Neomi Co-Founders on Person-First Grocery Carts: Where AI Should Draw the Line on Promoted Products

Foods for Health, Not for Shelves

Agentic commerce is loosening the grocery aisle's grip on the shopper

Neomi co-founders Dmytro Lylyk and Vladyslav Mehera argue AI should start with the person, not the product, building baskets from health goals and context rather than catalog searches.

thenextweb.com

Shoppers can order dinner from a phone and reorder a saved basket in a tap, yet the job of deciding what to buy is still there. An article The Next Web (TNW) published on September 14, 2026 opens with this paradox of online grocery.

It features Dmytro Lylyk and Vladyslav Mehera, co-founders of Neomi, a grocery shopping AI assistant. Lylyk sums up the company's philosophy as "foods for health, not for shelves," taking issue with how retailers and brands have optimized for being noticed through packaging and shelf placement.

According to Lylyk, shoppers are stuck choosing between searching a catalog and repeating past orders, and both reinforce a product-led system. Neomi says it instead treats health information, dietary restrictions, preferences and personal goals as signals that determine what belongs in a basket. The claim is about moving the starting point from the product to the person. This article checks it against public information and survey data, and looks closely at the handling of promoted products, which Lylyk called a line that must not be crossed.

What Public Information Confirms About Neomi

According to a June 2025 interview by Ukrainian tech outlet dev.ua, Neomi is a Ukrainian startup. The founders bootstrapped development and later received grants from the Ukrainian Startup Fund in fall 2024 and from the European Union in spring 2025. In June 2025, a pilot started on the website of supermarket chain VARUS, described as the company's first commercial partnership.

Neomi's current website presents it as a white-label AI storefront for grocers. It runs alongside a retailer's existing online store, reads the live catalog, stock and prices to build carts, and leaves checkout and customer data with the retailer.

Why Wine Ended Up in a Romantic Dinner Cart

Lylyk offers a concrete example. A Neomi user who selected a "Romantic Dinner for Two" experience found wine in the basket without having asked for it. The wine was not a requested product; its relevance came from the context of the occasion.

Someone picturing a dinner knows they need a meal but does not necessarily know every product it requires. Search only accepts product names the shopper already knows, while a description of the occasion can surface products they have not thought of yet. That is the core of person-first design.

Shoppers did not start out using it that way. According to Mehera, users initially treated Neomi like a search engine, asking for milk and then moving on to the next single item. Neomi began nudging them to describe their needs as a whole: what they eat during a week, a preferred cuisine, a dietary goal. Mehera sees this as a learned behavior that will take time to change.

Real requests appear in Neomi's case study of the VARUS pilot. "Write me a product list for 6 days with a $50 budget. Minimum cooking, no porridge, no milk, no fruit." "For a men's group watching football." None of these is a product name. They are constraints, occasions and budgets.

According to the case study, the median path from first request to finished cart took roughly four exchanges, and a large majority of steps added seven or more items at once. Neomi concludes that shoppers used the AI not as a search bar but as something to hand the decision to. Neomi's website adds that when a request is too vague, the assistant replies with a quick in-chat form, which reads as a way for the provider to lighten the learning burden on shoppers.

Demand-Side Data Supports the Person-First Approach

The McKinsey report TNW cites, The State of Grocery North America 2026, draws on surveys of nearly 5,000 consumers and more than 40 grocery executives in the United States and Canada. 47% of consumers prioritize specific functional benefits such as high protein or low sugar over general "healthy" claims, and nearly 55% (53% in the footnote) say the retailer that best supports their wellness offers personalized nutrition recommendations.

On the retailer side, grocers expect the share of fully personalized promotions to rise from 35% today to 55% within two to three years. Nearly half of grocers expect AI agents to assist with a third or more of transactions within five years, and major grocers already see more than 20% of online orders come from recommended or preassembled baskets.

Read the VARUS Numbers Together With Their Conditions

According to the case study, after the public launch on April 14, 2026, around 88% of sessions put items in the cart and around 15% ended in a confirmed order. Among purchasing sessions, the median time to a finished cart was about 2.5 minutes, against 20 to 30 minutes for a traditional VARUS online grocery session.

The launch conditions matter, though. The channel was web only, not promoted, deliberately kept out of the mobile app, and limited to a smaller sample for the first month. Neomi itself states that every number came from a channel that was being actively hidden, and that this condition cuts both ways. The roughly 15% order rate is shown next to the site's overall visitor-to-order rate of about 1%, but the two use different denominators.

The website also frames these as directional results from one public pilot. Effects on average order value, repeat usage, substitution rates and margins were not disclosed in anything we could check. Whether other retailers would see similar results has not yet been tested.

Promoted Products Only When They Fit Intent: Who Draws That Line?

The most pointed remark in the TNW article concerns retail media. Lylyk noted that AI could push products into the cart simply because they are promoted, even when they fit neither the cart nor the shopper's expectations, and called that "the borderline which must not be crossed." Promoted products, in his view, belong in the cart only when they genuinely suit the shopper's intent.

What makes this interesting is that the line sits inside Neomi's own pitch to grocers. Its website promotes the ability to place a grocer's priority products and weekly-ad items straight into the cart at the moment of decision, and the FAQ lists this as a reason Neomi pays for itself for grocers. It adds that shoppers always see those items and can remove them.

In other words, Neomi does not reject putting promoted products in the cart. They go in on three conditions: they fit the intent, they are visible, and they can be removed. The open question is who judges "fits the intent" and by what criteria, and for now that judgment sits inside the provider and the retailer that deploys it.

This matters for the whole industry. In the same McKinsey report, grocery executives say retail media contributes 3 to 10% of their total profit, and the report notes that as discovery shifts to conversational assistants, retail media will need to evolve from placement-based advertising to influence within algorithmic decisioning. Writing in California Management Review, Paul F. Accornero compared delegating shopping to AI agents to the principal-agent problem, arguing that an agent could favor partners that pay for placement in ways consumers would struggle to verify.

With outside AI agents, research suggests ad labels may even backfire. When a Columbia Business School team had GPT-4.1, Claude Sonnet 4 and Gemini 2.5 Flash shop on a mock e-commerce site, products marked "Sponsored" saw their selection rates drop, while those labeled "Overall Pick" gained. How promotions work inside a retailer's own assistant and how outside agents evaluate its products need to be designed separately. Attempts to put ad slots inside the conversation, such as Topsort's Sponsored Prompts, are moving in parallel.

Consumer sentiment cannot be ignored either. In The Shelby Report's coverage of a McKinsey webinar, willingness to use AI stood at 51% for assisted product search but fell to 20% for fully automatic ordering with no user review. McKinsey lists what would raise interest among non-users: turning the feature off easily, overriding its suggestions, and reaching human support if something goes wrong. Keeping promoted items "always visible and easy to remove" is in line with consumers' reluctance to give up control.

Instacart and DoorDash Are Moving the Same Way

Building carts from occasions and meal plans is becoming standard among larger players. Instacart's Clementine, launched across North America on September 9, turns requests like "high-protein easy dinners for two" into carts in seconds, and Instacart points to inventory signals from nearly 100,000 stores as its differentiator. Where Instacart runs a marketplace, Neomi pitches an AI storefront that sits on a retailer's existing systems and is owned by the retailer.

DoorDash launched Ask DoorDash, which builds carts from recipe photos, in June, and US grocer Schnucks has announced a shopping assistant built around health goals with VitalityIP. Albertsons reported that conversational AI lifted average order value by about 10% for conversational search and about 26% for its full assistant, but its first-quarter gross margin declined, a reminder that higher order values do not automatically mean higher profit.

What Grocery E-commerce Teams Should Prepare Now

Product data is the easiest place to start. Neomi says that when a retailer's catalog is missing dietary data, it fills the gap with its own tags. The feature is effectively designed for catalogs whose dietary attributes are incomplete.

If products can be retrieved by occasions such as "dinner for two" or "watching football," and by attributes such as "high protein" or "dairy-free," they are more likely to be picked up by both a retailer's own assistant and outside agents. Chris Selland told PYMNTS that in a world where agents choose products, structured product data, fulfillment information, trust signals and availability shape recommendations. Our analysis of Walmart Sparky recommendations likewise found that the retailer product listing accounted for only 27% of the citation surface.

Inventory and substitutions are just as essential. A cart built from an occasion falls apart if the suggested products cannot be bought. Neomi says it filters out out-of-stock items before suggesting them, and lists stock-aware recommendations and better substitutions among the next things to test at VARUS.

Health information needs clear boundaries as well. Neomi's FAQ states plainly that its handling of dietary needs "filters the catalog rather than making a medical guarantee." When carts are built around allergies or medical conditions, accurate product information and a flow that prompts shoppers to confirm remain the retailer's responsibility.

Then there are the rules for retail media. If promoted products are inserted into AI-built carts, retailers need to decide, alongside their supplier agreements, which products may be added for which intents, how they are labeled, and how removal rates are measured. The line Lylyk describes will be drawn where the logic of selling ads meets the logic of an AI that speaks for the shopper's intent.

The two founders advise retailers to test intent-based shopping with their own shoppers and inventory now, while the model is still taking shape. They envision shoppers expressing food needs through retailer websites, AI assistants, recipe platforms and diet apps, with technology translating those needs into products. Holding data that can translate a request from any of those entry points into your own products is how to prepare. For the other major thread in grocery agents, predictive replenishment, see our overview of grocery and agentic commerce.

Conclusion

What Neomi's co-founders describe is a shift in grocery AI from a tool for finding products to a tool that translates a person's circumstances into a cart. The wine that appeared in a romantic dinner basket shows how describing an occasion can surface products search would never reach.

At the same time, the published results come from a single pilot run through a hidden channel, and the criteria behind adding promoted products "only when they fit intent" are not yet visible from the outside. The things to watch next are results from retailers beyond VARUS, the effect on average order value and margins, and how far retailers and platforms will disclose to shoppers the rules for which promoted products enter AI-built carts.