What generative AI is
Generative AI is the umbrella term for models that create new content rather than pick an answer from a fixed set of options. The term went mainstream after ChatGPT was released to the public on 30 November 2022: language models gained an interface that anyone could use, not just developers.
For e-commerce, the most useful distinction is between generative and predictive models — they solve different problems and are evaluated differently.
| Predictive (discriminative) ML | Generative AI | |
|---|---|---|
| Output | A class, a probability, a rank | A new object: text, an image, code |
| Typical question | “Will this user buy?” | “Describe this product”, “Pick options and explain why” |
| Examples in a store | Recommendations, category sorting, churn prediction | Assistant, product descriptions, creatives |
| How to evaluate | Metrics on labelled data + A/B tests | Factual accuracy of answers + A/B tests |
Main types of generative models
- Large language models (LLMs) predict the next token and, as a result, write coherent text and code. Assistants and description generation are built on them.
- Reasoning models are LLMs that work through a chain of reasoning before answering; see Reasoning Model.
- Diffusion models generate an image by gradually removing noise from a random starting picture (Stable Diffusion is one example). They are used for banners and backgrounds.
- Multimodal models accept and produce several types of data: for instance, they can describe a product from its photo.
Where generative AI works in e-commerce
- Shopping assistant — answers product questions and suggests options in a conversation. Answers are grounded in catalogue data through RAG.
- Product descriptions — first drafts for product pages that have no copy, especially in the long tail of the range.
- Search — parsing long conversational queries such as “a gift for my dad who loves fishing, under $60”.
- Review summaries — a short digest of hundreds of reviews on the product page.
- Creatives — banner and headline variants for A/B tests.
Limits and how to manage them
- Hallucination. A model can confidently invent a specification or a price. The fix is grounding in the catalogue, rules in the system instructions and an explicit “nothing found” path.
- Cost and latency. Every answer is an inference call billed by the token. Request size is capped by the context window, and filling it well is a discipline of its own — context engineering.
- Memory. A model does not remember a shopper between requests; whatever needs to carry over, the application stores itself — that is agent memory.
- Quality control. Generated descriptions are spot-checked by people, while the assistant is tested on a set of typical questions before launch and with an A/B test after it.
Important: generative AI does not replace the store’s data. An assistant’s answers are only as good as the product feed: if a product record has no size or material, the model cannot answer honestly about them.