The path to an agent-first web

By Iain, 24 Mar 2026

For three decades, the web has operated on an implicit contract between the people who build websites and the people who visit them. You design pages for human eyes and organise information for human brains, monetising attention through ads, upsells, and sticky navigation patterns that keep visitors scrolling just a little longer. The browser was a viewport, the click was the unit of intent, and the entire economic architecture of digital commerce grew up around a single assumption, that your next visitor would be a person.

That assumption is breaking, not in some distant, speculative future but right now, in the spring of 2026, as AI agents learn to browse, compare, and purchase on behalf of the humans who deploy them. The shift from a human-first web to an agent-first web is the largest structural change to online commerce since the smartphone made the desktop optional. And the fight over who controls the terms of that shift is already producing lawsuits, protocol wars, and corporate panic that would be entertaining if the stakes weren’t so high.

What an agent-first web looks like

The simplest way to understand the agentic web is to watch what happens when you tell an AI assistant to buy you a pair of noise-cancelling headphones under £250. In the old model, you would open a browser, type a query, scan ten blue links, visit four retailers, compare spec sheets, read reviews, add something to a cart, enter your address, and complete the checkout. The entire process takes 20 to 40 minutes and exposes you to dozens of ads, recommendation widgets, and dark patterns along the way.

In the agentic model, you state your intent once and the agent does the rest. It parses your constraints (budget, battery life, brand preferences it has learned from prior purchases), queries product catalogues through structured APIs, cross-references aggregated reviews, compares delivery windows, and either presents a shortlist for approval or completes the purchase directly. You never visit a website, and you might not see a brand name until the box arrives. The discovery, evaluation, and transaction phases that used to occupy separate stages of a marketing funnel collapse into a single interaction between your agent and a merchant’s data layer.

The numbers behind this shift are not experimental, they are structural. McKinsey projects that agentic commerce could redirect between $3 trillion and $5 trillion in global retail spending by 2030, with roughly $1 trillion of that coming from the United States alone. AI-referred traffic to American retail sites grew 805% year-on-year on Black Friday 2025, according to Adobe data cited by MetaRouter. One in five Cyber Week orders involved an AI agent, representing approximately $70 billion in gross merchandise value, according to Salesforce.

The protocol layer underneath it all

Every time a technology shifts from experimental to inevitable, the interesting action moves from the product layer to the protocol layer. Email mattered less than SMTP, and web pages mattered less than HTTP. The agentic web will be defined less by any individual agent (ChatGPT, Gemini, Copilot) than by the protocols that determine how those agents interact with websites and with each other.

The stack that is emerging has three layers, and knowing how they fit together is worth the effort.

At the base of the stack sits the Model Context Protocol (MCP), originally created by Anthropic and donated to the Linux Foundation in December 2025. MCP is the plumbing that connects an AI model to external tools and data sources, the USB-C port for AI applications. Its SDK sees over 97 million monthly downloads, and every major AI platform now supports it. MCP handles the question “what can this agent do?” by giving it structured access to databases and external services.

Above MCP sits the Agent-to-Agent protocol (A2A), contributed by Google to the Linux Foundation with support from over 50 technology partners. A2A handles the question “how do agents talk to each other?” When your personal shopping agent needs to negotiate with a retailer’s fulfilment agent, A2A provides the handshake through a client-remote architecture with JSON-over-HTTP, agent discovery via manifest files, and asynchronous task management.

The newest and most consequential layer is WebMCP, announced by Google in February 2026 and available for early preview in Chrome Canary. WebMCP answers the question “how do agents interact with websites?” Instead of AI agents taking screenshots and guessing which blue rectangle is the submit button (a technique that is exactly as fragile and absurd as it sounds), WebMCP lets websites publish a structured “Tool Contract” through a new browser API called navigator.modelContext. A travel site can expose a search_flights(origin, destination, date) function directly, and a retailer can expose add_to_cart(product_id, quantity). The agent calls those functions with typed parameters, which means no DOM scraping, no visual interpretation, and no praying that last week’s CSS refactor didn’t break the bot.

WebMCP is jointly developed by Google and Microsoft as a W3C Community Group standard, and when Chrome and Edge, which between them account for over 80% of global browser market share, treat AI agents as first-class web participants, the web has shifted in a way that no amount of wishing otherwise will reverse.

The milestones, roughly in sequence

The agentic web did not arrive fully formed but has followed a progression that, viewed in hindsight, looks almost inevitable even though in real time it felt discontinuous and surprising.

The first phase, running from roughly late 2024 through mid-2025, was agent-assisted discovery. AI search tools like Perplexity, Google’s AI Overviews, and ChatGPT began answering product queries with synthesised recommendations rather than link lists. Users started asking “what should I buy?” rather than “show me search results.” Traffic patterns began to change. Previsible reported that traffic from AI sources surged 1,200% while traditional search traffic declined 10%. The funnel did not break, but it started leaking from the top.

The second phase, from mid-2025 through early 2026, was agent-mediated transactions, and it arrived with fanfare before hitting a wall. OpenAI launched the Agentic Commerce Protocol (ACP) alongside “Instant Checkout in ChatGPT” in September 2025, co-developed with Stripe, and US Etsy sellers went live immediately. Over a million Shopify merchants were announced as coming soon. Microsoft followed with Copilot Checkout in January 2026, integrating Shopify and Stripe alongside PayPal. Google announced the Universal Commerce Protocol (UCP) at NRF 2026, backed by Walmart, Target, Shopify, Best Buy, and a coalition of payment processors including Visa and Mastercard.

Then reality intervened, and it did so with the bluntness that commerce tends to reserve for overconfident technologists. In March 2026, OpenAI scaled back Instant Checkout, acknowledging that the feature “did not offer the level of flexibility that we aspire to provide.” According to The Information, only about a dozen of the promised millions of Shopify merchants had gone live. Users were browsing but not buying inside the chatbot, and OpenAI had not even built a system to collect state sales taxes. The company is now repositioning ChatGPT as a discovery and recommendation layer, with purchases routed to retailer apps (Walmart, Target, Instacart) or back to merchant websites. ACP lives on, but its scope has narrowed from “every Shopify store” to a small pool of large, well-connected retailers. The lesson is instructive and worth taking seriously, because discovery turns out to be the easy part of agentic commerce. Checkout, with its fraud detection, tax compliance, inventory synchronisation, and consumer trust requirements, is the part that breaks.

The third phase, which is where we are now, is agent-native infrastructure, and WebMCP is the clearest signal of it. Rather than treating agents as tolerated visitors who happen to be automated, the web platform itself is being rebuilt to serve them natively. Browsers are becoming agent runtimes, websites are exposing structured tool contracts alongside their HTML, and the distinction between “a site designed for humans” and “a site designed for machines” is dissolving into “a site designed for both.”

The fourth phase, still largely ahead of us, is agent-to-agent commerce at scale, where your personal AI negotiates with a retailer’s AI over price and warranty terms without either of you being involved. The protocols exist (A2A handles this), and the payment rails exist too (Google’s Agent Payments Protocol, AP2, handles authorisation through signed mandates with user consent). What does not yet exist is the consumer trust and regulatory framework to let this run unsupervised for anything beyond low-stakes, routine purchases, and that gap will not close this year.

The contest over agentic commerce

If the protocol layer is the plumbing, the contest over agentic commerce is the knife fight in the kitchen. And at the centre of that fight stands Amazon, wielding a lawsuit in one hand and a job posting for “agentic commerce partnerships” in the other.

Amazon controls roughly 40% of US e-commerce spending, and its advertising business generated $68.6 billion in 2025. That advertising revenue depends entirely on humans browsing and clicking, on being exposed to sponsored product placements. AI agents that skip the browsing, ignore the ads, and go straight to the best-priced item based on objective specifications represent an existential threat to Amazon’s fastest-growing revenue stream.

Amazon’s response has been to build walls, and tall ones at that. It added ChatGPT and Perplexity crawlers to its robots.txt in November 2025, blocking 47 AI bots in total, and then sued Perplexity in November 2025, alleging that the startup’s Comet browser disguised its automated agents as human Chrome users to bypass Amazon’s bot detection. In March 2026, Judge Maxine Chesney issued a preliminary injunction blocking Perplexity’s Comet from accessing password-protected areas of Amazon under the Computer Fraud and Abuse Act, giving Perplexity seven days to appeal.

The legal question at the heart of this case is genuinely interesting and will outlast the specific parties involved. Perplexity’s position is that an AI agent is an extension of the user and inherits the user’s rights. If I can browse Amazon and buy headphones, my agent should be able to do the same thing on my behalf. Amazon’s position is that the agent is a robot, subject to the site’s terms of service, which explicitly ban automated access, and the court has so far sided with Amazon.

But the hypocrisy in Amazon’s position is difficult to ignore. While suing Perplexity for scraping Amazon product pages, Amazon simultaneously launched Shop Direct (originally called “Buy for Me”), which gives Amazon’s own AI agent access to over 100 million products from more than 400,000 external merchants. CNBC reported in January 2026 that 180 businesses found their products listed inside the programme without consent. Amazon’s offer to those merchants was an opt-out email address. The company that sued a startup for scraping its pages is scraping independent merchants’ pages to populate its own AI shopping experience, and the irony of that position has not been lost on anyone paying attention.

Meanwhile, Amazon has conspicuously declined to join either of the two major open agentic commerce protocols. UCP has over 20 partners including Walmart, Target, and Visa, while ACP has Stripe and the Shopify network behind it. Amazon is absent from both, betting instead on its own closed tools (Rufus, Alexa+, Shop Direct) that keep the transaction inside Amazon’s walled garden. Jordan Berke, CEO of Tomorrow retail consulting, summarised the situation neatly. Amazon faces a “leader’s dilemma” where its dominant market share means it has the most to lose from any change.

The broader lesson here extends well beyond Amazon, because every incumbent whose business model depends on controlling the presentation layer (the ads, the recommended products, the checkout flow optimised for impulse purchases) has a structural incentive to resist agents. Agents strip out the presentation layer entirely, consuming data and returning decisions. A website’s carefully crafted brand story, its lifestyle photography, its strategic placement of premium products at eye level, none of it registers when the “visitor” is a language model parsing JSON.

This is why the agentic web is producing a tug of war between two visions. The open-web vision, championed by Google’s UCP coalition, Anthropic’s MCP community, and Shopify’s Agentic Storefronts, says merchants should publish structured data and let any agent transact on standard terms. The walled-garden vision, championed by Amazon and any incumbent with a profitable ad business, says platforms should control which agents get access and on what terms. The Amazon v Perplexity ruling is an early data point, not a final answer, but it establishes that platforms can legally block AI agents, a precedent that will shape the next decade of commerce.

What this means for site owners who are not Amazon

If you run a website that sells things, provides services, or depends on organic traffic for lead generation, the agentic web changes your operating environment in ways that are worth taking seriously now, before the 2026 holiday season.

Your website has two audiences now, one human and one consisting of a fleet of AI agents that will read your structured data, evaluate your product specs, check your shipping policies, and compare your offer against every competitor in milliseconds. If your product information is locked inside rendered HTML, buried in JavaScript-heavy single-page applications, or scattered across PDFs and image files, agents will skip you and find someone else’s data without a moment’s hesitation.

SEO is morphing into AEO. Answer Engine Optimisation, the practice of structuring your content so AI systems can parse and cite it, is becoming as important as traditional search optimisation. This means complete, accurate structured data (schema.org markup, clean product feeds, machine-readable shipping and returns policies). It means enriched metadata that answers the questions an agent would ask. Not “what does this product look like?” but “what are its specifications, what is its availability, and what do aggregated reviews say about its reliability?”

Your analytics are about to lie to you. When a customer discovers your product inside ChatGPT or Gemini and either completes the purchase through a retailer app or arrives at your checkout page through an agent referral, your traditional attribution models will miss the origin entirely, and you will need separate tracking for AI-mediated conversions to make sense of what is happening. That means UTM parameters for ChatGPT referrals, monitoring protocol-level checkouts through ACP or UCP, and accepting that the click-through rate is no longer the metric you should be watching. The unit of value in an agentic world is the transaction, not the visit.

Brand building changes, but it does not disappear. There is a school of thought that says agents will reduce every purchase to a price-and-specs comparison, making brand meaningless, but this is wrong or at least incomplete. Agents weigh aggregated reviews and reputation signals heavily when deciding which products to recommend. A strong brand with consistent quality and positive third-party endorsements will be cited more often and ranked higher by agents than a commodity alternative. What changes is where that brand building happens, shifting from your homepage to your data layer and the review profile that surrounds you.

How to prepare, practically

The good news is that preparing for the agentic web does not require ripping out your technology stack, only making what you already have legible to machines.

Start with your product data by auditing your structured data markup for completeness and accuracy. Every product should have complete schema.org Product markup including price, availability, shipping estimates, return policies, and specifications in machine-readable formats. If you sell services rather than products, apply the same discipline to your service descriptions and pricing.

Adopt the relevant commerce protocols as they become available for your platform, but be realistic about what they offer today. OpenAI’s retreat from Instant Checkout means that ACP, for now, is a discovery and recommendation channel rather than a complete transaction layer. If you are on Shopify, integration is straightforward but the payoff is in product visibility inside ChatGPT, not in-chat purchases. Google’s UCP is the more ambitious bet on end-to-end agent commerce, and if your platform provider announces UCP support, it is worth evaluating early. The conversion data from AI-referred traffic is promising (shoppers arriving from AI services are 38% more likely to buy than those from traditional channels), even if the purchase itself still happens on your own site.

Monitor your AI visibility by tracking how your products appear in AI shopping results. If an agent queries “best project management tool for agencies” and your product does not appear in the response, you have a discoverability problem that no amount of Google Ads will fix. Tools for monitoring AI visibility are emerging (Sanbi and Otterly among them), though the category is immature.

Keep a close eye on WebMCP as it matures, because when Google’s early preview programme opens more broadly, experiment with exposing structured tool contracts on your site. The Declarative API (adding toolname and tooldescription attributes to HTML forms) requires minimal engineering effort. The Imperative API (registering JavaScript tools via navigator.modelContext) is more involved but gives you fine-grained control over what agents can and cannot do on your site. Early adoption here is a competitive advantage, not a risk.

Clean up your third-party presence, because agents cross-reference reviews, directory listings, and third-party mentions when evaluating your business. Inconsistent information across platforms (wrong opening hours on Google Business, outdated pricing on a comparison site, unresolved complaints on Trustpilot) will count against you in ways that were previously just annoying but are now directly transactional. An agent that finds conflicting data will either skip you or flag you as unreliable.

Think about what you do not want agents to do, because the agentic web is not purely upside. If you sell high-margin products where the in-store or on-site experience is part of the value proposition (luxury goods, complex B2B services, bespoke anything), you may want to control how agents represent your offer rather than exposing everything as a commodity API call. The WebMCP framework lets you define exactly which actions you expose and which you keep behind human-only flows, and you should use that control deliberately.

The web we are building toward

In a 1999 essay titled “Weaving the Web,” Tim Berners-Lee described his vision for a web where machines could read and reason about content, not just display it, and he called it the Semantic Web. It did not happen the way he imagined, partly because the tools were not ready and partly because there was no economic incentive for site owners to structure their data for machine consumption. Twenty-seven years later, the incentive has arrived, not from academic standards bodies but from a generation of AI models capable enough to be useful consumers of structured information, and a generation of users who would rather delegate the tedious parts of shopping to software.

The agentic web will not replace the human web, and people will still browse, still window-shop, still click on things because they look interesting. But alongside every human visitor, there will be an increasing population of automated ones, running errands and completing transactions that their owners could not be bothered to handle themselves. The sites that thrive will be the ones legible to both audiences. The ones that fail will be the ones still optimising for a visitor who has already sent their agent instead.