What a 3x Conversion Gap Reveals — Walmart's ChatGPT Experiment and Emart's Next Move

Get a product recommendation inside ChatGPT, and complete the purchase right there. "Instant Checkout," unveiled by Walmart and OpenAI in October 2025, was one of the first real implementations of that vision. But a few months later, the numbers Walmart itself disclosed were unexpected. The conversion rate for products purchased directly inside ChatGPT came in three times lower than when the same recommended products were purchased after the customer clicked through to Walmart's website. Removing one click from checkout ended up selling less, not more. Understanding why this happened — and why Walmart's own agent posted the opposite result over the same period — reveals where the real battleground in agentic commerce actually lies. Shinsegae Group's experiment, run through Emart, is aimed squarely at that same battleground.

What Walmart Is Doing

Since summer 2025, Walmart has organized its AI strategy around four "super agents": Sparky, a shopping agent for customers; Associate Agent, which assists employee workflows; Marty, aimed at suppliers and sellers to manage listings, orders, and advertising; and Developer Agent, which supports internal developers' build, test, and deployment work. Rather than letting individual departments spin up their own separate chatbots, Walmart chose to orchestrate multiple specialized agents underneath four entry points organized by user type.

Of these, Sparky is furthest along. Introduced into the Walmart app in June 2025, Sparky handles product search, review summarization, and contextual recommendations, and has since expanded into in-store repeat purchasing, meal planning, and Spanish-language support. Walmart says it uses not just general-purpose LLMs but also retail-specific models trained on its own distribution data — a bid to make the most of its data assets in the fight for the customer touchpoint.

Key Insight #1 — The Real Cause of Low Conversion Wasn't "Recommendation Distrust"

Walmart didn't stop there. Last October 14, it formally announced its collaboration with OpenAI: a way to buy a product recommended inside ChatGPT without ever leaving for a separate shopping site, completing checkout on the spot. The service rolled out to a subset of users starting in November, covering around 200,000 products.

The results were underwhelming. Daniel Danker, Walmart's head of AI acceleration, product, and design, disclosed that the conversion rate for products purchased directly inside ChatGPT came in at roughly a third of what it was when customers were routed to the website. Taken at face value, that result invites an easy interpretation: "consumers didn't trust the AI's recommendation." But the cause Walmart actually pointed to was something else entirely.

Instant Checkout was fundamentally designed for single-item purchases. There was a risk that every product a customer checked out through ChatGPT would trigger a separate order and separate delivery, and it was difficult to combine such a purchase with items already sitting in a customer's Walmart cart. Perks like Walmart+ shipping benefits, minimum order thresholds, bundled-delivery terms, and loyalty points simply didn't function properly within this checkout path. For a customer trying to buy several household staples at once and have them delivered together, "checkout inside the chat" turned out to be an inconvenience rather than a convenience. In the end, the real story behind the low conversion rate wasn't a problem with the AI model's recommendation quality — it was a checkout workflow disconnected from the rest of the shopping journey.

Key Insight #2 — OpenAI Reached the Same Conclusion and Changed Course

Walmart wasn't the only one to acknowledge this problem. In March 2026, OpenAI itself admitted that the Instant Checkout model "didn't offer the flexibility it was aiming for," and shifted its approach. Instead of completing checkout entirely inside ChatGPT, it expanded its "Agentic Commerce Protocol" so ChatGPT focuses on discovering and surfacing products, while actual checkout stays under each retailer's own control within their own systems. Major retailers including Target, Sephora, Nordstrom, Lowe's, Best Buy, Home Depot, and Wayfair moved to this model.

Walmart introduced its own response around the same time. It embedded Sparky as its own app inside ChatGPT, so that even when a customer discovers a product through ChatGPT, login, cart sync, checkout, and membership benefits are all handled within an environment Walmart controls. As Walmart put it: "No matter where the journey starts, customers get the same Walmart assortment, value, and delivery speed." Rather than pulling back from external platforms, the direction is to bring its own transaction environment directly inside them.

Key Insight #3 — Over the Same Period, the Self-Controlled Agent Grew in the Opposite Direction

This is where an interesting contrast emerges. While Instant Checkout was underperforming, Sparky — the agent Walmart controls entirely on its own — was tracing a completely different trajectory. In fiscal Q1 2026 (ending late April), GMV flowing through Sparky rose 150% quarter-over-quarter, and the number of purchases made through Sparky more than quadrupled. Weekly active users (WAU) more than doubled within a single quarter, and the average order value of Sparky users was roughly 35% higher than non-users. Roughly half of Walmart app users had already tried Sparky.

Why did the results diverge so sharply for the same category of AI agent? The decisive difference between ChatGPT's Instant Checkout and Sparky wasn't model performance — it was who controlled the transaction workflow. Sparky is fully connected from the start to a customer's Walmart account, existing cart, membership benefits, delivery options, and inventory data. Instant Checkout, by contrast, moved only the checkout button into ChatGPT, with none of those connections in place. Ultimately, what determined performance wasn't how smart the agent was, but how seamlessly it was connected to the actual order, payment, and delivery systems.

Key Insight #4 — The Google Partnership Follows the Same Logic

Walmart also partnered with Google on January 11, 2026. The deal lets Gemini users discover Walmart and Sam's Club products in the course of conversation and complete checkout inside the chatbot via Google's new "Universal Commerce Protocol." Linking a Walmart and Gemini account enables personalized recommendations based on combined online and offline purchase history, and products found in Gemini can be merged with an existing Walmart cart.

What's notable is that this structure directly reflects the lessons from the ChatGPT failure. "Discovery happens on the external platform; account, cart, checkout, and delivery stay within Walmart's systems" — that principle now applies equally to the redesigned OpenAI arrangement and the new Google partnership. Instant Checkout's failure wasn't a one-off misstep — it's becoming a standard lesson the whole industry is converging on.

Business Impact — Agents Are Becoming a Revenue Model, Not Just a Touchpoint

The payoff from Walmart's agent strategy isn't limited to the shopping agent. Marty, the supplier- and advertiser-facing agent, has been running in beta since January 2026 inside Walmart Connect, letting advertisers build and optimize sponsored search campaigns conversationally, with a company-wide rollout to all advertisers planned within the year. 97% of advertiser queries are unique, reflecting how individualized each advertiser's consultation is. In the same quarter, Walmart's ad revenue grew 37% year-over-year, and its retail media business overall grew 33% — six times the company's overall revenue growth rate.

In other words, for Walmart, the agent isn't simply a device for improving customer experience — it's becoming a dual engine generating revenue from both customers (Sparky) and advertisers/suppliers (Marty) at once. In the fight over agentic commerce, the question of "who controls the customer touchpoint" is really closer to "who controls the transaction and advertising systems well enough to turn that touchpoint into revenue."

The Domestic Case — Where Does Emart Stand Right Now?

This isn't a story that only applies elsewhere. Shinsegae Group, through Emart, is moving in a similar direction, and the pace has visibly picked up over the past few months. That said, it would be an overstatement to say "Emart has already commercialized a customer-facing AI shopping agent the way Walmart has." Laying out what's actually confirmed so far: Emart has prioritized putting less flashy internal-facing AI to work on the ground before its customer-facing shopping agent, which remains at the planning stage and has already gone through a partner reshuffle.

The plan: end-to-end AI commerce from search to checkout — and a partner swap, twice over

On April 6, 2026, Shinsegae Group announced it had signed an MOU with OpenAI for AI commerce business cooperation. Starting with Emart, the vision was to build, by 2027, a fully integrated AI commerce experience where a customer could type "put together tomorrow's family dinner" into a chat window and have the AI assemble the needed products, fill the cart based on purchase history and preferences, and complete checkout and delivery scheduling. It's a picture that closely resembles the meal-planning, product-assembly, and purchase-execution flow Walmart's Sparky is aiming for.

But just 11 days after that announcement, on April 17, Shinsegae Group announced it was halting discussions with OpenAI and concentrating its efforts on a partnership with US-based Reflection AI instead. Alongside a major project to build a 250MW AI data center domestically, the group reorganized its strategy around applying AI to six areas: sourcing, ordering, pricing, logistics, inventory management, and customer management. Emart was assigned a leading role within the group for this project, but the specific launch timeline and final technology partner for the customer-facing shopping agent have not yet been announced. The "we're going toward AI commerce" strategic direction has held, but the implementation approach and partner changed once within less than two months.

What's already running: unglamorous but genuinely operational on-the-ground AI

What took hold before the customer-facing shopping agent was, ironically, the less visible internal-facing AI. Emart buyers are using Microsoft Copilot to build their own task-specific agents — in the 2025 Copilot Agentathon, an "unpaid vendor-listing alert agent" built by Emart's buyer team won first place. Notably, it was frontline buyers, not developers, who designed the agent by connecting it directly to their own operational data. AI agents are already applied to repetitive tasks like fresh-food price forecasting and trade-area analysis, with plans to extend into food labeling, finance, and promotional operations.

There's already AI running at the customer touchpoint too. Every Emart and Traders location runs a unified digital consultation platform combining AI chatbot, agent chat, and phone support, and the AI chatbot handles more than half of all customer inquiries. At checkout, an AI camera system that detects scan omissions or checkout errors in real time is running on 923 units across 58 stores nationwide as of the end of April 2026.

The most concrete performance figures, though, come not from Emart itself but from its convenience-store affiliate, Emart24. Its "AI product recommendation service," introduced in January 2024, analyzes store-level sales data to find ten stores with similar characteristics, then recommends to franchise owners products that sell well at those comparable stores but aren't yet carried at their own. In a direct-operated-store test across 17 locations from August to November 2023, 90% of recommended products sold out and were reordered. That result was possible only by restructuring more than 6 billion daily transactions and over 80,000 product records down to the individual store level.

Compared with Walmart

To sum up: Walmart has already brought its customer-facing agent (Sparky) to market, gone through a failed experiment with checkout on an external platform, and is now adjusting its next strategy. Emart and Shinsegae Group, by contrast, are first embedding and validating AI in internal, consultation, and checkout use cases on the ground, while the end-to-end agent that would let customers hand off everything from search to checkout remains at the planning and partner-reshuffling stage. The sequence differs, but the direction is the same. So far, Emart has chosen the path of automating narrow tasks where the data is clear and results are easy to validate, ahead of a flashy customer-facing chatbot, and building up success stories from there.

Practical Implications

• Commerce/UX strategy teams: The formula "fewer clicks equals higher conversion" doesn't always hold in agentic commerce. Especially in categories where customers buy multiple items at once, reflect in your design standards that consistency across cart, delivery, and benefits matters more for conversion than trimming checkout steps
• Data/API teams: Agent competitiveness comes not from model performance but from system connectivity — real-time inventory, pricing, delivery availability, and membership benefits. When partnering with an external AI platform, check first whether the integration syncs bidirectionally across cart, account, and order data, not just product exposure
• External platform partnership teams: External AI platforms like ChatGPT and Gemini are valuable as discovery channels for new customers, but handing over checkout, fulfillment, and membership entirely means losing control. Build the hybrid principle — "discovery externally, transactions on our own systems" — explicitly into partnership contract terms
• Advertising/retail media teams: Following Walmart's Marty example, it's worth looking beyond customer-facing use cases for areas where AI agents could serve advertisers and suppliers as well. Starting with repetitive, standardized tasks like campaign setup and billing inquiries is the realistic entry point
• Corporate planning: Don't evaluate agent adoption by "how natural the conversation feels." Manage it against metrics that show whether a transaction actually completed — conversion rate, average order value, repeat-purchase rate, rate of human agent intervention. The opposite outcomes Walmart's Instant Checkout and Sparky produced over the same period ultimately came down to exactly these metrics
• AI strategy/tech-partner teams: Following Emart24's AI product-recommendation approach — starting with narrow tasks where data is clear and results are easy to validate, and building success stories from there — carries less risk than aiming from day one for an end-to-end agent spanning search, cart, and checkout. As Shinsegae Group's partner switch from OpenAI to Reflection AI just 11 days after its announcement shows, building your own operational foundation connecting products, inventory, cart, checkout, and delivery matters more than the partnership with any specific LLM vendor

Conclusion

Walmart's Instant Checkout experiment failed, but reading the failure precisely turns it into a useful case study. The problem wasn't that the AI was inaccurate — it was that only the checkout button got moved over, without the cart, membership, and delivery systems that were supposed to come with it. The proof is that Sparky, under Walmart's own control over the same period, produced the opposite result: 150% GMV growth. The real battleground in agentic commerce isn't the AI model, nor the amount of data. It's who controls the transaction operating system that connects customer intent seamlessly through to actual orders, checkout, inventory, and delivery. Emart's choice to build out buyer operations, customer service, and checkout — areas that are easier to validate — ahead of a customer-facing end-to-end agent should be read as stemming from the same conclusion.