generated listings depends on three variables, in order of importance:
Input data quality
A language model cannot invent precise technical characteristics that are not in its input data. If the catalogue has the «material» field empty for 60% of rows, the generated listings will also have it empty or, worse, the model will fill it with plausible but incorrect information. Garbage in, garbage out: investment in data cleaning is what generates the highest return in this type of project.
Prompt design
A generic prompt produces generic text. A well-designed prompt includes examples of your own high-performing listings, explicit constraints (length, structure, forbidden words), the target buyer profile and the keywords that should appear naturally. Designing that prompt is an iterative task that takes 2 to 5 days in real projects.
Review process
AI makes mistakes that a human editor detects in seconds: confusing units (centimetres for millimetres), asserting properties the product does not have, or using incorrect technical terminology for the sector. Defining which types of error are critical (block publication) and which are minor (corrected in the next prompt iteration) allows the review to scale without sacrificing accuracy.
Common use cases in Spanish SMEs
The use of AI for product listings is not limited to large retailers. In 2025 and 2026 we have seen accelerating adoption across several sectors:
- Industrial distributors with catalogues of fasteners, tools or electrical components: thousands of references with very similar technical attributes, where AI differentiates the description without inventing properties.
- Fashion and footwear stores: generating listings by colour and size, adapting the text to the season without repeating the same paragraph for twenty variants.
- Gourmet food distributors: listings that highlight origin, production method and food pairing from supplier data.
- Marketplaces and dropshipping: transforming supplier listings into unique content to avoid duplicate content on Amazon, El Corte Inglés or Miravia.
Legal and ethical implications to consider
Using AI to generate commercial descriptions is subject to the same regulations as any advertising content. The Unfair Competition Act and the Omnibus Directive (transposed into Spanish law in 2022) prohibit misleading claims about product characteristics, even when they come from an automated system. Liability always rests with the merchant, not the AI model. This means the human review workflow is not optional: it is a legal obligation.
Additionally, the EU AI Act, in force since August 2024, classifies AI systems that generate content intended for the public as systems that must meet transparency requirements. Although product listings are not in the «high risk» category, they are subject to minimum transparency obligations if the system is made available to the public on a systematic basis. For companies that want to go beyond occasional use and build a sustainable content generation system, it is worth reviewing the requirements of the AI content generation service, which includes that compliance analysis.
How to measure return on investment
The return on an AI product listing project has three measurable dimensions:
- Time saved on copywriting: compare hours spent before and after producing listings of comparable quality. The typical saving is between 60% and 80% of copywriting time.
- Improvement in organic ranking: unique listings with relevant keywords increase visibility in Google Shopping and organic search. The impact becomes visible between 60 and 90 days after publication, depending on domain authority.
- Increase in conversion rate: more complete, benefit-oriented listings reduce buyer hesitation. In projects where the impact has been measured, the improvement in conversion on references with improved listings has been significant compared with the previous generic listings.
Frequently asked questions
Can AI generate product listings directly from the supplier catalogue?
Yes, but with caveats. AI can take the supplier's PDF or Excel as context and generate a differentiated listing from that data. The problem is that if the supplier catalogue is poor in information (only reference, EAN code and price), the generated listing will also be poor. The prior enrichment work — adding materials, dimensions, use cases, target audience — is what makes the result useful and unique.
What about SEO? Does Google penalise AI-generated content?
Google does not penalise AI-generated content for being AI-generated: it penalises low-quality content, regardless of whether it is human or AI in origin. Its quality guidelines (E-E-A-T: Experience, Expertise, Authoritativeness and Trust) apply equally to human content and AI-generated content. The key is that listings provide real and differentiated information, not interchangeable generic paragraphs. A well-calibrated workflow produces content that passes that test without difficulty.
How long does it take to get this type of workflow up and running?
An AI listing generation implementation project has three phases with indicative timelines: catalogue cleaning and structuring (1–3 weeks, depending on data quality), prompt design and calibration (1–2 weeks, with iterations on real samples) and generation, review and publication of the first batch (1–2 weeks). In total, a well-executed project can have the first listings published in 4 to 6 weeks from the start. The subsequent production rate depends on the speed of the editorial team's review.
Does an internal technical team need to maintain the workflow?
Not necessarily. If the workflow is built on no-code automation tools such as n8n or Make, a marketing manager with basic training can handle day-to-day operations: uploading the updated catalogue, launching generation and reviewing drafts. Technical pipeline maintenance (updates, model changes, prompt adjustments) can be outsourced. What does require internal presence is the review of the generated content: that editorial responsibility cannot be fully delegated to an external provider.