Since Aug. 18, Carrefour Belgium shoppers have been able to reach the grocer directly from ChatGPT. They can ask for recipe ideas, paste in a shopping list, pick from the suggested products and send them to their Carrefour cart. Agentic commerce is taking concrete shape in grocery.
The experience was built and is run by Mealz, a 20-person startup based in Lille, France. The company has kept a low profile despite integrations with several major French grocers (more on that below).
In practice, a user can paste a shopping list into ChatGPT and ask Mealz to find the matching items at Carrefour Belgium. They can drop the products that don't fit, then send the rest to the cart.
Users can also start from a much looser request – "suggest five dinners for a family of four" – and let the assistant come up with recipes.
The app goes beyond a chatbot. Mealz rebuilt some of the standard e-commerce moves inside ChatGPT, with interactive modules embedded in the conversation that let shoppers select, uncheck, compare and add to cart. "Even inside an AI assistant, conversation alone won't deliver a complete grocery experience," said César Tonnoir, co-founder of Mealz.
The journey stays tied to Carrefour Belgium. A shopper who leaves ChatGPT to check out on the grocer's site finds the cart already built. OpenAI approved the app after several months of review. It puts recommendation and transaction in the same interface.
Tonnoir doesn't see the interface as the main barrier to entry. "Anyone can build a chatbot for their site today," he said. The hard part is turning an intent or a recipe into relevant products that are actually in stock.
That is Mealz's original business. The company has spent seven years building its matching engine. For lasagna, the engine has to pick the right ground beef, in the right pack size and in line with the shopper's preferences, rather than any product with "steak" in its name, Tonnoir said. "Keep in mind that with an engine that's 90% accurate, a 10-ingredient recipe statistically produces one error. Ours is at 99%."
Mealz also made an editorial call. The assistant doesn't let ChatGPT make up recipes. It draws on a base of about 1,000 recipes created and tested by people, then matches each ingredient to the right product. The guardrail is meant to limit incoherent recipes, repetition and absurd substitutions. Those wear down the experience fast once shoppers move from a demo to weekly use.
For Mealz, making the catalog readable by an LLM is only part of the job. The model also needs enough domain-specific tools to complete a full task without errors. "The app was designed to be agnostic. Its core can easily be adapted to other AI assistants, and also built directly into retailers' sites and apps," Tonnoir added.
After ChatGPT, Mealz builds Coopérative U's AI assistant
Tonnoir calls this second route strategic. Usage of ChatGPT, Gemini and Claude is growing, but he expects the vast majority of grocery shopping to keep running through retailers' own interfaces.
Mealz is building a white-label AI assistant with French grocery cooperative Coopérative U, designed to sit inside the retailer's own digital properties. The project is underway. No official launch date has been announced. It resembles, with obvious differences, what retailers such as Carrefour (with Hopla), Walmart and Albertsons have already rolled out.
The assistant will use the same building blocks as the Carrefour Belgium app: understanding a complex request, suggesting recipes or products, building a list and adding it to the cart. Running it on-site changes the math for the retailer.
The first difference is that the grocer keeps control of the customer relationship. "Mealz operates as a white-label provider under a data processing agreement that lets the retailer control what information is sent to the model," Tonnoir said.
The second is visibility. Unlike a third-party app inside ChatGPT, an on-site assistant gives access to every request made in the retailer's environment. The retailer sees what was bought and also the intent expressed upstream: "five vegetarian dinners for four," "the best deals in my store" or "a holiday meal for under €80."
That lets the retailer build a real learning curve on shopper behavior. In agentic commerce, longer phrases and questions replace keywords. Those requests can also reveal more about customers, from household size to average budget and dietary restrictions. That only works if the retailer builds a consent and governance framework that matches how sensitive the data is.
On-site integration carries a cost that a ChatGPT app largely avoids: every request consumes compute. Mealz plans to bill on usage. It is trying to contain costs by running an off-the-shelf LLM on its own infrastructure.
The plan is to open the assistant to a fraction of users first, track adoption and revenue, then widen the rollout. "We're going to offer our current partners a free six-month pilot to help them get started," Tonnoir said. The pilot should also show them how to monetize the new engagement tool.
Mealz's legacy business offers a preview of what advertising in these assistants could look like. On the sites of French grocers Intermarché, Monoprix and U, the company already runs recipes whose ingredients can all be added to the cart in one click. Brands get two ad formats. They can sponsor a recipe or attach their product to an ingredient. When a recipe calls for spreadable cheese, for instance, the advertiser's product can be suggested automatically.
The second format accounts for 70% to 90% of Mealz's media sales depending on the period, the company says. It is sold on performance. The budget is drawn down as products are actually added through recipes, against a ROAS commitment set upfront. Mealz says it only counts sales directly tied to that add-to-cart, not later purchases by shoppers who simply saw the recipe.
"In an AI assistant, the first ad opportunity will look fairly similar: when a user asks for a recipe, a shopping list or a product, a campaign can favor a sponsored item, as long as it stays relevant," Tonnoir said. Conversational interfaces go further. The future unit of sale could be an intent or an entire theme: "deals in my store," "back to school," "hosting six people," "Christmas dinner" or "a week of high-protein meals."
Advertising would then be bought around a consumption moment, not just a placement or a keyword. Brands could step in at the inspiration stage, while the assistant builds its recommendation and the cart that goes with it. The idea, Tonnoir said, is to "sell inspiration in context rather than space on a site."
AI assistants are often cast as a threat to on-site retail media. If consumers plan purchases in ChatGPT, they visit fewer retailer pages and see fewer banners and sponsored products. Mealz's launch at Carrefour Belgium points to a more nuanced outcome. Traffic can leave the retailer's site while its catalog, cart, ad formats and sales measurement stay accessible from ChatGPT.
For retailers, the fight is over the entry point. A presence in the big assistants lets them reach those users and test new journeys cheaply. An assistant built into their own site or app gives them more control over the experience, the data collected and the ad formats on offer.
Mealz is moving on both fronts, with a Carrefour Belgium app in ChatGPT and an assistant built directly for U. The two setups will let the company compare usage, costs and monetization potential. In both, retail media goes beyond a sponsored product in a digital aisle. It kicks in as soon as the shopper states a need and follows the cart as it fills. The open question is how to make that ad influence clear enough that the assistant doesn't turn into a salesman in disguise.
