The First Reader Is a Machine

For most of my career, the work has started with one assumption: someone is going to show up. A person lands on a homepage, looks around, and decides in a few seconds whether to stay. We design for that moment. We argue about it in workshops. Whole teams exist to make those few seconds count.

Lately I've noticed how often that moment never happens. Someone asks a question, gets a clean, confident paragraph back, and moves on. The site we built was read, carefully even, just not by them. A model read it, took what it needed, and the person never arrived.

We talk about AI as a tool we use. I think the bigger shift is quieter than that. AI is becoming the place we go. Not a feature inside the internet, but the front door to it. And once the front door changes, everything behind it gets rebuilt to fit.

Here is where I think that leads. Within a few years, most of what gets published online will be written by language models, and most of it will be read first, and often only, by language models. People will still be at both ends of the chain. A person asks a question. A person gets an answer. But the long middle, the part we used to call "the web," will increasingly be machines writing for machines.

The Interface Moved

For twenty-five years, the deal was simple. You typed words into a box. You got a list of places to go. You went to one of them, read what someone wrote, and decided for yourself whether it answered your question. The search engine was a librarian pointing at shelves. The reading, and the judgment, were yours.

Answer engines change the deal. You ask a question in plain language. You get an answer in plain language. The shelves are still there, but you never walk to them. The machine walks for you, reads a dozen pages in a second, and hands you its synthesis.

This was already happening before generative AI. A large share of searches were ending without a click, answered by a snippet or a box at the top of the page. Generative answers simply finish the job. When the summary is good enough, the source becomes a footnote. Sometimes not even that.

And it isn't only search. The same pattern is showing up in the assistant on your phone, the chat window inside your software, the agent that books the trip, compares the products, drafts the reply. Every one of those is an answer engine with a different face. Every one of them reads on your behalf.

So the most important reader of almost anything published today is no longer a person. It's the model that decides whether that thing gets folded into an answer, and how.

Writing for the New Reader

When the reader changes, the writing changes. That's always been true. We wrote differently for newspapers than for radio, differently for radio than for television, differently for all of them once search arrived and we started writing for crawlers and keywords. SEO was the first time we wrote, at scale, for a machine audience. It just wasn't a very smart one.

The new machine audience is smart, or at least fluent, and it rewards different things. It doesn't care about your hero image. It doesn't scroll, so it doesn't care about your layout. It cares about whether your content is clear, structured, specific and easy to extract. It cares whether you state the claim plainly and support it. It cares whether you answer the question someone is likely to ask, in words close to the ones they'd use.

The early signs are already visible. Marketers have started talking about "generative engine optimization." Researchers have published work on which kinds of content get cited more often by AI answers. There are proposals for files like llms.txt, written not for human visitors but for models, a kind of guided tour for a reader who will never see the homepage.

Notice what that means. Organizations are starting to publish things whose primary audience is not human. That's the first half of the loop.

Written by Machines, Too

The second half is supply.

Writing used to be the expensive part. It took time and skill, so there was a natural limit on how much of it existed. That limit is gone. A model can produce a competent article in seconds, and a thousand of them in an afternoon. Studies of new web content already suggest that a large share of freshly published articles are machine-generated, and that share has been climbing fast.

Put the two halves together and the economics are hard to argue with. If the main reader is a model, and a model can write what that reader wants faster and cheaper than any person, then a lot of organizations will do exactly that. They'll use one model to write content shaped for another model to read. Product descriptions, help articles, comparison pages, explainers, FAQs. The connective tissue of the web.

None of this is sinister. Most of it will be fine. Some of it will be better than what it replaces, because a lot of what it replaces was thin and rushed and written to hit a keyword count. But step back and look at the shape of the system. A person asks. A model reads content written by a model, synthesizes it, and answers. The human sits at the edges. The middle runs on its own.

When the Loop Closes

This is where I get cautious.

In an earlier essay I argued that fluent isn't the same as smart. Language models are extraordinarily good at producing the form of good thinking. The sentences land. The structure holds. That's exactly why it's easy to mistake the output for understanding. Now imagine a system where fluency is talking mostly to fluency. Every step sounds right. Very few steps involve anyone checking.

There's research behind the worry. Work published in Nature showed that models trained repeatedly on their own generated output start to lose the edges of what they knew. The rare cases fade. The distribution narrows. The researchers called it model collapse. You don't need to accept the strongest version of that finding to see the general risk. A web that increasingly feeds on itself drifts toward the average. It gets smoother and more confident and less surprising.

And surprising is where the value lives. The odd case. The local knowledge. The expert who disagrees with the consensus for good reasons. The first-hand account. Those are precisely the things a summary tends to sand away, because a summary is built from what most sources say.

I keep coming back to the choice at the center of this whole series: mirror or oracle. Used as a mirror, AI shows us our own thinking more clearly and makes us sharper. Used as an oracle, it hands us answers and we stop asking how they were made. An answer engine is an oracle by design. That's its appeal. You ask, it tells. When most of the public reaches the digital world through an oracle, the question of what it's drawing from stops being technical. It becomes a question about what a whole society takes to be true.

What Gets Lost, and What Gets Scarce

Some things will be lost. Browsing, mostly. The detour that turned out to be the point. The second page of results where you found the strange, useful thing nobody else had seen. The experience of reading a source and feeling, from the way it was written, whether to trust it. Language has always made thinking visible. When we only see the synthesis, we lose sight of the thinking behind each source.

But when something becomes abundant, something else becomes scarce. And scarcity is where strategy lives.

If competent, generic content is free, it's worth nothing. What becomes valuable is everything a model can't produce from what it has already read. Original data. Primary research. Real experience. A clear point of view held by an identifiable person. A framework that organizes a messy problem in a way nobody had organized it before.

That last one matters more than people think. In an earlier essay I argued that expertise is structure: the expert doesn't just know more facts, they know how the facts fit together. Answer engines are very good at retrieving facts and reasonably good at stitching them together in familiar ways. They are much weaker at originating new structure. The organizations that win the next decade of the web will be the ones that keep feeding genuinely new structure into the system, because that's the part the system can't make for itself.

Put simply: in a web written by machines for machines, the scarce thing is a human who actually thought about something.

When Nobody Visits

This is the part I find hardest to write, because it's about my own work.

The digital experience industry is built on the visit. Information architecture, interaction design, content strategy, conversion paths, personalization. All of it assumes a person moving through a space someone designed. If that person increasingly stays inside the answer engine, the space doesn't disappear. Its job changes.

I think the website splits in two. One half becomes source material: structured, precise, written so a machine can understand it and represent it accurately. Less a destination than a reference. The other half becomes the place people come only when the answer isn't enough. When they want to decide, to buy, to trust, to get a feel for who they're dealing with. Fewer visits, but each one matters more, because the person making it has already been briefed by a machine and is arriving to confirm or commit.

That second half is still experience design, maybe the most important kind. But the questions change. Not "how do we get them to the page?" but "what does someone need when they arrive already knowing the basics?" Not "how do we hold attention?" but "what can only a visit give them?"

There's a quieter loss, too. A brand's voice lives on its own pages. Inside an answer, it gets paraphrased. Careful positioning becomes one clause in someone else's sentence, summarized by a model that has also read every competitor. Organizations that ignore this will find their identity flattened into the average. The ones that pay attention will make their point of view so specific and so consistent that it survives the paraphrase.

What This Means for Organizations

If you lead a brand, a content team or a digital strategy, here's how I'd start thinking about it.

Write for two readers. Every piece now has a machine reader and a human reader. The machine needs clarity, structure and a plain statement of the claim. The human needs a reason to care and a reason to trust. These aren't in conflict. Clear writing serves both. Vague writing serves neither.

Publish what only you know. Your proprietary data, your customer patterns, your hard-won lessons, your honest opinions. If a model could have written it from the existing web, it adds nothing to the web, and the answer engine has no reason to cite you.

Make your structure legible. Clean information architecture, consistent terminology, structured data, content that answers specific questions directly. You are teaching a reader who will never see your design. Make the meaning survive without it.

Design the visit that's left. Assume the people who still arrive have already been briefed. Skip the introduction they don't need and give them what an answer can't: proof, specificity, a sense of who you are, and a clear way to act.

Measure presence, not just traffic. Clicks will keep falling for a lot of informational content, and that isn't necessarily failure. Take the questions your customers actually ask, put them to the major answer engines on a regular schedule, and track three things: whether you appear, what gets said about you, and whether it's accurate. It's crude today. It's still better than measuring a shrinking channel and calling it performance.

Protect a human core. Use AI to scale the connective tissue if you want. But keep a center of work that is unmistakably authored: named people, real reasoning, a recognizable voice. That core is what gives everything else credibility, and it's what the machines will increasingly lean on.

Decide what you want to be in the loop. Are you a source, or are you noise? Most organizations haven't asked. The ones that do will make different choices about what they publish and why.

Possibility, Not Prediction

I've written this as if the outcome is settled, so let me be honest about what isn't. Answer engines could stumble on trust. Regulation, licensing and lawsuits over training data could reshape the economics. People may decide they miss the open web and go back to it, the way some people went back to vinyl. Models may get much better at telling original work from recycled work and start rewarding it heavily, which would be the best outcome of all.

But the direction is clear enough to plan around. The interface is becoming conversational. The first reader is becoming a machine. The cost of writing is going to zero. Those three facts all pull the same way.

The question I'd ask isn't whether that happens. It's whether, at the very start of the chain, anyone is still doing the thinking. Language models are remarkable at carrying meaning forward. They are not the place meaning comes from. That still has to be us.

The front door has moved. Make sure there's something worth finding behind it.