The Agent-Shaped Web
AI Overviews now eat 58% of clicks. Responsive design optimizes for every screen size humans own. Agents don't care about either. What does an agent-first web actually look like?
Viewport Gymnastics
A friend sent me a motorcycle site the other day. I pulled it up on my 49-inch monitor and the hero image looked blurry. My Webflow instincts kicked in—is this a bug I’ve seen before? I didn’t even scroll at first. Just started poking at the page, resizing the window, watching the layout stretch and compress.
It’s actually a cool site. Not putting it down. But the investigation got me thinking about all the random nits I’ve dealt with over the years. Mobile edge cases. Ultrawide breakpoints. Container queries to make grids not look sad at certain widths. The hours spent on details nobody consciously notices.
In 2023, I jokingly tweeted about viewport fragmentation.
Doing my daily viewport gymnastics 🤸♂️
"Instead of making design decisions on strict, limited breakpoints, keep in mind the sheer amount of fragmentation there is in viewports."
Good reading:
viewports.fyi
The data from viewports.fyi showed over 2,300 unique viewport sizes. Responsive design had become a combinatorial nightmare.
That time isn’t wasted. The craft matters. This site I’m building now is full of opinionated edge cases—problems so seemingly irrelevant that you’d never notice I spent hours on them. But it’s an experiment in starting with the right primitives, the same way you’d write code with the right principles. Content first. Structure first. Then iterate on the vessel.
The thing is: the medium for consumption is changing. Not that you stop caring about the human experience—but you might need to shift focus.
Agents don’t care about any of this. They don’t see hero animations. They don’t appreciate grid systems. They read content, extract what they need, and summarize it for someone who never visits at all.
58% Gone
Ahrefs published updated data this week. AI Overviews now reduce position 1 CTR by 58%. Nine months ago, in April 2025, that number was 34.5%. The trajectory is steeper than anyone projected.
For every 100 clicks that historically went to the top search result, Google now keeps 58. Your site still ranks first. Users just don’t need to visit anymore. The AI read it for them.
Easy to frame this as Google hoarding traffic. But AI Overviews are genuinely useful. Search results have been SEO-gamed for years. The first page is often content farms optimized for algorithms, not humans. Recipe blogs with 2,000 words of backstory before the ingredient list. Listicles that exist solely to rank for long-tail keywords. AI Overviews cut through that.
People don’t read. They skim, scroll, bounce. Now they ask agents, and they read what the agents say.
Your beautiful landing page? The agent summarized it in three bullets. Your carefully crafted conversion funnel? The user got the answer without entering it. Your A/B tested CTA button with the exact right shade of orange? Never rendered.
The marketing industry spent two decades optimizing for human attention patterns. Heat maps. Eye tracking. Above-the-fold optimization. All of it assumes a human is looking at your page. When the agent is the one looking, none of that matters. The agent doesn’t care about your visual hierarchy. It parses your DOM, extracts structured data, and moves on.
Natural Language Everything
Guillermo Rauch posted something today that crystallized what I’d been circling around:
Terminals used to accept shell commands, now it’s natural language.
IDEs used to accept code, now it’s natural language.
Search engines used to accept keywords, now it’s natural language.
Your sites & apps accept clicks, drags, hovers, taps… soon it’ll all be natural language
Terminals went from arcane commands to natural language shells. IDEs went from menu-diving to “explain this function.” Search went from keyword stuffing (“best restaurant NYC Italian cheap 2024”) to conversational queries (“where should I eat tonight?”).
Websites are the last holdout. We’re still building elaborate input systems for humans—hover states, tap targets, scroll animations, hamburger menus, infinite scroll, parallax effects, microinteractions. All the tricks that make the web feel alive. Meanwhile, the actual information retrieval increasingly happens through agents that bypass all of it.
Influencer marketing is the extreme case. Performative monkeys creating dopamine hits for scrolling addicts. It works because humans are susceptible to parasocial relationships and FOMO. We buy things because someone we feel connected to uses them, even though we’ve never met them and they were paid to hold the product.
Agents don’t experience FOMO. They don’t care about parasocial dynamics. They don’t feel inadequate when they see someone’s curated life. They just want structured data they can parse and summarize. “What’s the best protein powder under $50?” The agent doesn’t care that a fitness influencer recommended it while doing a sponsored workout in their home gym. The agent cares about the protein content, price, and third-party testing results.
The influencer economy is human-shaped in a way that doesn’t translate to agent consumption. Structured data does. Product specifications, verified reviews, comparison tables—these are the inputs agents actually use. The emotional manipulation layer that makes influencer marketing work becomes invisible to machine readers.
Building for Agents
Your content should be able to stand on its own without a rendering layer propping it up.
Structured content is the scaffolding. Markdown, frontmatter, metadata, schema markup. Start with something that makes sense without a browser, then build the human experience on top. The mechanisms for serving this to agents are already multiplying, from content negotiation at the CDN to structured entry points to API endpoints that skip rendering entirely. No single standard has won, and it probably doesn’t need to. The point isn’t which file you add or which header you set. It’s that your content carries its own meaning.
When you mark up your product page with proper schema, agents can extract price, availability, and specifications without guessing. When you structure your FAQ with explicit question-answer pairs, agents can answer user queries directly. Structured data has always helped search rankings. We just treated it as an optimization rather than the main event.
Humans read this structured content too. They already do, just not raw. An LLM trains on it, or an agent pulls a section by header and streams it back as rich text in a chat. The original form disappears, but the content is still doing work. Content written “for agents” ends up in front of people anyway, reshaped by whatever touched it last. The rendered website is one shape among several. The agent-facing version might end up being a genuinely different shape, not just a stripped-down copy. But the structured content underneath is where both start.
But What About Design?
What about brand? Design? The emotional experience of visiting a website?
Brand still matters when humans visit. They just visit less. Design becomes a rendering layer on top of content, not the starting point. Your design system inherits from your content, not the other way around.
Think about it like responsive design, but inverted. We used to design desktop-first, then adapt for mobile. Now we design content-first, then render for humans.
The pretty landing page isn’t dead. It’s just no longer the primary interface. It’s one of several ways your content gets consumed, and increasingly not the most common one.
Messy Middle
We’re in the messy middle right now. You still need a website because humans still visit. You can’t go full agent-only because the infrastructure isn’t there yet. Most sites aren’t structured for agent consumption, and most agents are still brute-forcing their way through HTML.
But you can see which direction the wind is blowing. Build for agents first, render for humans second. Not because humans don’t matter, but because the agent layer is where discovery increasingly happens.
It’s the same pattern as the mobile transition. In 2010, you had a desktop site and maybe a mobile site—two separate codebases, two separate experiences. By 2015, responsive design meant one codebase, multiple outputs. You designed once and rendered appropriately for the device. By 2020, mobile-first was the default mental model. You started with the mobile constraints and scaled up.
We’re at the 2010 stage for agents. Separate concerns, unclear best practices, lots of experimentation. Some people are adding llms.txt files. Others are investing heavily in structured data. A few are rethinking their entire content architecture from scratch.
In five years, agent-first will be the default mental model. In ten, we’ll laugh at the idea of building a website that agents couldn’t parse—the same way we now laugh at sites that aren’t mobile-responsive.
Two Vessels
That obsessive craft still matters. Content needs a people-shaped vessel. That’s where design and branding and taste come in. The pretty landing page isn’t dead.
But there’s a parallel vessel now, and it’s agent-shaped. Structure. Readability. Organization. Utility. These two might end up looking nothing alike, built for fundamentally different kinds of readers.
The structured content underneath is where both start. Not because it’s the canonical version, but because it’s good scaffolding. Get the bones right, and both vessels have something solid to build on.
The pretty landing page isn’t going anywhere. Humans will still visit, and they deserve a good experience. But it’s no longer the only vessel. It’s one of two.