Building a shop with AI: what actually exists

What can AI genuinely do for an online shop?

Four families of task, with one thing in common: they're formatting tasks, where the context is supplied and the result is checkable at a glance.

Recognising, first. From a photo, a current model identifies an object, a material, a trade without difficulty. That's what makes it possible to build the right shop: a potter and a caterer don't need the same screens.

Writing, next. Brand name, tagline, product description, about text, a standard answer to a frequent question, translation: this is the ground where the gap between AI and a blank page is widest. Composing too: making five colours agree, choosing two typefaces that work together, producing a simple logo and a cover — template work, done fast and properly.

Filing, finally, and this is the invisible part that matters most later: classifying a product into a category, extracting its attributes, producing a structured page that search engines and shopping assistants can read.

• Recognising a trade and a product from a photo.
• Writing titles, descriptions, about texts, standard answers, translations.
• Composing an identity: palette, typefaces, logo, cover.
• Structuring a catalogue: categories, attributes, machine-readable pages.
• Retouching and generating staged visuals from a real photo.

What's marketing rather than technology?

Four promises come up everywhere, and none holds. "A profitable shop in thirty seconds": the shop can indeed exist in a few minutes, profitability depends on what you sell and who buys it — two variables no model controls. "AI finds your customers": it writes ads and suggests audiences, it has neither budget nor audience to lend you.

"A turnkey shop that sells by itself" is the most expensive one. It usually accompanies an offer to resell generic products: a thousand identical shops, generated the same day, with the same catalogue and the same copy. The problem isn't the AI, it's that there's nothing to sell that nobody else is selling.

"AI optimises your search visibility automatically", finally, conflates two things: producing well-structured pages, which it does well, and earning authority in the eyes of search engines, which takes time and sources that talk about you.

A simple marker: a credible promise is about PRODUCTION — copy, images, structure. A doubtful promise is about the RESULT — sales, traffic, ranking, profitability. Nobody can guarantee the second.

What won't AI do for you?

The price, first. A model knows neither your production cost, nor your margin, nor the time a piece takes you, nor what customers in your area will pay. A generated estimate is wrong silently, and you carry the gap. That's exactly why Mopsee forbids its AI from inventing an amount: with no price from you, the offer ships as "on request" rather than at a plausible, wrong one.

The facts about your product, next. A model with only a photo will happily describe a twenty-centimetre glazed stoneware piece — and be wrong about both the material and the size. Everything measurable has to come from you: dimensions, weight, composition, duration, capacity, exact contents.

And finally everything that binds you: your terms of sale, your real lead times, your returns policy, your tax obligations. An AI produces text that looks like terms and conditions; whether it fits your country and your activity isn't guaranteed, and it's your signature underneath.

• The price of your offers — a commercial decision, never a generation.
• Measurable data: dimensions, materials, weight, durations, capacities.
• Your lead times and returns policy, which depend on your real production.
• Your legal and tax obligations, which depend on your country.
• The audience: no AI lends you customers.

How do you judge an AI shop-building tool?

Five questions are enough, and they're asked before testing anything.

Is everything editable afterwards? A tool that generates and locks produces a shop like a thousand others, with no way out. Does the tool refuse to invent what it doesn't know? A system that fills every field at all costs mostly fills the wrong ones. Does going live require your explicit consent? Nothing should become public unless you say so.

Who owns the address and the customers? That's the question that decides what you're left with if you leave. And finally: what does the tool produce for machines? In 2026 a growing share of buyers go through assistants that read structured data and product feeds, not decorated pages. A tool producing none of that makes you invisible to them, whatever the quality of its copy.

• Everything stays editable, including what the AI wrote.
• The tool leaves a field empty rather than inventing a value.
• Publishing is an explicit act, never automatic.
• The address, the copy, the photos and the customers stay yours.
• Structured data and product feeds are generated out of the box.

Where does Mopsee sit in this landscape?

On a precise case: getting started. The photo comes in, an understanding step decides the trade and therefore the kind of shop, then identity and first offer are generated in parallel and appear as they land. You correct by asking, at any moment.

Three guardrails live in the product rather than in a sales pitch. The AI has no right to invent a price — with no amount from you, the offer stays "on request". Typefaces come from a short pre-vetted list rather than the whole web, because a badly matched typeface damages a brand for a long time. And the shop stays private until you press the go-live button yourself.

Finally, every shop automatically publishes what machines read: structured data on every product page, a sitemap, the catalogue in plain text for assistants, and product feeds for shopping engines. That's the work nobody wants to do by hand, and it increasingly decides whether a product is offered or skipped.

What it doesn't do deserves saying just as plainly: it doesn't find your customers, it doesn't set your prices, and it doesn't replace an offer someone actually wants to buy.