Sellers generate product cards with AI: 7,500 examples over eight months

Over eight months, a marketplace card-generation service processed more than 7,500 requests. The Habr team analyzed seller behavior and found some unexpected patterns: users think in diffs, not prompts; activity peaks in the evening; and under the hood runs a cascade of language models with Redis pub/sub for real-time feedback.

How sellers actually work

It doesn't start with a perfect prompt. Sellers don't write polished instructions. They upload a draft card, see the AI's output and fix it through diffs in the interface.

• More than 7,500 generations over eight months of the service's operation
• Peak activity: 18:00–22:00 (evening, when sellers generate after their main work on marketplaces)
• Users prefer editing finished output (diffs) to rewriting prompts from scratch

This means sellers don't plan content. They work reactively: upload the product → click generate → look at the result → edit the lines they need. The cycle takes 2–3 minutes. There is no prompt refinement: they take the first draft and whittle it down.

Architecture under the hood

The service uses fan-out: it splits the generation of a single card into parallel subtasks: title, description, tags, specifications, price recommendations. Each subtask goes out as a separate call to different models.

The key feature is a model cascade with fallback. If the primary model is overloaded or returns low-quality output, the system switches automatically to an alternative with no visible delay. Real-time progress is pushed to the frontend via Redis pub/sub: the seller sees in the browser which parts of the card are ready and can start editing the title while the description is still being generated.

This is critical for UX. A seller will tolerate waiting 3–5 seconds. At 30 seconds, they close the tab. If the system queried models sequentially, response time would be unacceptable for small sellers.

Why prompts are a myth

The marketing of AI tools insists that a perfect prompt equals a perfect result. In practice, sellers settled on a different philosophy long ago. They don't write perfect instructions.

Instead: generate → look → edit → submit. AI serves as an intelligent assistant for the first draft, while the seller remains the final editor, responsible for the result. This behavior explains the demand for tools with a good editing UI (diffs, change highlighting, version history).

What the audience looks like

The data comes from a live service. Its users are sellers running small and mid-sized shops on Yandex.Market, Ozon and Wildberries. The 18:00–22:00 activity peak suggests this is a tool for small business owners, not for the content departments of large marketplaces: people generate cards in their free time, after their main work.

This sets the requirements for price (a seller pays $5–10, not $100) and speed (generation in 3–5 seconds, not a minute). The seller's time is worth more.

What this means

AI for marketplace content works not as a full autopilot but as a work assistant. Sellers are ready to delegate the first draft to a neural network, but the final say stays with them. This explains the growing demand for tools with a smart editor: sellers are willing to pay not for perfect generation, but for the convenience and speed of editing.