AI in E-Commerce: Why Product Data Is More Important Than P…

Many retailers believe that a clever ChatGPT prompt solves the problem of product communication. The reality looks different. AI in e-commerce only reaches its full potential when the foundation is right — and that foundation is not the perfect prompt, but the structured product data model. Jan Kaiser, founder of Xanevo and AI practitioner since 2018, explains in the podcast "Behind the Scenes in Marketing" why artificial intelligence in e-commerce stands or falls with clean data. In this article, we summarize the key insights and translate them into concrete recommendations for marketing managers and e-commerce teams.

From classic machine learning to modern AI in e-commerce

Jan Kaiser is no newcomer to the world of AI. He comes from a classical computer science background, built his first neural network in 2019 — a reinforcement learning system for pharmaceutical quality control — and has built AI departments in corporations in the pharmaceutical and fintech sectors.

To the podcast episode:

▶

This background shapes his view of the current market. While today anyone can build AI workflows for 20 euros a month, just a few years ago even six-figure budgets were on the agenda for prototypes alone.

"As of today, people already flinch when a prototype costs five figures. That shows how rapidly the development has taken place."
— Jan Kaiser

This democratization of AI technology is both an opportunity and a risk: it opens the door for many new providers — including those who sell superficial solutions that fall apart at the first real-world application.

Why "just using ChatGPT" falls short for AI in e-commerce

A recurring pattern in Xanevo's consulting practice: retailers throw supplier data into the ChatGPT interface and expect structured, shop-compliant product data as output.

The result is usually disappointing — or worse: incorrect. Because a language model is trained to be as helpful as possible. It will therefore fill in fields even when the necessary information is not present in the source data at all. This phenomenon is called hallucination.

"An AI is trained to be incredibly helpful and always wants to give an answer. That means it sometimes hallucinates values because it tries to complete a task as precisely as possible."
— Jan Kaiser

The real problem lies not in the prompt, but in the data foundation. If you don't give the model a clear target data model — that is, precisely defined attributes, fields, and permitted values — you get outputs that may sound plausible but can be factually incorrect. A vacuum cleaner that is incorrectly attributed a wet cleaning function leads directly to returns and customer complaints.

Optimizing product data: The true foundation of any AI strategy

Before any AI tool is deployed, a clear product data hierarchy is needed. Anyone who wants to optimize product data should draw on established industry standards:

• E-Class: A hierarchical category tree that classifies products from a top-level category down to a specific leaf node (e.g., "Garden tools" → "Electric hedge trimmer")
• ETIM: An attribute and data standard that defines which properties a product category should have and which values are permitted

These standards do not need to be adopted one-to-one. They serve as a starting point that can be tailored to your own shop structure, SEO requirements, and target audience.

"The simplest thing you can do if you don't yet have a clean product data hierarchy is: take one of these standards, adopt the branches that work for you, cut off what doesn't fit, and then look at how you transfer it into your front-end category tree."
— Jan Kaiser

Why clean product data decisively improves AI quality

The more precisely a data model is defined, the more targeted artificial intelligence can be used in e-commerce. In concrete terms, this means:

• The model knows exactly which fields it should fill in
• It can recognize when a field must remain empty because the information is missing from the source data
• Quality assurance steps can be described programmatically and automated

Anyone who wants to optimize product data simultaneously creates the prerequisites for scalable AI workflows — and saves considerable manual rework in the long run.

Small language models vs. large language models: What to use when in AI e-commerce?

A central topic is the question of which model is best suited for which purpose. Jan Kaiser draws a clear distinction between two scenarios:

Small, specialized models (small language models)

When useful: When the use case is narrowly defined, the product range changes little, and high volumes at low costs are critical.

Real-world example: For text recognition in PDF documents, using a large model via API costs around 25 cents per document. A specially trained small model handles the same task for 0.01 cents — with comparable accuracy.

Typical use cases:
— Text recognition in documents
— Rule-based mappings for stable product categories (e.g., fashion with fixed seasonal ranges)
— Grammatical adjustments and templating

Large commercial language models (GPT-4, Gemini, Claude, etc.)

When useful: When input data is variable (sometimes CSV, sometimes PDF, sometimes images), the range changes frequently, and no proprietary models are to be trained.

Typical use cases:
— Product text generation
— Translations into many languages
— Data onboarding for broad, evolving product ranges

"If I have uncertainty about what's coming in, and output uncertainty because the data model can change, then it's naturally a great fit to use a large language model and adapt the content process with prompts, rather than training your own models."
— Jan Kaiser

The multi-LLM approach: Security through majority vote

A particularly practice-relevant method from the Xanevo toolkit is the multi-LLM approach. Here, the same task is processed in parallel by multiple language models and the results are compared with one another.

The principle can be compared to an expert panel: if nine out of ten specialists agree on a statement, that statement can be trusted far more than when opinions are evenly split.

Practical implementation:
— If four out of five models agree → the result is automatically accepted
— If only three out of five agree → a human reviews the case
— This approach is used for mappings, translations, and text generation

Measured result: In a concrete OCR project, it turned out that the AI achieved a precision of over 90% — while the human process that served as the reference was only at 72%. The supposed gold standard was the human, not the machine.

When does automation with artificial intelligence in e-commerce pay off?

Jan Kaiser outlines a clear scale that companies can use as a guide:

"As of today I would say: if you're operating in the three-digit range, AI already makes sense — but more as an assistive tool, not as full automation."
— Jan Kaiser

The decisive question is not only the volume, but also the rate of change of the product range and the variance of input data.

Internationalization with AI in e-commerce: Why the big-bang approach fails

According to Jan Kaiser, internationalization projects frequently fail due to a structural error: a budget is set, content is produced in one language, everything is rolled out all at once — and then people wonder why results are not forthcoming.

His counter-proposal is the iterative test approach:

1. Segment the content: What needs to be translated? Navigation and legal texts absolutely require human review. Technical product data (e.g., dimensions of a dowel) can safely be translated fully automatically.
2. Test different quality levels: For 100 articles, have a native speaker write the content; for another 100, work with ChatGPT; for another 100, use a multi-LLM approach.
3. Measure results: Which variant achieves better rankings, more traffic, higher conversion rates?
4. Scale incrementally: What works gets expanded. What doesn't work gets discarded.

Localization vs. translation: An underestimated factor in AI e-commerce

Automatic translation and genuine localization are not the same thing — a frequently underestimated aspect when deploying artificial intelligence in e-commerce.

Concrete examples from the conversation:
— In Switzerland, a bicycle is called "Velo" — anyone advertising with "Fahrrad" loses visibility
— The term "Checkout" is usually not translated as "Kasse" in German shops, even though that would be the correct translation
— Hedge trimming regulations are governed by law in Germany — such compliance texts may not exist in other countries at all and should not simply be carried over

"There are EU-standardized hazardous substance notices with H-statements and P-statements that are already available in translated form in standards databases. I don't want to translate those with AI at all — someone else has already translated them correctly for me."
— Jan Kaiser

Compliance texts and marketing texts: Always keep them separate

In the CMS or PIM system, a clear distinction should be made between:

• Marketing texts (product descriptions, benefits, usage instructions) → AI translation with optional human review is suitable
• Compliance texts (warning notices, hazardous substance information, legal notices) → Use of standards databases or mandatory human review

Fine-tuning or prompting: What actually makes a difference in AI e-commerce?

A question that in practice