What's Agentic Commerce? The End of the Traditional Search Journey.

The traditional search-to-click web journey is dying. Within a decade, autonomous AI agents will research, negotiate, and complete purchases on behalf of consumers, and the economic opportunity sits somewhere between USD 3 trillion and USD 5 trillion globally by 2030. That's not a typo. I've spent years auditing ecommerce stacks on platforms like Shopify and Magento, and in that time the search bar has gone from the centre of the shopping experience to something that feels increasingly like a legacy relic. The shift isn't theoretical. 45% of consumers already use AI in some part of their buying journey, and that number was captured before the latest wave of agentic tools hit the market. What's replacing search isn't just a better search bar. It's a fundamentally different model where the human stops browsing and the machine starts transacting. The shift shows up clearly in the 2026 ecommerce statistics: AI-referred sessions are compounding at 4,700% year on year off a tiny base. The power dynamics between merchants, platforms, and AI intermediaries are being rewritten, and not necessarily in merchants' favour. Fast.

The Death of the 'Search-and-Click' Model

The 'search-and-click' loop has dominated online retail for two decades: type a query, scan results, click a listing, compare options across tabs, add to cart, checkout. Every step fed data back to merchants and platforms. Every click was monetisable. Google built a USD 300 billion advertising business on the assumption that loop would last forever.

Agentic commerce breaks it. Instead of a human clicking through ten product pages, an AI agent gets a brief ("I need a waterproof cycling jacket under GBP 150, breathable, available in large, delivered by Friday"), then researches options, compares specs, checks stock levels, and executes the purchase. No tabs. No comparison sites. No banner ads.

Think of it as the shift from a browsing librarian model, where you walk the aisles and pull your own books, to a personal concierge model, where you describe what you want and someone handles the logistics end to end. The concierge doesn't care about shelf placement or promotional endcaps. It doesn't care about your brand story. It cares about matching your criteria. Full stop.

That's a structural problem for anyone whose revenue depends on the old journey. Search platforms, affiliate networks, and merchants who invested heavily in front-end conversion optimisation all lose influence when the customer never visits the site. The IBM Institute for Business Value pegs the global opportunity at USD 3 trillion to USD 5 trillion by 2030, but that value has to come from somewhere. A lot of it will get redirected away from the businesses that currently own the click.

How Agents Handle Discovery and Evaluation

Agentic systems in 2025 bear almost no resemblance to the chatbots from 2020. Those were reactive, keyword-matching scripts bolted onto FAQ pages. Current systems can reason, plan, and act across multiple services simultaneously. They keep context across multi-step tasks, call external APIs, and adjust their approach when they hit a dead end.

Discovery is where the difference is sharpest. A traditional shopper depends on SEO rankings, paid placements, and review aggregators. An agent bypasses all of that. It queries product data directly through APIs or structured feeds, pulls pricing from wherever it can reach, cross-references reviews, and checks everything against what the user actually asked for. SEO title tags and meta descriptions don't matter when no human is reading the results page.

That said, agents aren't magic. They're only as good as the data they can actually get to. If a merchant's product catalogue isn't available through a well-documented API or a structured data format, the agent simply won't find it. This comes up repeatedly in Shopify and BigCommerce audits: stores with beautifully designed front ends but no programmatic access to their inventory data are invisible to agentic workflows. Our CSV-first architecture guide for Shopify covers how to structure catalogue data so an agent (or anything else that talks to an API) can actually read it. The deeper technical playbook, llms.txt, JSON-LD, and the JavaScript blind spot that leaves most Shopify PDPs invisible, sits in our LLM-ready Shopify storefront guide. The bar is moving well beyond a pretty storefront now.

Trust calibration is the catch. When a human browses, they apply judgement, reading between the lines of a suspiciously glowing review, noticing that a product photo looks different from the one on Reddit. Agents don't have that instinct yet. They can be gamed by manipulated structured data or artificially inflated ratings in machine-readable formats. The fraud surface is different, not smaller.

On evaluation, agents excel at objective criteria: price, specs, delivery times, return policies. Subjective ones are harder. Build quality feel, brand reputation nuances, the tacit knowledge a seasoned buyer carries. For commodity purchases, agents are already better than most humans. For considered purchases, they're a useful filter. Not a replacement.

Standardising the Transaction: The Agentic Commerce Protocol

ACP is an open-source standard, community-designed under the Apache 2.0 licence, that defines how AI agents communicate with merchants to browse catalogues, initiate purchases, and handle payments. For agents to transact at scale, they need exactly this kind of common language. And right now, ACP is the only candidate.

OpenAI is the first AI platform to implement ACP within ChatGPT. Stripe is the first compatible payment service provider, enabling agents to process transactions by passing secure payment tokens. That pairing matters because it creates a reference implementation other platforms and PSPs can follow. Not a theoretical spec. An actual working example.

The merchant-of-record model sits at the centre of how ACP works. Even when a transaction originates inside a third-party AI app, say a user asking ChatGPT to buy running shoes, the business remains the merchant of record. The merchant controls product presentation, pricing, fulfilment, and customer service. The agent is an intermediary, not a retailer. That distinction matters for tax compliance, warranty obligations, and brand control, and it's one of the more sensible calls in the spec, frankly.

On security: ACP handles PCI-compliant payment token passing without exposing sensitive credentials to the agent layer. The agent never sees a card number. It passes a tokenised reference to the PSP, which handles the actual charge. Right architecture. Still requires merchants to integrate with a compatible PSP and maintain token lifecycle management on their backend.

Here's a simplified view of what ACP needs from merchants:

The catch is implementation. This isn't a plugin install. Supporting asynchronous purchase flows means rearchitecting how your checkout works. Most ecommerce platforms (Shopify, Magento, BigCommerce) assume a synchronous, browser-based session. Removing that assumption requires backend engineering, webhook infrastructure, and testing against agent behaviours that don't follow predictable user paths. Merchants running on tightly customised themes with fragile checkout scripts will feel this most. Considerably.

Limitations and Friction Points

The protocol spec makes this look cleaner than it's.

Data loss is the first real concern. When a customer buys through your website, you capture session analytics, browsing behaviour, referral source, and email addresses for remarketing. When an agent buys on a customer's behalf, you get the transaction, but not much else. Maybe an order note. The rich behavioural data that fuels retention marketing disappears, and merchants become dependent on whatever the AI platform chooses to share. No guarantee that data will be granular, or even available.

Then there's the walled garden problem. If agents favour specific platforms, and right now OpenAI's ChatGPT is the only major implementation, small merchants who can't meet strict integration criteria get excluded. We've seen this pattern before, with Google Shopping, Amazon Marketplace, and Facebook Shops all doing versions of it. The platform owns the distribution, sets the rules, and takes a cut. Agentic commerce risks repeating the same consolidation under a different name, with AI platforms replacing search engines as the gatekeepers of consumer attention.

Fragmentation is real, not theoretical. Not all AI platforms currently talk to all retail systems. A merchant who integrates with ACP for ChatGPT still can't reach customers using Claude, Gemini, or any other AI agent framework (many of which speak MCP rather than ACP) that hasn't adopted the protocol. Cross-platform interoperability is aspirational at this point. I've audited enough multi-platform setups to know that "works everywhere" rarely does on day one.

Cost is the quiet variable. Backend engineering for async flows, PSP integration, feed maintenance, and ongoing compliance with protocol updates. None of this is free. For a mid-market merchant running on thin margins, it's not obvious yet whether any of this is worth the investment.

The Future of Merchant Control

Merchants who prioritise data portability and API maturity will hold onto more control than those who've been focused on how their storefront looks. The store isn't dying. It's shifting to wherever the intelligence lives, whether that's inside a chat window, a voice assistant, or an agent running quietly in the background.

Treat your product data as the product. Structured, accessible, accurate. Those merchants will show up when an AI agent queries a catalogue at 2am on behalf of a customer who never opened a browser tab. Everyone else will be hoping the old search bar still works.

Start with your APIs. Not next quarter. Now.