Every shop, marketplace and marketing team needs a lot of short, persuasive and correct text: product titles, descriptions, bullet points, ad headlines, newsletter teasers. Manufacturer data is often thin or inconsistent, and writing by hand does not scale to thousands of SKUs and several languages. LLMs turn structured product data, photos and a brand guide into channel-specific copy in seconds. The work shifts from writing to data quality, brand rules, claim checking and approval. This is one of the most widely adopted generative AI use cases and is now built into most shop, marketplace and PIM software.
Problem & value
- Today: copywriters or category managers write texts per SKU and channel. Supplier data is copied in unchanged, so you get duplicate content, missing attributes and weak SEO. Long-tail products often have no description at all, and new markets wait for translations.
- AI value: coverage of the full catalogue, faster time-to-shelf, consistent tone, channel-specific variants (character limits, marketplace style guides) and many ad variants for testing.
- Measured examples (self-reported by vendors and agencies):
- A Polish building-materials retailer with 10,000+ products cut writing time from 15 minutes to under 2 minutes per product. In early review, 50% of AI drafts were approved as they were, 45% needed minor edits and 5% were regenerated [1].
- Amazon reports that more than 900,000 sellers have used its generative listing tools. Merchants accept the generated content without edits more than 90% of the time. TechCrunch notes this may partly mean sellers do not check outputs closely [2][3].
- axite (AX Semantics, Stuttgart) reports 52 million texts for XXXLutz in 13+ languages and a +50% conversion at babyartikel.de. These are vendor claims and were not independently verified [4].
- Vendor conversion claims vary widely and are rarely backed by controlled tests, so plan your own A/B test before you count on a revenue uplift.
Variants
- SMB online shop / marketplace seller: one-click generation inside Shopify, Shopware, WooCommerce, Amazon Seller Central, eBay or Etsy, from a few keywords, a photo or a product URL [3].
- Retail / brand manufacturer at scale: PIM-driven bulk generation for 10k–1M SKUs with attribute completeness checks, per-channel templates and translation, e.g. Akeneo, Salsify, Pimcore [5].
- B2B / manufacturing: turning technical data sheets and ETIM/ECLASS attributes into readable catalogue and distributor texts. The data enrichment itself is in DATA-010.
- Performance marketing: headline and description variants for Google/Meta ads, newsletter subject lines and landing-page sections. Bidding and creative optimisation are in REC-007.
- Tourism & hospitality: descriptions of hotels, holiday flats and experiences from structured property data, for booking portals and destination marketing organisations.
- Private use: listings for second-hand sales (Kleinanzeigen, Vinted, eBay) from a photo.
- Out of scope here: long-form SEO articles (LANG-005), social posts (LANG-007), rewriting existing texts into the brand voice (LANG-011), regulated claim review (LANG-024) and imagery (GEN-001).
How it works
- product data (PIM/ERP/feed) + photos + brief
- data check (required attributes present? units normalised? conflicts?)
- grounding pack (attributes, data sheet facts, brand guide, approved examples via retrieval)
- LLM / VLM generation per channel template (length, tone, keywords, structure as JSON)
- automated checks (facts only from source data, banned/regulated claims, length, duplicates, language)
- human review & approval (sample or 100% depending on risk)
- publish to shop/marketplace/feed (API)
- A/B test & performance feedback
- Grounding is the key design choice. The model should write only from the attributes it receives. Prompt it to output "missing" rather than invent, then check every number and material in the output against the source record. A structured output (JSON per field) makes this testable.
- Brand voice comes from a style guide plus 10–50 approved examples retrieved by similarity (embeddings), and less often from fine-tuning.
- VLMs read product photos to fill gaps (colour, style, visible features). Treat anything inferred from an image as a suggestion that needs confirmation.
- Human-in-the-loop: approval of new templates, full review of regulated categories (cosmetics, food, health, toys, green claims) and sample checks on the rest.
Data preparation
The raw material is product data: attributes, materials, dimensions and certificates in ERP, PIM or shop; photos and supplier data sheets (often PDFs) on file shares or in e-mail; plus the brand voice guide, banned-claims list and a few approved example texts, which usually exist only in people's heads or scattered Word files.
Typical gaps: missing or inconsistent attributes, the same product with different spellings across systems, data sheets only as scans, no written list of forbidden claims (health, "best", sustainability), and no approved examples per channel.
Preparation steps: - Export product data from ERP or shop and clean it into one tidy table (see PREP-011 and PREP-005); fill the mandatory attributes first, because the model must not guess them. - Run supplier PDFs and scans through OCR (see PREP-006, DOC-014). - Write down the brand voice, claim rules and 5 to 10 gold-standard texts per channel (see PREP-016, PREP-015). - Check the legal basis if reviews or customer data are used (see PREP-025).
Rough share: about 40 to 60 percent of the total effort. The model setup takes days; fixing product data takes weeks. See GUIDE-003.
Implementation paths
Which path for whom
- SMB: Path A. Shop, marketplace and PIM tools ship mature, EU-hostable copy generators at SMB prices, and product copy is a standard process.
- Enterprise: Path B. Thousands of SKUs, many channels and an existing PIM call for API-based generation integrated into the product data flow, with brand and claim checks of your own.
- Developer: Path B. Cloud APIs with structured outputs are the fastest way to build feeds and plug-ins for several clients.
- Private: Path A. Hobby sellers and small shops get what they need from the built-in assistants of their marketplace or shop system.
Path A — Buy (recommended for most SMBs)
- Built into commerce platforms: Shopify Magic, Shopware AI Copilot, Amazon listing tools (built on Bedrock [2][3]), Google Merchant Center Product Studio, eBay "magical listing". Free or included. Good enough for small catalogues.
- PIM / product content platforms: Akeneo (FR) "Text and Image Enrichment" can connect to your own LLM for governance [5]. Others include Salsify and Pimcore (AT, open-source core).
- Specialised e-commerce text generators (EU): axite / AX Semantics (DE) combines LLM generation with structured-data guardrails in 110+ languages [4]. Retresco (DE) does data-to-text.
- Marketing writing assistants: neuroflash (DE). It stores texts on German servers in the Open Telekom Cloud and routes generation to GPT, Gemini, Claude or Mistral [6]. Others include Jasper, Writer, Copy.ai, DeepL Write (DE, rewriting) and Microsoft 365 Copilot / Gemini in Workspace.
- Choose A when: the catalogue fits standard templates and the shop or PIM already offers it. Check where the model runs, whether your data is used for training, and whether you can enforce banned terms.
Path B — Build on cloud APIs
- Models: Claude, GPT and Gemini (each offers EU data residency or EU regions via Azure OpenAI, Vertex AI or Amazon Bedrock Frankfurt), and Mistral (FR) Medium 3.5 / Large 3 via La Plateforme [7]. Smaller, cheaper models (e.g. Mistral Small 4, Gemini Flash class) are enough for bulk descriptions. Use frontier models for headline creativity and difficult categories.
- Pattern: batch API jobs over the PIM export, with JSON-schema structured outputs and a second "checker" prompt or rule engine that compares claims with the source attributes. Translate with DeepL API or the LLM itself (LANG-009).
- Integration: n8n / Make / Power Automate or a small service (Python + LangChain or plain SDK). Read from the PIM/shop API (Akeneo, Shopware, Shopify Admin API, Amazon SP-API), write back drafts with status "to review". One agency build for 8,205 products used OpenAI API + LangChain + Retool as the review UI [1].
- Choose B when: you have custom templates, many channels or languages, or want your own checks.
Path C — Self-host (open source)
- Models: Qwen3-30B-A3B-Instruct (Apache 2.0, 3.3B active parameters, fast on one GPU) [8], Mistral Small 4 / Ministral 3 (Apache 2.0) [7], Gemma 4 (12B, 26B MoE, 31B; multimodal, can read product photos) [9]. Larger: Mistral Large 3 (Apache 2.0) or Qwen3-235B class on HW3/HW4.
- Stack: vLLM or Ollama, structured outputs via vLLM guided decoding / Outlines, a small embedding model (e.g. multilingual-e5 or bge-m3) for example retrieval, n8n self-hosted for the workflow, and LanguageTool (self-hostable) for spelling and grammar.
- Hardware: HW2 (one 24–32 GB GPU) runs a ~30B MoE or 12–14B dense model quantised. That can produce tens of thousands of descriptions per night. HW3 if you want ~70–120B quality for German marketing tone.
- Choose C when: product data is confidential before launch, or volumes are large enough that per-token cost matters. German style quality of small models is good but needs review.
Path D — Custom training
Worth it only at large scale with a stable house style: LoRA fine-tuning a 7–30B model on a few thousand approved, high-performing texts per language cuts prompt length and improves voice consistency. Edits from the review queue and A/B winners are the natural training data. For most organisations, prompting plus retrieved examples is enough.
State of the art (2026)
- Commodity in platforms: generation from keywords, a photo or a URL is standard in Amazon, Shopify, Google Merchant Center and major PIMs. Amazon moved from writing new listings (2023) to enhancing existing listings with missing attributes (May 2025) [2][3].
- Multimodal input: open and closed models (Gemma 4, Mistral 3 family, Qwen3-VL, Gemini, GPT, Claude) read images, so photo-to-listing works without separate captioning [9].
- Guardrailed generation: vendors combine LLM fluency with deterministic checks against product data ("intents instead of prompts" [4]). Agentic PIM features flag data gaps before generation [5].
- Generative engine optimisation: descriptions are increasingly written to be quoted correctly by AI shopping assistants, not only for Google ranking (see REC-011).
- Limits and failure modes: invented specifications (materials, certificates, compatibility), generic "AI-sounding" copy that all looks the same, unlawful claims ("eco-friendly", "clinically proven"), wrong units after translation, and reviewers who stop checking. The high acceptance rates are not proof of accuracy [2].
Shared resources & combinations
- Same LLM endpoint / GPU (llm-general, HW2–HW3): LANG-002 email drafting, LANG-005 SEO content, LANG-007 social posts, LANG-011 brand-voice rewriting. A single shared brand-voice pack (style guide + example store in a vector index) serves all of them.
- Upstream data: DATA-010 product data enrichment (PIM, ECLASS, ETIM). Good attributes are the biggest quality lever for generated copy.
- Translation: LANG-009 for market localisation of the generated texts.
- Downstream / adjacent: LANG-024 marketing compliance review (claim checks), GEN-001 product imagery, GEN-027 landing pages, REC-006 email personalisation, REC-007 ad creative optimisation, REC-011 generative engine optimisation.
- Workflow engine (n8n / Power Automate) and the review UI can be reused for every LANG.WRITE case.
Effort, cost & KPIs
- Path A: days. Often included in the shop/PIM plan or a per-seat writing tool subscription.
- Path B: 2–6 weeks for a pipeline with PIM integration, checks and review UI. Token cost per description is typically a fraction of a cent to a few cents, depending on model and length.
- Path C: 4–8 weeks plus GPU server and operations. This pays off with very large or confidential catalogues.
- Cost drivers: data cleanup, number of channels × languages, review effort in regulated categories.
- KPIs: catalogue coverage (% SKUs with complete copy), time-to-shelf, first-pass approval rate, factual error rate in sampled audits, edit distance, conversion rate and add-to-cart rate (A/B test), organic impressions, ad CTR, returns attributed to wrong descriptions.
People & change
Start with management (budget, brand and claims policy), the marketing or e-commerce lead as process owner, and the data owners of ERP and PIM, who must release the product data. Involve IT for interfaces and accounts, and legal or compliance early for claim rules (UWG, health and environmental claims). Staff who write the texts today are the key group: pick two or three of them as a pilot group and let them define the review checklist.
Works council: usually no, as the tool processes product data and does not monitor staff. It still depends on the product. If it logs individual users' prompts and output, §87(1) no. 6 BetrVG may apply, and §90(1) no. 3 requires timely information about AI-planned work processes. See GUIDE-004. DPO and DPIA are only needed if customer reviews or personal data flow in; conclude an AVV (Art. 28 GDPR) with every provider.
AI literacy: support staff with a short training on prompting, hallucinated claims and review duty (Art. 4 AI Act, see GUIDE-006). Typical concern: "AI replaces copywriters." Address it openly: the work shifts to quality and brand control, and nothing is published without human approval (GUIDE-008).
Risks, regulation & compliance
- EU AI Act: not high-risk (no Annex III category). Art. 50 applies from 2 August 2026 [10][11] (Digital Omnibus, Reg. 2026/1744: systems already on the market before 2 Aug 2026 have until 2 Dec 2026 for Art. 50(2) marking).
- Art. 50(2), providers: whoever provides the generative system (the model vendor, and you if you build your own generator under Path B, C or D) must mark outputs machine-readably. The exemption for assistive or standard-editing functions is narrow. The Commission's voluntary Code of Practice on marking and labelling (final June 2026) is the reference [11].
- Art. 50(4), deployers: the labelling duty for AI text covers only text that informs the public on matters of public interest, and it falls away after human review with editorial responsibility. Ordinary product and ad copy is usually outside it.
- Art. 4: an AI literacy duty applies to staff.
- Consumer law (main risk): the seller is liable for misleading copy whether a person or an AI wrote it (UCPD, in Germany the UWG). Directive (EU) 2024/825 applies from 27 September 2026. It bans generic environmental claims without recognised excellent environmental performance, product claims based on emissions offsetting, and sustainability labels without certification [12]. Put these terms on the banned list.
- Sector rules: health claims for food (Regulation 1924/2006), cosmetics claims (Regulation 655/2013), German Heilmittelwerbegesetz for health products, the textile labelling regulation, and product safety information in online offers under the GPSR (Regulation 2023/988). Route these categories to full review or to LANG-024.
- Copyright & marketplaces: do not feed competitor texts in for paraphrasing. Follow marketplace content policies.
- GDPR: product data is not personal. Personalised copy, customer reviews used as input, or CRM segments are. Use an EU-hosted provider with a data processing agreement and no training on your data.
- Mitigations: generate facts only from source attributes, use automated claim and banned-term checks, require mandatory review for regulated categories, keep an audit log of AI draft vs. published text, and run regular sample audits.
Getting started
- Pick one category with 200–500 SKUs and clean attribute data. Measure the current coverage, time-to-shelf and conversion.
- Write a one-page brand voice guide plus a banned-claims list (green claims, health claims, superlatives), and collect 20–50 of your best existing texts as examples.
- Test the generator already in your shop/PIM (Path A) against a Path B prototype on the same SKUs. Have two reviewers score factual accuracy and tone.
- Add automated checks (attribute consistency, length, banned terms) and a review status in the PIM before anything goes live.
- Run an A/B test of AI vs. existing copy for 4–8 weeks, then roll out by category and language.
Sources
- SUNZINET — case study: AI content generation for 10,000 products (construction-materials retailer) — https://www.sunzinet.com/en/references/scale-10000-products-ai-content-generation
- TechCrunch — Amazon's newest AI tool is designed to enhance product listings (8 May 2025) — https://techcrunch.com/2025/05/08/amazons-newest-ai-tool-is-designed-to-enhance-product-listings
- Amazon — Generative AI-powered product listings for sellers — https://www.aboutamazon.com/news/innovation-at-amazon/amazon-generative-ai-powered-product-listings
- axite (AX Semantics GmbH, Stuttgart) — product content platform and customer results — https://axite.io/en/
- Akeneo — product experience platform, AI text & image enrichment — https://www.akeneo.com/
- neuroflash — AI copilot for marketing teams (data hosting statement) — https://neuroflash.com/neuroflash-for-marketing-teams-businesses/
- Mistral AI — models overview — https://docs.mistral.ai/getting-started/models/models_overview/
- Qwen3-30B-A3B-Instruct-2507 model card — https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507
- Google — Gemma models overview — https://ai.google.dev/gemma/docs/core
- EU AI Act, Article 50 (transparency obligations) — https://artificialintelligenceact.eu/article/50/
- European Commission — Code of Practice on marking and labelling of AI-generated content — https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content
- European Commission — Directive (EU) 2024/825 Empowering consumers for the green transition — https://transition-pathways.europa.eu/legislation/directive-eu-2024825-empowering-consumer-green-transition
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj