Most vendor case studies are written to end an argument. This one is written to start a better one, so it includes the parts a case study normally leaves out.
VARUS is a Ukrainian grocery chain operated by Omega LLC. It opened its first store in Dnipro in 2003 and now runs more than 120 supermarkets across several formats, alongside a national online grocery business.
What shipped, and when
Roughly a year of development, starting with internal use inside VARUS. A closed pilot ran through the summer of 2025. The public launch was on 14 April 2026, branded as the VARUS AI Helper and powered by Neomi.
How it was launched, which matters more than it sounds
Web only. Not promoted. Deliberately kept out of the mobile app. Held to a smaller sample for the first month so VARUS could watch real behaviour before opening it up.
Every number below came from a channel that was being actively hidden. That cuts both ways, and we say so further down, but it is the condition under which the results were produced and it belongs next to them.
What we measured
The pilot tracked the shape of the session, not just its outcome, because the argument for an AI storefront is about how the work gets done rather than a single conversion event.
Add to cart. Around 88% of sessions put items in the cart.
Session to order. Around 15% of sessions ended in a confirmed order. The store's overall website visitor-to-order rate is around 1%.
Time to a finished cart. A median of about 2.5 minutes among purchasing sessions, with some completing in under 90 seconds. A traditional online grocery session at VARUS runs 20 to 30 minutes.
Prompts to checkout. A median of roughly four exchanges from the first request to a completed cart.
Items per step. A large majority of steps added seven or more items at once, which is the assembly effect the product is built for.
Voice. Voice sessions showed a checkout rate around 25%, against roughly 15% across all sessions.
On growth, the Retail Association of Ukraine reported that after the April launch, the number of users grew by roughly 260% month on month, and that in May the turnover of orders assembled through the AI Helper grew by about 250% over the previous month.
The pattern across those numbers is more informative than any one of them. Sessions were short, dense, and ended in a cart. That is what you would expect if the assembly work moved from the shopper to the storefront, and it is the thing worth testing on someone else's traffic.
What shoppers actually asked for
This is the part we find most persuasive, and it is not a number.
Pick something for my husband so he can easily cook balanced, varied meals during the week. I already have side dishes.
Write me a product list for 6 days with a $50 budget. Minimum cooking, no porridge, no milk, no fruit.
Everything for a balanced diet, but without carbs because I already have grains. Needed for the week.
Romantic dinner for two.
For a men's group watching football.
None of those is a product search. They are constraints, occasions and budgets, handed over whole.
Shoppers did not use the AI Helper as a better search bar. They used it to delegate the decision.
That matters against the baseline it replaced. More than 45% of VARUS online purchases previously came through the catalogue, where a shopper finds items one at a time.
What surprised us
Two things did not match the internal expectation.
The first was how much of the value came from multi-item steps rather than clever single-item matching. We built for the hard matching problem and the wins came from volume: a shopper describing a week of meals and receiving most of it at once.
The second was how much richer inputs mattered. Sessions that started from something other than typed text, including voice, images and recipe links, converted well above their share of sessions.
What we are testing next
The first phase answered whether shoppers would use this at all. They will. The next phase is about how much value an AI-assisted session can create, which comes down to two numbers: average order value and conversion to a confirmed order.
Discoverability. The first month deliberately hid the channel. Placing it closer to high-intent moments, including search, recipes, catalogue browsing and cart building, is the obvious lever and an untested one.
Complete carts rather than partial lists. Weekly planning, recipe bundles, family meals and occasion baskets, tested against AOV.
Basket quality. Guiding recommendations toward what is actually in stock, better substitutions and sensible alternatives.
Voice and mobile. While voice inputs were used in 7% of scenarios, they ended up in having 25% conversion compared to 15% with text inputs. A great insight to promote dictating AI what to purchase.