Roughly 61 million American consumers now begin their retail shopping research using AI. They have swapped the search box. What they have not swapped is where the money goes. Six in ten AI-assisted purchases still close on Amazon, according to new PYMNTS data. The search routine changed. The reward did not.
This is the central tension in agentic commerce right now, and it has a name borrowed from consumer packaged goods. PYMNTS CEO Karen Webster calls it the Febreze moment.
The Reward the Industry Has Not Built
In the late 1990s, Procter & Gamble launched Febreze expecting a blockbuster. The product eliminated odors instead of masking them. It flopped. The reason was not the technology — it worked. The reason was that most people who needed it could not smell the odor in their own homes. There was no cue to reach for the bottle, and when they did spray, they noticed no difference. No cue, no reward, no habit loop.
P&G fixed the problem by changing what it sold. It stopped selling odor elimination and started selling the fresh-scent reward at the end of cleaning. Febreze banked more than $230 million in its first year after relaunch. The technology stayed the same. Everything around it changed.
Webster argues agentic commerce is stuck in the same place. The industry is selling smart, seamless checkout. Consumers are not looking for a new way to click a button. They are looking for a guarantee that when something goes wrong — wrong size, wrong seller, refund that never shows — someone will make it right. “Trust is what’s left after a habit has paid off so many times that you stop checking and second-guessing,” she writes. Amazon has spent two decades building that habit. The agent has not built it yet.
What the Numbers Actually Show
A PYMNTS Intelligence survey of 2,191 US consumers in August 2026 found that 23% now start their retail shopping research with AI. That is 61 million people. Average AI-assisted spend rose from $889 in June to $1,039 in August — up 17% in two months. The discovery part is working.
But the buying part is not moving. Sixty percent of AI-assisted purchases still complete on Amazon. Big box and department stores get 20%. Brand sites get 6%.
The most telling data point comes from a controlled test PYMNTS ran. Consumers were told the AI assistant would find the exact same item at the same price, with the same shipping and the same returns, from three places. Forty percent still chose the marketplace. Twenty-two percent chose the department store. Sixteen percent chose the brand.
When asked why, the top answer was easier or more reliable returns, at 37%. Better selection, 35%. Habit, 33%. Better customer service, 32%. They did not believe the alternative was equal, even when told it was. Behavioral economics explains why: losses feel roughly twice as painful as equivalent gains. A bad purchase costs money. A bad search answer costs only time.
OpenAI learned this the hard way. It launched Instant Checkout in September 2025 and discontinued it by March 2026. Only about 30 Shopify merchants ever went live. Walmart said purchases made inside ChatGPT converted at one-third the rate of those sent to its own site. The “smart and seamless” pitch failed because it described the status quo for someone who already shops on Amazon. It was odor removal for someone who could not smell the odor.
Two Sides of the Same Gap
This consumer trust trap connects directly to the supply-side problem Forkast has been tracking. Only 11% of SMBs currently qualify as agent-ready, according to the PYMNTS/Visa GDSI Merchant Edition. Only 15% of merchants have the structured product data agents need to evaluate what is being sold. Only 23% can distinguish AI-driven traffic from human users. Merchants cannot see the traffic, and consumers will not leave the familiar. Agentic commerce is stuck on both sides.
The infrastructure layer is arriving. Stripe built full-stack agent commerce capabilities with 288 launches at Sessions. The merchant landscape is splitting: Shopify opened its entire catalog to Muse and Shop Pay, while Amazon blocked the agent the same day. The plumbing is getting built. The trust is not.
Webster’s prescription is blunt: match Amazon on the bad day and beat it on the all-in price. “The intermediary can open the door,” she writes. “The reward has to be the merchant’s own.”
Sixty-four percent of consumers check who the seller is when they buy on Amazon. They do not trust everything on the platform. They trust the platform to fix it when things go wrong. That distinction — between trusting the seller and trusting the resolution — is the gap no checkout protocol has closed.
What Changes the Equation
The payment networks are positioned to own this gap, if they can make the promise visible. Zero liability protections already made a generation comfortable shopping online. The chargeback is the original “someone will make it right.” As Webster puts it, “top of wallet becomes top of agent.” The issuer that makes its guarantee travel with the credential wherever the agent shops — wrong item, wrong seller, refund that never shows — could make a brand site as safe to try as the marketplace.
There is a gap between what consumers say would move them and what the agent currently delivers. A saved digital wallet on a brand site would move 24% of marketplace shoppers. An AI assistant handling the end-to-end checkout would move 12%. But in the controlled test, the marketplace still won at 40%. That 20-point gap between the 40% who default to the marketplace and the 20% who say everything-equal would choose differently is the prize. It is what comes into play when consumers actually believe the alternative is as safe as Amazon. Today they do not, and a survey cannot make them. Only experience will.
Until that experience materializes, 61 million consumers will keep asking AI what to buy. And then they will buy it on Amazon.