Selling through AI, redirecting a product strategy on proof rather than conviction

A legitimate request, unverified foundations

In spring 2025, the market sends converging signals: the announced tie-up between Shopify and ChatGPT, Copilot’s merchant program, announcements around agentic commerce and the Model Context Protocol. Conversational assistants stop being a demo gadget and become an entry point to purchase. The request arrives with its priority: demonstrate the feasibility of a complete purchase journey through AI, from dialogue to payment, so that the products of PrestaShop Enterprise merchants are findable and buyable from LLM interfaces.

That request was legitimate, ambitious, aligned with the market. It had one flaw: no one had verified what it rested on. Executing meant investing on unknown foundations. Refusing meant opposing a stated priority, with no hierarchical authority and no data to argue with. The discovery was run to get out of that dilemma through proof.

The mapping weakened the starting hypothesis

Rather than rushing to a prototype, the first cycle mapped the system: the buyer who delegates their search to an assistant, the merchant’s reality, the ecosystem of players already moving. This work was not documentation, it was de-risking. Each map asked the same question from a different angle: what has to be true for an AI purchase to actually work?

This mapping, led end to end, surfaced the uncomfortable answer. The initial hypothesis, ‘an end-to-end technical purchase journey will generate new sales for merchants’, rested on two pillars no one had verified: the quality of the product catalog and the performance of the merchant’s site.

A few days of testing rather than a battle of opinions

A team session turned that doubt into an action plan. The ambition was collectively narrowed: make products visible in external LLM interfaces, in an acquisition logic. The idea of an AI embedded directly on the store, which answers a different need, was set aside. One central, binary hypothesis remained: the products visible in LLMs today are visible because they flow through a structured feed, Google Merchant Center type.

Rather than debating it, the hypothesis was put through a targeted micro-experiment, run by tech. The verdict came within days: well-configured catalog feeds are enough to generate visibility, even without paid campaigns. The channel already existed. What was missing was the data.

Refuse the request, build the fuel before the road

The decision was prepared, not imposed. Three hypotheses were formalized, each with its underlying problem: the end-to-end purchase journey, the initial request; an AI catalog preparer, to diagnose product data quality and optimize it with AI assistance; an optimized landing experience, to take care of the arrival after the handover from the assistant. Two concepts were prototyped with Figma Make, to make the options discussable before any technical investment.

The dilemma was set. Executing the request meant building a road over a void: the journey rested on the two weakened pillars. Optimizing the landing page meant polishing a stretch of the road without touching the fuel. The reasoning came down to one image: product data is the fuel, the purchase journey is the road. A fast page displaying data that is poor or unreadable for AI does not convert. Solving data quality at the source creates the most durable value and lays the foundation that makes the next innovations possible.

The decision was made collectively at the end of the cycle: refuse the request as it had been formulated, prioritize the AI catalog preparer, open a second dedicated discovery cycle. Its consequences went beyond the choice of a track. The editorial quality of the catalog became a product subject in its own right, on par with the technical feeds. And the team’s vocabulary shifted from keyword-centric SEO to GEO, centered on conversational intents: a product page written for AI does not say ‘red sneakers size 36’, it tells ‘a pair of red sneakers for a 13-year-old girl who plays outside’. The usage situation becomes the unit of meaning.

A paused discovery is not a lost discovery

The second cycle was framed: a merchant empathy map centered on the catalog, identified touchpoints, including an AI readiness dashboard, an audit report and assisted writing help. Then the subject was paused before user testing. The second cycle’s hypotheses were not validated, the Go/No Go was not called. No measured result can be claimed here, and this case study claims none.

What remains comes down to two gains. A confirmed redirection: the effort no longer targets a spectacular journey but the foundation that would make it possible. And a reading of the problem that every market announcement reinforces rather than expires: when the subject reopens, the question will no longer be ‘should we sell through AI’ but ‘do our merchants have the data to exist in this channel’. The cycle’s bottom line fits in one sentence: the problem was not at the surface, it was at the source.