At the start of 2025, about 1 in 200 visits to a website came from an AI bot. By the end of the year it was 1 in 31. Over those same months, human traffic actually fell quarter over quarter.
I want to be clear about what that number is saying. The composition of who, or what, visits your website is changing in real time. The next major user of the internet isn’t going to be a person. It’s software acting for a person. And that’s not a forecast I’m asking you to take on faith. It’s already happening, and you can watch it in the traffic, in the money flowing into agent infrastructure, and in the standards war over how agents buy things.
This is the agentic web. Here’s what it is, why it’s moving faster than the coverage suggests, and what it changes if you actually run a business.
What is the agentic web?
The agentic web is the version of the internet where AI agents, not humans, do a growing share of the browsing, deciding, and buying. An agent is software you give a goal to. You say “find me trail-running shoes under $150 that arrive by Friday,” and it handles the discovery, the comparison, the checkout, and the follow-up. You set the parameters. It does the work.
For thirty years we built everything for a human deciding where to click. Websites, storefronts, ads, checkout flows, all of it assumed a pair of eyes and a cursor. The agentic web breaks that assumption, because the thing on the other end is now a program that just wants the data and the action.
The traffic already shifted
The 1-in-31 number comes from web traffic data across 2025. On the shopping side it’s even starker: Adobe measured generative-AI traffic to US retail sites growing roughly 4,700% year over year. That’s not a rounding error. That’s a new audience arriving while most companies are still optimizing for the old one.
Aaron Levie, Box’s CEO, put the mechanism plainly: agents won’t use your interface, they’ll talk to your APIs. Which means software eventually has to go “headless.” The buttons and menus we built for human eyes become optional, because the actual user is a program. Levie goes as far as calling agents the primary users of the internet, and he isn’t alone. Yann LeCun and Zapier’s Wade Foster have made versions of the same argument.
The money backs it up. People are building infrastructure specifically for agents. Anchor Browser is a cloud browser whose entire job is letting agents use websites in secure, logged-in environments. Onyx Security raised $40 million to govern the agents running inside companies, treating an agent like an employee that needs least-privilege access and monitoring. And the one that sounds like a joke until you check: Meta acquired Moltbook, a social network built only for AI agents. When a platform like that gets bought, the category is real.
Underneath most of this sits Model Context Protocol (MCP), Anthropic’s standard for letting an agent plug into tools and data the way USB lets a device plug into anything. Google is adding MCP support to its consumer agent. When the plumbing standardizes, the technology is usually about to get serious.
Your website has a new reader, and it isn’t a person
Here’s the part most business owners haven’t absorbed yet. If an agent is the one reading your site, then your site has to be legible to a machine, not just attractive to a human.
This is where answer engine optimization (AEO) comes in. Classic SEO was about ranking for a person who reads a results page and clicks. AEO is about being the source an agent or AI answer engine actually pulls from. Clean structured data, clear product and service information, machine-readable everything. If your pricing, availability, or service details live only inside a pretty layout and not in structured fields, the agent can’t reliably use them, and you get skipped before a human ever sees you.
The companies investing in this now are the ones whose operations will be readable to agents first. Everyone else is building for a reader who’s slowly leaving.
How agents actually buy: the protocol war
The most underreported piece of all of this is that agents are starting to buy things, and there’s a genuine fight over how. This is “agentic commerce,” an agent handling a whole purchase from a goal instead of a click.
The “standards” battle is crowded. OpenAI and Stripe have ACP. Google has UCP, backed by a coalition that includes Walmart, Target, Visa, and Mastercard, plus AP2 with PayPal. Anthropic’s MCP handles discovery. Visa has its own Trusted Agent Protocol. Everyone is trying to own the rails before they’re poured.
The payment mechanics are clever, and worth understanding if you sell anything. OpenAI and Stripe use a Shared Payment Token: a credential locked to one merchant, one amount, single use, time-limited. You’re not handing an autonomous agent your card. You’re handing it a one-time key for a specific purchase. The card networks built their own version, “agentic network tokens,” where you vault your card once and the agent spends within your stated intent. They’ve even bolted buy-now-pay-later onto it through Affirm and Klarna.
There’s also a weirder problem hiding here. A normal card transaction carries a fixed fee around 30 cents. Fine for buying shoes. Fatal when an agent needs to pay tiny amounts thousands of times, per API call or per second. That’s why stablecoin rails like Coinbase’s x402 keep showing up in a conversation that’s otherwise about Visa and OpenAI. They get the per-transaction cost low enough for machines to actually use.
And it left the lab. McKinsey projects AI agents could be responsible for around a trillion dollars in US transactions by 2030. Mastercard and Santander already completed Europe’s first live, end-to-end agent payment inside a regulated bank. Production, not slides.
What the agentic web means for your business
I’ll give you the honest version, because the timeline is where most coverage oversells. The direction is clear. The speed is not. There’s good research showing experienced developers don’t just hand the wheel to AI agents; they stay in control and insist on quality. One randomized trial even found AI slowed experienced open-source maintainers by about 19%. So anyone telling you to fire your team and let the agents run loose this quarter is selling something.
But “uncertain timeline” is not the same as “ignore it.” Here’s what actually changes for an operator:
Your customer might arrive as software. The systems you build now should assume some share of the people evaluating you are doing it through an agent, which means structured, legible, API-accessible information beats a beautiful but opaque site.
Your internal operations are the bigger near-term win. The same agent technology that’s learning to buy things is already good at running repetitive, multi-step work inside a business: intake, routing, follow-up, reconciliation. We did exactly that for [CLIENT NAME — TODO], automating month-end reconciliation across eight co-packing plants running three different ERPs. That’s where the return is right now, and it’s the part you control.
This is the gap I keep noticing in the coverage. Almost everyone is writing about the agentic web from the platform’s side, telling retailers to make their store agent-ready for someone else’s checkout. Far fewer are talking about the business that needs a custom system built around its own operations. That’s the side I care about, and it’s the practical one.
How to prepare your business for AI agents
You don’t need to predict the exact timeline to start. A few concrete moves:
Make your public information machine-readable. Structured data on your products, services, pricing, and availability is the new front door. Treat AEO as seriously as you treated SEO.
Find the repetitive workflow that’s quietly costing you. The fastest payback from agents right now isn’t customer-facing, it’s the multi-step internal process nobody enjoys owning. We build these systems through AI engineering, and you can see one in production in our 4AM Media build, where agents now handle 90% of support tickets and recovered $200K a month, or across the rest of our e-commerce work.
Decide where a human stays in the loop. The research is clear that experienced people keep quality up by staying in control. Design that in from the start instead of bolting it on later.
If you want to see what an agent working on a business’s behalf looks like in production rather than a demo, our case studies walk through the actual builds, the numbers, and what broke along the way.
The web’s next customer might not be human. Right now, in public, everyone from OpenAI to Visa to Coinbase is racing to build the road it’ll drive on. The businesses that start making themselves legible to that road, and that put agents to work on their own operations, are the ones who’ll be ready when it’s paved.
