OpenAI's Shopping Agents Flopped — Here's What They Got Wrong

OpenAI's shopping agent flopped because it tried to replace the entire purchase funnel with a single centralized model, ignoring the messy, trust-dependent reality of how people actually buy things online. Walmart tested roughly 200,000 items through OpenAI's "Instant Checkout" feature -- an in-chat purchasing flow tested since November 2025 and discontinued in early March 2026. The conversion rate never reached parity with Walmart.com, and Daniel Danker, Walmart's EVP of Product & Design, described the in-chat purchasing experience as "unsatisfying."

That word choice stunned me. If an AI agent can't even beat a distracted human scrolling on their phone during lunch, something is structurally broken.

The core mistake: treating shopping like search#

OpenAI built Operator's shopping mode the same way they built ChatGPT -- as an answer engine. You ask, it fetches. But shopping is not search. Search has a correct answer. Shopping has preferences, trade-offs, regret, and returns.

When Operator suggested products, it optimized for relevance to the query. Not for the user's budget constraints, brand loyalty, shipping speed requirements, or the fact that they bought the same moisturizer three times already and just need a reorder button. Search Engine Land reported that Instant Checkout converted at roughly 70% of Walmart.com's native rate -- better than raw ChatGPT browsing, but still well below parity.

Shopping requires context that no centralized model has. Your purchase history, your sizing data, your household's dietary restrictions, your budget this month. OpenAI tried to infer all of this from a chat window. It doesn't work.

Retailers fought back harder than expected#

Here's what most coverage missed. Retailers actively blocked Operator.

A Walmart spokesperson told CNBC: "We learned that our customers want consistency across every touchpoint." So Walmart replaced Instant Checkout with "Sparky," their own chatbot, now embedded directly inside ChatGPT. Sparky converts at roughly 70% of Walmart.com's rate -- an improvement over raw Instant Checkout, but still not parity.

Target reportedly declined a similar OpenAI partnership altogether. Other major retailers followed suit with various forms of Operator throttling.

This was entirely predictable. Retailers have spent billions building their own conversion funnels. They're not going to hand that over to OpenAI for free. The incentive alignment was wrong from day one.

The affiliate economics don't pencil out#

OpenAI reportedly planned to monetize shopping through affiliate commissions, taking 3-8% on completed purchases. But affiliate commerce only works at scale when you control the demand funnel. Google Shopping works because Google controls search intent. Amazon works because Amazon controls the catalog and fulfillment.

OpenAI controls neither. They're a middleman with no inventory, no logistics, and no purchase data moat. PYMNTS reported that Walmart shut down the agentic commerce experiment entirely, pivoting to Sparky as a retailer-controlled alternative. The affiliate model never reached the scale OpenAI projected.

Compare this to what's working in agent commerce. Visa's new agent payment CLI isn't trying to replace the shopping experience. It's giving agents their own financial rails to transact within existing merchant ecosystems. That's the difference between disrupting and enabling.

Personal agents beat centralized agents every time#

The fundamental problem isn't that AI shopping agents are a bad idea. It's that centralized AI shopping agents are a bad idea.

A shopping agent that works for OpenAI has misaligned incentives. It needs to drive affiliate revenue. It needs to satisfy OpenAI's retail partners. It needs to balance user experience against monetization. The user comes third.

A shopping agent that works for you has one job: get you the best deal on the thing you actually want. It knows your purchase history because it lives in your data. It knows your budget because it has access to your financial context. It doesn't care about affiliate commissions because it's not monetized that way.

This is why the shift from chatbots to personal agents keeps accelerating. When the agent serves you instead of a platform, the entire incentive structure flips. An agent marketplace like Shopify's agentic storefront model gets this right -- agents negotiate with storefronts on behalf of users, not on behalf of Shopify.

What a working shopping agent actually looks like#

The agents that are quietly succeeding at commerce aren't trying to replace the entire shopping experience. They're handling specific, high-friction tasks:

• Reordering: An agent that knows you buy the same coffee every 3 weeks and just does it. No browsing required.
• Price monitoring: An agent that watches 15 retailers for a specific item and alerts you when the price drops below your threshold.
• Returns management: An agent that tracks return windows, initiates returns, and follows up on refunds. This alone saves hours per year for frequent online shoppers.
• Comparison shopping with constraints: Not "find me a laptop" but "find me a laptop under $1,200 with at least 32GB RAM that ships in 2 days to 94110."

These are narrow, well-defined tasks where agents dramatically outperform humans. They're also tasks where the agent needs your context, not OpenAI's affiliate partnerships.

If you're building agents that handle real tasks -- shopping, scheduling, monitoring -- RapidClaw lets you deploy personal AI agents in under 60 seconds, with the context and integrations that centralized platforms will never have.

The trust problem nobody solved#

There's a deeper issue beneath the economics. People don't trust an AI to spend their money.

A survey by Bain & Company in January 2026 found that only 9% of US consumers said they'd be "comfortable" letting an AI agent complete a purchase over $50 without human confirmation. For purchases over $200, that number dropped to 3%.

OpenAI tried to solve this with a confirmation step before checkout. But that defeats the entire purpose. If the human has to review every purchase, the agent is just a fancy product search engine, which ChatGPT already was.

Trust in agent commerce will come from agents that build track records over time. Your personal agent that has correctly reordered your groceries 40 times in a row earns the right to handle a $200 purchase autonomously. A centralized agent you just met does not.

What happens next#

OpenAI isn't going to abandon shopping. They'll iterate -- Walmart's Sparky integration and a planned Google Gemini + Sparky integration expected in April 2026 show the direction. But the structural problems -- misaligned incentives, retailer resistance, no data moat, low consumer trust -- aren't bugs. They're features of the centralized model.

The companies that will win agent commerce are the ones building infrastructure for personal agents to transact, not the ones building a single mega-agent that tries to shop for everyone. Visa, Stripe, and Shopify understand this. OpenAI, so far, does not.

The shift will happen, but it'll look more like agents with their own payment credentials operating within merchant ecosystems than like a single chatbot trying to be the Amazon of AI.

---

Frequently Asked Questions#

Why did OpenAI's shopping agent fail?#

OpenAI's "Instant Checkout" feature -- tested with Walmart across roughly 200,000 items since November 2025 -- struggled because it used a centralized model that lacked personal shopping context, faced retailer resistance, and never converted at parity with native retail apps. Walmart's EVP of Product & Design called the in-chat purchasing experience "unsatisfying," and the feature was discontinued in early March 2026.

Can AI agents actually buy things online?#

Yes, but the agents that work best are personal agents with access to your purchase history, budget, and preferences. Centralized agents like OpenAI's Instant Checkout that try to serve millions of users with no individual context converted at roughly 70% of native retail rates -- better than raw browsing, but still below parity.

What is agentic commerce?#

Agentic commerce refers to AI agents autonomously discovering, negotiating, and completing purchases on behalf of users. It includes everything from automated reordering to price monitoring to full autonomous procurement. The infrastructure is being built now by Visa, Stripe, Mastercard, and others.

Will AI replace online shopping?#

AI agents will handle an increasing share of routine and high-friction purchases like reorders, price-sensitive items, and returns. But browsing, discovery, and high-consideration purchases will remain human-driven for the foreseeable future. Only 9% of consumers currently trust AI to complete purchases over $50 without confirmation.

How is personal agent shopping different from ChatGPT shopping?#

A personal agent runs on your behalf with access to your data, your preferences, and your financial context. ChatGPT and Operator run as centralized services optimizing for platform revenue. The incentive alignment is fundamentally different, which is why personal agents convert at higher rates for their owners.