This three-part series will chart the rise of the ‘two-tier internet’ and its impact on the future of commerce. In this first edition, Nathalie Lethbridge, Founder, and Jess Carlin, Partner at advisory firm Atonik, explore the infrastructure moment: the technical shift that’s allowing AI agents to operate as a parallel layer of discovery and transaction.
One Tuesday this past September, a surprisingly mundane product update landed with tectonic force. OpenAI enabled Instant Checkout, a feature that lets ChatGPT users buy products from Etsy, Shopify, and now major US retailers like Walmart and Target, all without loading a single webpage. No browsing, no tabs, no search results. Just: tell the AI what you want, it surfaces it, you hit “Buy.”
It’s been building for the last year, but with that one product release, the web, the one we’ve spent 25 years clicking, scrolling, and optimising for, split in two. A two-tier internet is here. One tier built for humans who browse. Another tier built for agents who ask, interpret, and transact. It’s early, but the signals are unmistakable.
The Checkout Heard Around the Web
But what does this actually look like in practice? The concept is relatively straightforward: ChatGPT could now display products from Etsy and Shopify merchants directly in the chat window. As long as those merchants enabled the Agentic Commerce Protocol (ACP)—the direct “rails” through which an agent can easily access product—ChatGPT would handle the checkout inside the conversation. A new kind of shopping infrastructure just activated.
But of course, the proof lies in the adoption metrics. There are graveyards of products that sound great on paper but fail to engage with consumers. While it’s still early days, this doesn’t seem to be one. During the 2025 Black Friday period, AI-driven shopping traffic jumped 805 percent, according to Adobe. For a behaviour that didn’t meaningfully exist a year ago, the uptake is proving substantial. Shopify revealed that AI-led traffic is up sevenfold since January 2025, with AI-driven orders up elevenfold in the same period. Sixty-four percent of shoppers expect AI to influence their purchasing decisions.
It’s not just OpenAI either. Perplexity launched a paid-for agentic shopping assistant back in November 2024 and recently rolled out virtual try-on features alongside a PayPal partnership. Google rolled out its own agentic checkout system, and a Gemini-powered shopping concierges that compare prices, track coupons, and forecast deals. Amazon launched Rufus, its walled-garden shopping agent trained exclusively on Amazon’s own ecosystem in February 2024.
The platforms are racing to own the rails. And the shift isn’t confined to North America. In India, now the world’s largest real-time payments market, the National Payments Corporation of India (NPCI) and Razorpay launched AI-led payments directly inside ChatGPT, allowing consumers to complete UPI transactions within the conversational interface. In a country where UPI processes more than 20 billion transactions a month, this integration signals something major. This is becoming global infrastructure.
Agents at the Gate
How these two different infrastructure modes are now operating simultaneously tells us something important about the emerging commerce layer. For two decades, the web has operated on a “hunting” model. This is the one we all know and have become comfortable with. You search, scroll, click, compare, bounce, search again. Entire industries have been built on navigating this: SEO, SEM, retargeting, cart abandonment recovery, product detail page optimisation, affiliate marketing.
The “browsing” mode remains visual, exploratory, and tab-driven. It’s optimised for pages, clicks, and funnel analytics. This tier isn’t disappearing.
But the second layer, one powered by agents and conversational commerce, is a completely different beast. In the conversational commerce mode, agents operate on natural-language queries, so intent, constraints, and preferences get encoded directly into the request, and agents navigate product data to match. It’s what makes ‘shopping at the point of inspiration’ possible: the moment you think of something is now the moment an agent can find it and buy it.
This mode doesn’t route consumers through the traditional paths that we are used to seeing in the browsing mode, which means that it reduces dependency on the homepage, search results, and category navigation. Instead, it creates a direct interface between intent and transaction execution.
This is why the Shopify data is significant. AI systems are funnelling traffic directly to products without consumers ever visiting a store’s homepage or engaging with traditional discovery mechanisms. You ask for an item under £30 that ships fast and matches specific criteria. The agent retrieves and surfaces the relevant products. This is a completely different internet logic, operating on invisible infrastructure most people never see.
The Agent Layer: Infrastructure at Work
Understanding what infrastructure actually enables this new mode of commerce to function requires looking at the technical architecture that sits beneath these transactions. This is where the open agentic layer enabled by the Agentic Commerce Protocol, set up by OpenAI and payment provider Stripe, comes into play. In structural terms, it’s a cross-platform layer of rails that sits above traditional websites and allows agents to easily access product information.
The entire shift rests on this concept of an open agentic infrastructure, an architecture where LLMs plug into tools, APIs, memory, and structured product data. This is what lets an AI move from generating text to performing transactions. This push toward open infrastructure was recently formalised through the new Agentic AI Foundation (AAIF), created under the Linux Foundation to steward open, interoperable standards for agent systems as they move into real-world production. This distinction matters. Previous conversational commerce experiments operated on chatbots that could discuss products but couldn’t actually transact or access real-time inventory data.
This new architecture changes the game completely.
This is why OpenAI calls its Agentic Commerce Protocol “open rails.” And why Amazon is fighting to keep its rails closed. Whoever controls the rails controls discovery, trust, economics, and ultimately the advertising opportunities embedded in these transactions. The internet’s power centre is shifting from search engines (which operate on keywords and user navigation) to agents (which interpret intent directly). It’s a complete reorganisation of power in digital commerce infrastructure.
The Web Didn’t Die – It Split
So the question now moves on from whether conversational commerce is replacing traditional retail. it won’t. There will always be consumers who navigate through browsing, who search on Google, who read reviews across platforms.
The bigger story is the emergence of this parallel internet later, where two infrastructure modes serve different purchase contexts.
The human web remains essential for product categories and purchase contexts that benefit from visual discovery, extended narrative, and exploratory interaction. The agentic web will be essential for functional purchases, deal finding, repeat buying, time-sensitive tasks, and straightforward consumer problem-solving.
Humans will use both. The answer for brands and platforms is that you must serve both. The structural reality is that once consumers have access to a direct agentic pathway for a particular purchase context, traditional web-based discovery and navigation become secondary.
For brands and media buyers, this fractures the visibility landscape in ways that demand different strategies operating simultaneously.
For media buyers, your traditional spend on Google Shopping, Meta, and Amazon remains effective within the browsing tier. But agent-mediated commerce exists outside that ecosystem entirely. You can’t buy placement in a ChatGPT recommendation (at least not yet, it’s coming!). Your measurement frameworks will need to account for a visibility channel you can’t directly control, one where success depends on structured data quality rather than creative spend. You’re effectively operating two parallel campaigns with different metrics, different levers, and different ROI models.
For brands, you’re no longer competing on a single surface. You compete for attention through traditional channels (creative, messaging, audience targeting), whilst simultaneously competing for agent recommendations through data completeness, accuracy, and operational reliability. These require different teams and different investments. Your homepage design doesn’t help you win in an agent interface. Your product data structure doesn’t help you win with Google’s Shopping Graph.
The practical challenge then becomes reconciling these competing priorities and budgets within a single organisation, building capabilities while the entire industry is still learning, and maintaining coherence across two very different competitive landscapes.
Who Guides What We Buy
When the infrastructure controlling discovery changes, the implications for how commerce operates become immediately apparent.
First, the homepage’s role is shifting as more discovery begins upstream in conversational interfaces. If users access products primarily through agents rather than websites, traditional site navigation and information architecture become less critical. Structured data, product attributes, fulfilment reliability, and metadata quality become the most important elements of visibility instead.
Second, discovery is being intermediated at a new layer entirely. Rather than consumers finding brands through search engines or marketplaces, agents select which brands and products appear in response to queries. The ranking logic determining these selections isn’t controlled by merchants. It’s embedded in agent systems operated by platform companies.
Third, a new battle over open versus closed rails is defining platform strategy. It’s fundamentally about control: who controls the shopping graph, the ranking logic, the economic leakage, and the advertising opportunities that follow. OpenAI wants interoperable, multi-platform commerce agents that can operate across ecosystems. Amazon wants a vertically integrated, end-to-end, closed agent economy that keeps commerce within its walled garden. Google is defending search from becoming obsolete as a discovery mechanism.
This is the real competition reshaping digital commerce. It’s not about mobile or social. It’s about who owns the infrastructure that sits between intent and transaction.
The question now is whether your strategy is built around this fractured reality, or whether you’re still optimising for the internet as it was.
In the next edition, we’ll explore the human side of this shift: how intent becomes a conversation, how agents become decision-makers, and how consumer behaviour changes when the web starts answering instead of asking.
Read more…
Part 2 – The Two-Tier Internet: How the Holiday Season Showed Shopping Is Changing
Part 3 – Strategy in a Two-Tier Internet: How You Actually Compete When Discovery Splits in Two
