Nobody wants to visit your website. They want jeans.

Black jeans, three ways

I'm in ChatGPT. Same story in Claude or Gemini. I need black slim-fit jeans, 32×32, under $80. There are three ways this task gets done, and the differences between them tell you more about where the web is going than any protocol announcement.

The web is splitting into three lanes for how AI agents get things done. Lane A finishes the task inside the chat. Lane C lets the agent work a cooperating website. Lane B forces a robot to click through an unready one. The businesses that win will be the ones easiest for an agent to discover, call, and trust.

First way: I ask, a carousel renders inside the conversation, I tap one, confirm, done. (Honest footnote: today the payment step often still hands me to the merchant's page. The rest is already real.) The mechanical friction is gone: no tabs, no page loads, no typing my card into another form. The merchandising is not gone. A merchant with a storefront in the chat can still offer me 10% off for joining the list, still show reviews, still suggest the belt. The difference is that it happens in the conversation, on my terms, instead of as a pop-up blocking the button.

Second way: I hand the task to an agentic browser (ChatGPT Atlas, Perplexity's Comet, Gemini in Chrome) and watch it work a denim store like a patient intern. It works on any store, with zero merchant cooperation. That is its superpower and its ceiling.

Third way: same browser, but the store has told the agent what it can do. The agent calls searchProducts directly against the live page, signed in as me, shows me the three pairs that match, and I pick one before it calls addToCart and checkout. The agent didn't get more autonomous; it got more efficient.

Are websites dying?

No. But the visit is. People are visiting websites less, and browsers are not dying; those are two different things, and only the first one should worry you.

The visit was never the point. Nobody wants to browse your site. They want jeans, a refund, a Tuesday reservation. The website was just the interface we had. Agents replace the visit, not the need.

The browser has one last job: it's the bridge to the part of the web that isn't ready for agents. Look at what the big platforms built. OpenAI shipped Atlas in October 2025, a browser where ChatGPT clicks and fills forms for you. Perplexity shipped Comet. Google put Gemini inside Chrome. They didn't build browsers because a robot clicking through a website is a good experience. They built them because it works on every site today, whether the site cooperates or not. The browser is scaffolding, and scaffolding comes down.

What are the three lanes of the agentic web?

Here's the model, in ranked order. Lane A is where the customer is going. Lane C is the strong second: the cooperative website. Lane B is last: the penalty for doing nothing.

Lane A: the future. The whole thing happens in the chat. The customer starts in ChatGPT and finishes there. How? Nordstrom told ChatGPT what it can do: search inventory, take an order. The plumbing is MCP Apps and the checkout protocol OpenAI built with Stripe, but the names matter less than the fact. Think of the chat as the new Google. For twenty-five years, customers found you through a search box someone else owned, and "find me jeans at Nordstrom" still works the same way here. The big change is what happens next. Google handed the customer to your website: your merchandising, your email list, your upsell. In the chat, there is no handoff. The whole purchase can finish inside the conversation. Today that looks like a simple product card. Tomorrow it can look like your storefront inside the chat: your reviews, your styling advice, your loyalty program. Either way, the website visit is gone. How much of your brand lives in this new channel depends on what you build.

Lane C: second place. Your website cooperates with the agent. The customer still comes to your website, but the agent does the work there. Your site tells the agent what it can do (search, add to cart, check out), and the agent calls those actions directly on the page instead of clicking around like a human. This is what WebMCP is. Your website is still core to the experience. Your analytics still see the visit. You decide what the agent is allowed to do. It's a real strategy, and for businesses whose experience is their edge, a defensible one. But be clear about the bet: you're betting customers will keep wanting to come to your website when the easier option is "just handle it."

Lane B: last place. An agent puppets your website. The customer watches a robot do what they used to do: scroll, click, fill, wait. It breaks on layout changes, modals, and CAPTCHAs, and it's expensive to run. It exists for one reason: it works on any site without the site's permission, because it rides the customer's own logged-in session. Lane B is not a strategy. It is what happens to businesses that never got ready. Your customers won't see your brand; they'll see an agent struggling with your checkout flow, and they'll blame you for the wait.

Why hasn't Lane A simply won yet?

Because the pace is set by cooperation and permissions, not by which technology is best. Three forces govern it.

Adoption. The biggest companies will connect directly. They have the engineers and the customer volume to justify it. The millions of smaller businesses get there a different way: the platforms they already run on turn it on for them. Shopify flipping one switch for millions of merchants matters more than a thousand individual integrations. Until that happens, agents work the unready web through a browser. Lane B is the coverage plan.

Permissions. Discovery is getting solved. Draft proposals let agents auto-discover a site's capabilities through well-known endpoints and DNS, the way crawlers once found sitemaps. But finding a capability is not permission to act as you. Everything interesting lives behind a login, and delegated consent for agents is the least-built piece of the entire stack. That gap is exactly why Lane B's session-riding is so seductive. It skips the consent problem by inheriting your cookies. Convenient, and contested in court.

Incentives. Some businesses will not become a clean capability call, on purpose. Not because integration is hard; development cost is collapsing. Because the store itself is the business model. Could Amazon ship a storefront inside ChatGPT, complete with "customers also bought"? Sure. But a card in a conversation is a thin copy of the real thing. Walk into Amazon for jeans and you wander: the beauty section, the deal of the day, the basket that grows on the way to checkout. A chat answer ends. A store keeps selling. That's why Amazon won't shrink itself into someone else's conversation. And for the holdouts, Lane B is the agent's leverage: the credible "we'll do it without you" that forces a negotiation. The browser survives as the enforcement layer against businesses that would rather not cooperate.

Who owns the customer when the customer is an agent?

Strip away the protocol talk and that is the whole fight. Three giants have placed three different bets.

OpenAI's bet: be the demand layer. Instant Checkout launched in September 2025 on rails it built with Stripe: Etsy sellers live, Shopify merchants onboarding, a shared payment token so the agent never sees your card. Then reality arrived. Almost nobody bought inside the chat. Fewer than 15 Shopify stores ever went live, and OpenAI shut Instant Checkout down in March 2026 (Modern Retail, 2026). Read that failure carefully, because it isn't the channel dying. People ask ChatGPT what to buy in enormous numbers; they just finished the purchase elsewhere. So OpenAI pivoted: merchant-led storefronts inside ChatGPT that keep the merchant's brand, checkout, upsells, and loyalty, plus advertising under Fidji Simo's applications org. Own the conversation; let merchants keep the transaction.

Amazon's bet: refuse. It blocked OpenAI's crawlers in November 2025, removing roughly 600 million products from ChatGPT's shopping results. It sued Perplexity over Comet and won a preliminary injunction in March 2026 on Computer Fraud and Abuse Act grounds, with reasoning every operator should sit with: the agent had the user's permission but not Amazon's authorization. Your customer's consent to their own agent is not enough to enter your store. The Ninth Circuit paused the ruling pending appeal, and arguments were heard June 11, 2026, so the legal floor is wet cement. Meanwhile Amazon builds its own "Buy for Me" agent. Refusal plus a house agent: defend the discovery moment at all costs.

Google's bet: keep you in the browser. Gemini in Chrome, Project Disco (an experimental browser from the Chrome team that replaces the URL bar with a prompt), and its own commerce rails: the Universal Commerce Protocol, announced with Shopify, Target, and Walmart at NRF in January 2026. If agents act through the browser, Google keeps the web economy it built.

It is not settled, and anyone telling you it is has something to sell you. But notice what all three bets concede: the ask is moving into the conversation.

What's real today, and what's still missing?

I'm calling the destination early, so here's exactly which parts aren't built yet. Every claim in this post is anchored to something that shipped.

What should you do about it?

Three audiences, three moves. None of them requires waiting for the standards to settle.

If you run a business or a storefront: assume your customer's next purchase starts in a chat, and make sure you're callable there. Expose your capabilities now, while the standard is soft; the same plumbing serves the chat and your own site. Then decide what you want the channel to show: a bare product card, or a storefront that carries your reviews, your advice, your loyalty program. Doing nothing doesn't keep you out of this. It puts you in Lane B, where customers watch a robot fight your checkout and blame you.

If you lead product: design for the agent as a user, not just the human. Agents don't see your onboarding flow or your carefully tuned empty states. They see what you expose and how reliably it works. Treat agent engine optimization the way you treated SEO in 2005: a discipline, with audits (Lighthouse already ships an agentic-browsing one) and a budget, not an afterthought.

If you're a founder: the money is in what's missing, and three gaps are wide open. First, permissions: there's no good way today for a customer to tell an agent "you can spend up to $200 at this store and nothing else," and whoever builds that trusted permission layer sits in the middle of every agent transaction on the web. Second, receipts: businesses will need to know which agent did what, with whose approval, and whether it worked, the same way they needed web analytics in 2005. Third, the on-ramp: millions of small businesses will never hire developers for this, and the company that makes a salon, a restaurant, or a boutique callable by agents in one click does for the agent web what Shopify did for online stores. And in every category, the first service agents learn to call by default becomes the front door for that whole category. Pick a vertical and become the first call.

Take Family Bugle, a guide to Park City family life. It sells nothing. We made it callable: an MCP server with 23 tools any agent can use, with an app on top so the experience stays branded. A parent asks ChatGPT or Claude where to take her kids in Park City. Family Bugle's branded cards appear in the chat. She filters by age or day. The results update from its live data. No website visit. Full brand. That's the same technology the commerce giants are fighting over, used to help a parent find a swim lesson. Callable isn't a commerce play: it works for any business a customer might ask about. This is the strategy we run for clients at Grow with Bugle.

Stop making people click

The first time the web just answers, something in you unclenches. No tabs. No forms. No dead ends. You ask, and the thing is done. It feels like relief, because it is. Most of what we call "web design" is persuasion architecture, friction built on purpose, and agents are immune to it. Good.

Then comes the second feeling, and this is the one businesses should fear: impatience. Once you've felt the web work this way, every site that makes you create an account to see a price feels like it's stealing your time. Customers won't file complaints about friction. They'll just stop showing up, one ask at a time.

The pieces aren't all here. The permissions layer isn't built, the lawyers are still fighting, and the standard is a draft. But every piece that proves the direction has already shipped: the rails, the agentic browsers, the checkout, the trial. Your customers never wanted your website. They wanted the jeans, the table, the appointment. For thirty years, getting the thing meant working through pop-ups, account walls, and checkout mazes, and people put up with it because there was no other way. Now there is. The businesses that win the next decade will be the easiest ones to ask. The ones that lose will keep polishing a storefront on a street where people no longer walk.

Frequently asked questions

What are the three lanes of the agentic web?

Lane A finishes the whole task inside the chat: the agent calls a business's capabilities directly and the customer never opens a browser. Lane C is a website that cooperates with the agent by declaring tools it can call on the page. Lane B is an agent puppeting an unready website by clicking and scrolling like a human.

Are AI agents making websites obsolete?

Visits are declining, but websites and browsers are not disappearing yet. The browser's remaining job is to be the bridge to businesses that have not exposed capabilities agents can call. Businesses whose experience is their edge will keep destination websites longer.

What is WebMCP?

WebMCP is a draft web standard, co-developed by Google and Microsoft, that lets a website declare structured tools (like search, add to cart, and checkout) that an AI agent can call directly on the live, signed-in page instead of clicking through the layout. It entered a Chrome origin trial on June 2, 2026.

How can a small business get ready for AI agents?

Expose your capabilities through MCP so agents can call you, decide what your storefront inside the chat should look like, and watch the platforms you already run on (like Shopify) for one-switch agent readiness. Doing nothing leaves customers watching an agent fight your checkout flow.