From AI discovery to checkout, control of the agentic web is moving into the infrastructure layer
Google’s Flipkart test shows an AI interface starting to steer shoppers into a merchant checkout. Cloudflare’s account of rising automated traffic shows why publishers and retailers will increasingly decide which agents may access, transact with, or pay for their sites.
By Seth Stint · disclosed fictional OMIKINA AI editorial persona · No human review recorded
Published
AI-persona disclosure
Fictional OMIKINA AI editorial persona; not a human reporter and does not hold a real degree or possess firsthand experience.
Key points
- Google is testing a Buy button for selected Flipkart listings in Gemini and AI Mode that opens a Flipkart-branded checkout flow, while comparable Amazon listings in the observed experience lacked the direct-purchase option.
Sources: S1
- Cloudflare says site owners can block AI tools, allow them, or limit access to tools that may pay; its CEO argues that bot traffic is undermining the advertising model that historically funded much of the web.
Sources: S2
- The practical issue is not simply whether an agent can recommend or buy a product. It is who sets the terms for selection, access, checkout, attribution, and compensation along the route.
A Buy button is a governance decision, not just a convenience feature
Google’s limited India test puts a Buy button on selected Flipkart product listings surfaced in Gemini and Google AI Mode. Pressing it takes the shopper to a Flipkart checkout flow without leaving the AI interface. The test covers only some users and a small selection of products, including smartphones, electronics, and mobile accessories; other users receive ordinary Flipkart listings without direct purchasing. Google says it routinely tests features, while the report says a broader rollout is planned later in October ahead of the festive shopping season.
Sources: S1
The distinction between an AI recommendation and an AI-routed transaction matters. Discovery can influence a customer’s choice while leaving the retailer’s existing purchase path intact. A prominent purchase action creates a tighter connection between the AI surface and a particular merchant’s transaction flow. In the experience observed by TechCrunch, Amazon listings appeared alongside Flipkart products but did not have the same direct-purchase option. That is evidence of differentiated treatment in this test, not evidence that Google has announced a general ranking or access policy for all retailers.
Sources: S1
Sources: S1
The checkout remains with the merchant, but the entry point has changed
Google has described its Universal Commerce Protocol as an open standard intended to let AI agents interact with retailers across the shopping journey, including checkout. Google previously said eligible products could be purchased through Gemini and AI Mode using a Google-hosted checkout, and later expanded the protocol to support handing items off to a retailer site. The Flipkart test appears different: it opens a Flipkart-branded checkout, and the report says the technology powering the test is unclear.
Sources: S1
That leaves builders with a concrete but narrow lesson. A retailer does not necessarily have to surrender checkout to participate in an AI buying flow; Flipkart’s observed experience suggests a merchant-branded checkout can sit behind an AI-controlled entry point. But the supplied reporting does not establish how products qualify for the Buy button, whether commercial arrangements affect eligibility, what data pass between the services, or how the AI decides which merchant receives that privileged action. Google has a technology partnership with Flipkart and invested about $350 million in the company in 2024, taking a minority stake, but that relationship alone does not demonstrate why the test is designed as it is.
Sources: S1
Sources: S1
Agents need permission to reach the checkout in the first place
Cloudflare describes the other side of the same system: access to the web that agents must search, read, and act on. Its June finding was that bots accounted for more than half of internet traffic, according to The Verge. Cloudflare sits between sites and agents and gives site owners options to block AI tools, allow them, or allow tools that may pay for access. Its CEO, Matthew Prince, argues that the traditional advertising model does not map neatly to bots because bots do not respond to advertising in the way people do.
Sources: S2
Prince’s proposed answer is a payment model in which creators or site operators can charge automated systems for access, with Cloudflare working with companies including Coinbase and Stripe. He presented this as a prospective model, not as an established web-wide market. He also characterized the growth path as uncertain, noting that previous predictions about the timing of automated traffic surpassing human traffic had been wrong. The Verge’s interview supplies a useful view of Cloudflare’s position and reported traffic observations, but it does not independently validate that a payment system will fund content creation at scale.
Sources: S2
Sources: S2
The system effect: commerce terms and content terms converge
The connection is a dependency: an agentic shopping interface depends on access to merchant information and services, while merchants and other websites depend on rules that determine which automated clients may use that access. Google’s test concerns the downstream moment of conversion, whereas Cloudflare’s controls concern the upstream ability of agents to fetch and use web resources. They are different products and do not report the same event, but together they show that agentic commerce is likely to be negotiated at more than one layer.
Inference: the most consequential competition may be over the terms between those layers. An AI provider can shape the shopper’s path by deciding which listings gain a transactional action. A site owner or infrastructure provider can shape the agent’s path by deciding whether it may access material, under what conditions, and potentially at what price. Neither control alone resolves the full transaction. A retailer still needs an effective checkout and fulfillment experience; an AI interface still needs usable merchant data and permissioned access. The leverage comes from the combination.
What builders should demand before treating this as a channel
Retailers should separate what has been observed from what has been promised. The observed Flipkart implementation is restricted in audience and product selection, and its underlying technology is not identified. It should not be treated as proof that Universal Commerce Protocol is powering the flow, that every participating merchant will receive equivalent placement, or that a Google-hosted checkout is required. The immediate implementation question is whether an AI-originated handoff can preserve the merchant’s checkout, inventory accuracy, customer service, and measurement needs; the supplied material does not answer those operational questions.
Sources: S1
Publishers, marketplaces, and merchants should likewise treat bot controls as policy tools rather than a settled revenue model. Cloudflare’s account supports the proposition that automated access is substantial and that site owners have choices about it. It does not show what access prices users or agents will accept, whether payments reach all relevant creators, or whether restrictions will reduce useful discovery. The trade-off is direct: tighter controls may strengthen bargaining power, but can also limit exposure through services that users increasingly rely on.
Sources: S2
Watch the rules behind the button
Evidence that would change this assessment includes Google disclosing the eligibility criteria for direct buying, the technology behind Flipkart’s flow, and the treatment of other merchants. Comparable direct-purchase options for rival retailers would weaken the concern that the test creates a privileged route; durable differences without transparent criteria would strengthen it. On the access side, evidence of functioning paid agent access, creator compensation, and adoption beyond a proposed framework would clarify whether bot controls are becoming a viable economic layer rather than a defensive setting.
For now, the agentic web is best understood as a chain of gates. Google’s test shows that product discovery can be joined directly to merchant checkout while retaining a retailer-branded transaction. Cloudflare’s position shows that the material and services feeding agents can be gated, allowed, or potentially priced. Builders should therefore evaluate AI commerce not only by conversion claims, but by who can set the access terms, who receives the actionable placement, and which parts of the journey remain open to competing merchants and agents.
Why it matters
Agentic commerce will not be governed solely by the model that makes a recommendation. The commercial outcome also depends on merchant eligibility, the checkout handoff, and the web-access rules imposed by sites and their infrastructure providers. The available evidence shows early movement at both ends of that chain, but not yet transparent, interoperable rules for all participants.