August 3, 2026

The agent does not need to complete the booking to improve the sale. More than a year ago, Balaji Srinivasan described AI's current shape in terms that fit this market unusually well. “AI doesn't do it end-to-end. It does it middle-to-middle. The new bottlenecks are prompting and verifying.” He was writing about AI broadly, not shopping, but commerce is now demonstrating the point. Last week's digest showed gateways gaining economic power while buyers still approved the final commitment. Commerce has demonstrated the value of discovery for months. This week made its economic structure clearer. Hotel brands exposed live inventory inside assistant interfaces, retailers gained tighter control over agent-readable catalogs, and payment networks pushed verification into the rails. The middle is becoming a market of its own.

Radisson Hotel Group supplied the clearest design. Its new hotel discovery app in ChatGPT lets a traveler describe a trip in ordinary language and search more than 1,000 hotels across more than 100 countries. Results include live inventory and rates, location context, amenities, hotel details, and a map. When the traveler is ready, the app sends them to Radisson's website to reserve. In-chat booking, loyalty recognition, and reservation changes remain future plans. Even without those capabilities, Radisson now shapes the shortlist and receives the traveler back inside its own booking environment. The agent makes the brand easier to discover without displacing the place where the commercial relationship begins.

Travel makes the emerging gateway economics visible because discovery and distribution have always carried a price. Air New Zealand CEO Nikhil Ravishankar posed the strategic question directly. The airline knows how to build brand love with humans, he said, but does not yet know how to do that with an agent. One contributor's analysis in HospitalityNet maps three possible routes. An OTA can enter the assistant and preserve its familiar commission, an MCP aggregator can refer the traveler to a hotel site for a new fee, or a hotel can operate its own infrastructure and reach the assistant directly. None is settled. That is the opportunity. Brands, marketplaces, and new intermediaries can still compete on relevance, service, attribution, conversion, and cost instead of inheriting one inevitable tollbooth.

Merchants are also gaining practical control over what the agent sees. Salesforce's July B2C Commerce release shows what that can look like across two surfaces. Retailers decide which products enter the OpenAI feed, how catalog fields map to its format, and how titles, prices, images, and inventory synchronize each day. On their own storefronts, natural-language search handles longer requests while merchandising rules can emphasize strategic inventory or suppress products that are out of stock. The agent gains cleaner data and a better chance of understanding the shopper, while the seller retains authority over availability and presentation. Better discovery should make a catalog easier to navigate, not make the merchant invisible.

Usage is beginning to justify that work. CX Dive, reporting Amazon's second-quarter earnings call, said active Alexa for Shopping users nearly doubled in the quarter from a year earlier while interactions rose fivefold. Amazon CEO Andy Jassy said more than 350 million shoppers had used the assistant during the previous twelve months. Those are company-reported engagement figures, not proof that the assistant caused a purchase, but they establish growing use at meaningful scale. The commercial case does not depend on every session ending with autonomous checkout. When an assistant becomes a regular place to ask about products, the merchants and platforms that answer well move closer to the moment when money changes hands.

Reaching that moment consistently will require product understanding that goes beyond polished conversation. Onton's vendor-built home-decor benchmark gives the problem a useful contour. On 90 curated, intent-heavy queries scored by three independent LLM judges, the company reported precision among the top ten results of 0.630 for Onton, 0.543 for Google Shopping, and 0.469 for Amazon. All three judges ranked the systems in that order, though their agreement on exact scores was moderate. The narrow test targets difficult queries, and Onton still lost ground where Amazon's category data better captured functional specifications. The lesson is not that one vendor has solved product search. The result suggests that taste, context, negation, and constraints can become searchable product knowledge rather than friction the shopper has to resolve alone.

Payment networks are treating verification as an extension of existing rails rather than a reason to wait. PYMNTS, reviewing second-quarter disclosures from the largest U.S. card networks, reported that Mastercard is scaling Agent Pay through live agentic transactions and working with more than 30 industry participants on machine-speed permission, orchestration, and settlement. Its Verifiable Intent work with Google is designed to let a party challenge a transaction it did not authorize and use existing chargeback processes for the dispute. These are network claims reported by PYMNTS, not technical validation of every edge case. They still show why bounded autonomy can advance quickly. The value is not only moving money. It is connecting an agent's action to recognizable permission and recourse.

This week moved commerce forward without moving shoppers out of the decision. Discovery became more actionable, merchants retained room to shape presentation, and payment networks made verification part of the transaction path. Each improvement shortens the distance between a shopper's intent and a product that actually fits. The best gateway will earn its place by making that path faster, more accurate, and more useful while leaving shoppers in command and giving brands meaningful ways to compete. Full autonomy may come later. Commerce does not have to wait for it to get better.