Селлеры генерируют карточки товаров через AI: 7 500 примеров за восемь месяцев

Over eight months, the product card generation service for marketplaces processed over 7,500 requests. The Habr team analyzed seller behavior and uncovered unexpected patterns: users think in diffs, not prompts; peak activity occurs in the evening; and under the hood, a cascade of language models with Redis pub/sub enables real-time response.

How Sellers Actually Work

Everything doesn't start with the perfect prompt. Sellers don't write polished instructions — they upload a draft card, see the AI result, and edit it through diffs in the interface.

• Over 7,500 generations in eight months of service operation
• Peak activity: 6:00 PM–10:00 PM (evenings — sellers generate after their main marketplace work)
• Users prefer editing the finished product (diffs) over rewriting prompts from scratch

This means sellers don't plan content. They work reactively: upload a product → click generate → view the result → edit the necessary lines. The cycle takes 2–3 minutes. No prompt refinement: take the first draft and work through it.

Architecture Under the Hood

The service uses fan-out — it breaks a single card generation into parallel subtasks: title, description, tags, characteristics, pricing recommendations. Each subtask goes as a separate call to different models.

The key trick is a cascade of models with fallback. If the primary model is overloaded or produces low-quality results, the system automatically switches to an alternative without visible delay. Real-time progress is sent to the frontend through Redis pub/sub: the seller sees in the browser which parts of the card are already ready and can start editing the title while the description is being generated.

This is critical for UX. A seller waits 3–5 seconds — they can handle it. They wait 30 seconds — they close the tab. If the system polled models sequentially, the response time would be unacceptable for small sellers.

Why Prompts Are a Myth

AI tool marketing insists: the perfect prompt = the perfect result. In practice, sellers have long adopted a different philosophy. They don't write perfect instructions.

Instead: generate → look → edit → send. AI is used as an intelligent assistant for the first draft, and the seller remains the final editor, responsible for the result. The demand for tools with good editing UI (diffs, change highlighting, version history) is explained by this behavior.

What the Audience Looks Like

Data is collected from a real service: its users are sellers of small and medium shops on Yandex.Market, Ozon, Wildberries. Peak activity from 6:00 PM–10:00 PM shows this is a tool for small business owners, not for content departments of major marketplaces: people generate cards in their free time, after their main work.

This determines pricing requirements (a seller pays $5–10, not $100) and speed (generation in 3–5 seconds, not a minute). The seller's time is expensive.

What This Means

AI for marketplace content works not as full automation, but as an assistant in work. Sellers are willing to delegate the first draft to a neural network, but the final word remains theirs. This explains growing demand for tools with smart editors — sellers are willing to pay not for perfect generation, but for convenience and editing speed.