The Customer Who Never Visits Your Website

Picture a Tuesday evening. Someone opens an app and types: “Trail running shoes, wide fit, under $130, delivered before Saturday. Use my usual card.” Ninety seconds later a push notification asks them to approve a purchase. They tap “Yes.” Done.

They never saw a homepage. They never scrolled past a hero banner, never read a product description, never noticed a retargeting ad. A piece of software did the browsing, the comparing and the checking out. That scenario stopped being a keynote demo this month. And if you sell anything online, it quietly rewrites the rules you’ve been playing by for twenty years.

September 2026: The Month Agentic Commerce Got Real

A lot happened in three weeks:

Sept 8 - Meta launched Muse, a consumer AI agent that doesn’t just answer questions but acts: it manages email, books reservations, audits subscriptions, tracks prices and buys things (checkout runs through Stripe Link). It reportedly passed 2.5 million downloads in 13 days . Mid-September - Shopify plugged into Muse so agents can check out via Shop Pay, and PayPal announced support for Muse-powered shopping.

Sept 20 - Amazon started blocking Muse. Shoppers saw a pop-up saying access by an “unauthorized AI agent” violates Amazon’s Conditions of Use. Amazon says Muse didn’t identify itself and handled customer credentials in ways it wouldn’t accept. It’s the latest in a series: Amazon has pushed back on Google and OpenAI shopping agents too, and sued Perplexity over its Comet browser in late 2025.

Sept 22 - Six banks (ING, NatWest, ASB, Bank of America, Capital One and Commonwealth Bank of Australia) published a joint set of Agentic Commerce Principles covering transparency, safety, privacy, choice and interoperability.

Take any one of these alone and it’s a news item. Put them together and you get the shape of a new market: consumer agents that can pay, platforms picking sides, and banks trying to set the rules before somebody else does.

Three Things Changed at once

It helps to split commerce into three jobs, because agents are taking over all three, and each one hits businesses differently.

1. Discovery: who finds the product?

Search used to mean putting a link in front of a human. Then generative AI arrived, and the goal became getting a model to mention your brand. Now the agent doesn’t hand over a list at all. It reads catalogs, checks stock, compares prices and shortlists on its own. The “search results page” becomes an internal ranking you’ll never see. ** 2. Decision: who picks the winner?** 3. A human can be won over by a good photo or a clever headline. An agent is won over by clean attributes, accurate prices, real delivery dates, return policies it can parse, and reviews it can trust. Persuasion gives way to legibility.

3. Execution: who presses “buy”?

This is the newest part, and the one with the most money and liability attached. Letting software move money for you needs new plumbing: scoped tokens, signed intent, agent identity, and a clear answer to “who pays when it goes wrong?”

The Protocol Map (It’s crowded)

If you’ve lost track of the acronyms, here’s roughly how the stack looks today:

• ACP (Agentic Commerce Protocol) - OpenAI and Stripe. Powers ChatGPT Instant Checkout. The key idea is a Shared Payment Token: single-use, tied to one merchant, one amount, and a time window.

• UCP (Universal Commerce Protocol) - Google and Shopify’s open standard for agents to discover products and complete checkout across merchants. Shopify made agent registration self-serve in June.

• AP2 (Agent Payments Protocol) - Google’s payment layer, launched with 60+ partners. It uses cryptographically signed “mandates” so there’s proof the human actually approved this purchase.

• Visa Intelligent Commerce & Trusted Agent Protocol and Mastercard Agent Pay - the card networks’ answer: agent-specific tokens and ways to tell a trusted agent from a malicious bot.

•MCP and A2A - not commerce protocols as such, but the connective tissue that lets agents call tools and talk to other agents.

• x402- HTTP-native stablecoin payments, aimed at machine-to-machine micro-transactions.

The lesson isn’t “pick one.” Most merchants will reach these through their commerce platform and payment provider. What matters is knowing which doors you’ve opened, to whom, and on what terms.

Walled Gardens vs. Open Doors

Amazon and Shopify have now taken opposite bets, and that split is the most important strategic story in commerce right now.

Amazon’s bet: the customer relationship, and the ad business built on it (well over $60 billion a year), is worth defending. If an outside agent does the shopping, Amazon loses sponsored placements, cross-sell and first-party data. So outside agents are kept out, and Amazon pushes its own (Rufus, Buy for Me).

Shopify’s bet: millions of independent merchants gain more from being reachable by every agent than they lose in control. So Shopify is making itself the easiest place for agents to transact.

Put simply, the fight has moved on from “can agents pay?” to “who controls discovery, authorization and execution?” For a brand, agentic commerce is a permissioned channel. Your visibility is only as good as the access you grant.

The Trust Gap is the Real Bottleneck

The adoption numbers look huge until you split them apart:

• About 132 million Americans have bought something with AI’s help (PYMNTS).

• Yet only 14% trust an AI to actually place the order (higher among Gen Z and millennials, at roughly 30%).

• 59% of AI-assisted purchases still end at Amazon.

So AI has won discovery but hasn’t yet won checkout. That’s why Muse makes users approve every irreversible action with a push notification. It’s also why banks are writing principles, and why AP2’s signed mandates exist.

Security is the other half of trust. The biggest unsolved problem is prompt injection: instructions hidden in a web page, product review or email that an agent mistakes for commands. Meta’s own guidance, the “Agents Rule of Two,” says an agent should never combine all three of these in one session:

1. Reading untrusted content,
2. Accessing sensitive data (credentials, payment methods), and
3. Taking actions that change things or send data out.

Any two is manageable. All three together is where a malicious product listing can talk an agent into leaking your card or buying the wrong thing. This is a mitigation, not a cure, and the industry knows it: Accenture found 78% of financial institutions expect AI-agent-linked fraud to rise, and Visa has reported sharp spikes in malicious bot transactions.

The Numbers worth Anchoring on

Forecasts vary, but they all point the same way:

Bain: $300–500B in US agentic commerce by 2030 (15–25% of e-commerce).

Gartner: 20% of digital commerce transactions executed through AI platforms by 2030.

McKinsey roughly $1T in US retail and $3–5T globally by 2030.

Adobe AI-driven visits to US retail sites grew about 4,700% year over year in 2025 . **Shopify **(early data): AI-assistant traffic converts about 50% better than organic search, with ~14% higher order values.

Treat the 2030 figures as direction, not destination. The conversion data is the part to act on now: agent-referred shoppers arrive further down the funnel and with more intent.

A Practical Playbook for Brands and Agencies

Here’s what I’d do this quarter, before the holiday rush turns agentic shopping from experiment into habit:

1. Treat your product feed as your storefront. Titles, variants, sizes, materials, compatibility, stock levels, shipping cut-offs, return windows: everything structured and machine-readable. Pages with structured data are cited ~3x more often in Google’s AI Overviews. For an agent, missing attributes mean you’re invisible.

2. Write an explicit agent policy. Decide which agents may read your prices, which may check out, and under what identity. Allow, block or negotiate, but decide. Look at agent-verification standards (Web Bot Auth, Visa TAP) rather than blanket bot-blocking that also locks out paying customers.

3. Integrate through your platform, not from scratch. If you’re on Shopify, UCP and Shop Pay routes are largely switched on. If you’re on Stripe, look at the Agentic Commerce Suite and ACP. Your job is configuration and governance, not building protocols.

4. Fix attribution before it breaks. Agent purchases often show up as “direct” traffic paid with virtual or tokenized cards. Tag agent-originated orders, track them as their own channel, and start measuring share of agent recommendation the way you once measured share of search.

5. Make price and inventory accuracy a KPI. An agent that gets burned by a phantom “in stock” or a price that changes at checkout will learn to skip you. Humans forgive stale data. Agents quietly drop you from the shortlist.

6. Design for conversation, not just clicks. Messaging-based agents (WhatsApp Business agents, for one) are moving to usage-based pricing: per message and per million tokens. Short, well-designed flows that spot purchase intent early and hand off cleanly to local payment methods will beat long, chatty bots on both cost and conversion.

7. Red-team your own content. Assume someone will plant instructions in your reviews, Q&A or marketplace listings to hijack agents. Moderate user-generated content with that in mind, and never let your own customer-service agent combine all three legs of the Rule of Two.

8. Keep a human brand worth choosing. When agents shortlist on specs and price, brand still decides ties and drives the “Buy from X” instructions humans give their agents. Being the brand a customer names out loud becomes the ultimate moat.

What to Watch over the Next 90 Days

Holiday 2026: the first peak season with mainstream consumer agents that can pay. Watch conversion, return rates, and how many shoppers actually let agents check out.

Amazon vs. everyone: does the blockade hold, turn into paid access deals, or end up in court?

Agent access fees: expect platforms to start charging agents for catalog and checkout access, a new “toll road” line item for marketing budgets.

Liability and regulation: who refunds a purchase an agent got wrong? The bank principles are a start, and regulators will follow.

Protocol consolidation: too many standards is a temporary state. Watch which ones payment providers and platforms make the default.

The Bottom Line

For two decades, digital marketing meant winning a human’s attention on a screen. Agentic commerce adds a second customer: a machine that doesn’t see your ads, doesn’t care about your homepage, and reads your data with ruthless precision.

The brands that win this shift won’t necessarily be the loudest. They’ll be the most legible, the most trustworthy, and the most deliberate about which agents they let through the door.

Your next customer may never visit your website. Make sure they can still find you.

Author: Mohamed Haseel, Chief Architect and Head of Technology Practices

This blog was first published on Medium.com. Watch this space for the next one in the series.