When an AI shows up to read your site for a customer, what kind of experience does it actually have?

It's a question we've been asking a lot lately at MCD, as the way folks search and find information continues to push towards agents. It's news to no one that instead of searching Google and clicking through a few sites, a growing majority of people are turning directly to ChatGPT or Claude for their results.

That shift is bigger than it sounds. Industry trackers like Ahrefs have measured a steady drop in click-throughs as AI-generated answers resolve more questions on the spot, and a growing share of searches now end without anyone clicking out to a website. For a brand, the first impression is increasingly something you don't control. It's whatever an AI managed to piece together about you from your content. Which is exactly why what that AI finds, and the experience you give it, is worth getting right.

It's easy to treat this as a brand-new discipline, but it really isn't. UX has always been about reducing friction for whoever is trying to get something done, and the only thing that has changed is who that someone is. There's even a name for it now. Netlify's CEO, Mathias Biilmann, introduced the term "Agent Experience" (AX) in 2025, arguing that the principles we've applied to human users for two decades now apply to AI agents too. We're not trying to claim that term. Plenty of people have written about it already. We're interested in one specific corner of it that tends to get overlooked.

The idea that stuck with our team is that an AI agent isn't so much a new audience as a stand-in for one. Someone gave it a task, like "find me the best option" or "is this a fit for me," and it went off to do the reading. It has no agenda of its own. It's carrying a persons intent along.

That changes the stakes in a useful way. If you give the agent what it needs, you're really serving the person who sent it, so good agent experience turns into good user experience with just one step removed. And if you hand the agent content it can't parse or trust, the person waiting on the other end gets a worse answer about you and never learns why. You won't see it happen. It just shows up later as a customer who quietly went somewhere else.

If you work in marketing, some of this will sound familiar, because there's already a fast-growing field around it called GEO and AEO, short for Generative and Answer Engine Optimization. It's the practice of getting your content cited and recommended by AI systems when people ask them questions, and if your team isn't paying attention to it yet, that's worth fixing.

It's worth noticing the lens, though. GEO and AEO grew out of SEO, and they tend to inherit SEO's main concern, which is visibility. That matters, but it treats the agent as a placement to win rather than a user to design for. Meanwhile the experience side of our industry, the people who think about users all day, has barely connected its own craft to any of this. In our experience the interesting work happens when you bring what UX already knows about designing for an audience into the room where everyone has been optimizing for visibility. When you do that, you land on something that looks a lot like personalization.

Good personalization was never really about formatting. It's about showing the right content to the right audience based on what that audience actually needs. We already do this for people. A first-time visitor and a returning, high-intent buyer don't get the same page, because they aren't looking for the same things.

An agent is simply another audience in that mix, with needs of its own. Once you see it that way, the approach follows naturally. You write for the agent the way you'd write for any audience whose needs are distinct.

This is where a lot of "make your site AI-friendly" advice stops a little short. The common technique, content negotiation, serves the agent a cleaner format of the same content, structured text instead of cluttered HTML, so the model spends fewer tokens and makes fewer mistakes. That "could" be useful (I'm still not convinced), but it's still just a better wrapper around the same content.

What we've been piloting goes a step past format and into depth. An agent can comfortably take in context that would overwhelm a person. For example this could be an exhaustive FAQ, the full specification table, the edge cases, or the honest "here's who this is and isn't for" detail you'd never lead a human landing page with. So the person gets the clean, confident version of a page, and the agent gets the longer, fuller one. It's the same product and the same facts, written at a depth that suits a different kind of reader. That's what we mean by Agent Personalization.

This isn't keyword stuffing. We're not salting pages with terms to trip a matching algorithm. It also isn't ranking in the search-results sense. We're not trying to climb a list of blue links or win a position on a page.

This is about how we present the same content more efficiently, along two axes at once. Structurally we want to give the agent clean, well-organized content it can parse without guessing, so it understands you accurately instead of approximately. Contextually we want to take advantage of the agents ability to absorb far more context than a person wants to read, and give it that fuller depth when it asks for it. Better structure and the right level of detail, matched to the reader. That's the whole idea.

The reasonable worry, if there is one, is whether "different content for agents" create to different of a message or experience. It's a fair thing to keep an eye on, and the line is simple to hold. The agent should get more of the same truth, never a different version of it. More depth, more structure, more of the context a person wouldn't want to wade through is all fair game. What you don't do is let the agent's version say something the human's version wouldn't, with different claims or prices or facts. As long as the agent is getting a fuller, better-organized account of the same reality your customers see, you're presenting content well, not hiding it. That's an important line worth holding to, especially for anyone trusting you with their brand.

For a long time, messy information architecture was survivable, because people are forgiving. They squint past clutter, guess what a vague label means, and keep scrolling until they find what they came for. A confusing site cost you a little conversion, not your place in the conversation.

Agents aren't forgiving in that way. They don't guess at your intent (although they're getting better at it) or give you the benefit of the doubt. If your content is well structured, they understand you. If it isn't, they approximate you, or move on to a source that's clearer. The craft of building clean, efficient experiences, the part that was easy to deprioritize when humans would tolerate a mess, is now what decides whether you're even legible to the system shaping your customer's decision. If anything, that craft matters more now than it used to.

We're piloting this now with a few clients across different verticals, and taking advantage of our expertise in headless content management platforms to quickly deploy our tests.

The majority of our clients run on these headless stacks, where content is kept separate from presentation, and many are already using it to serve personalization to their customers. The same separation that lets a team deliver a tailored experience to different human audiences is exactly what lets us serve a clean, structured version to an agent and a rich visual experience to a person from a single source of truth, without maintaining two parallel websites. Platforms like Contentful are even building personalization directly into the product, which means the muscle for serving the right content to the right audience is already there. Pointing it at a new kind of audience is a smaller step than it sounds. The teams that invested in flexible architecture are the ones who can move on this now.

Working with platforms like these is also what makes the delivery practical. Because we can segment agents from people based on an agent's header information at the edge, we can serve an agent-facing layer of content without touching the human experience. That's where the depth comes in: expanding FAQs, adding context to common questions, filling in product detail, or publishing an llms.txt that points agents to the authoritative source.

So far, early signs are encouraging. Agents are surfacing noticeably more accurate detail in our test queries across ChatGPT and Claude. To be clear about where this stands, it's all small-scale pilot testing right now, and we're still working on capturing more rigorous, quantitative results.

We're also working through some real questions. How do we define and measure success? How do we surface these changes in an analytics layer that wasn't built to see agent traffic? How often should the agent-facing content refresh, and where does added depth stop helping and start becoming noise? We'd rather share this while it's in progress than package it as a tidy case study, because it isn't finished, and that's the part worth talking about.

AI agents are a real audience today, not a few years out. They're already reading on your customers' behalf and shaping a first impression you no longer fully own. The teams that handle this well will treat the agent the way they've always treated their best users. Give the costumer the content it needs, structure it clearly enough to be understood, and keep it honest enough to be trusted.

This is the work we're in the middle of at MCD. If you're looking at your own site and wondering what an agent actually sees when it arrives, or whether your current stack could even do this without a rebuild, we're always happy to compare notes.