Your Next Reader Is a Machine

On June 3, 2026, the CEO of Cloudflare posted something that should have been bigger news than it was.

Cloudflare sits in the plumbing of the internet — a huge share of the world’s web traffic passes through it. So when Matthew Prince posted on X, he was reading off a dashboard most of us never get to see:

“Welp, that happened faster than I predicted. Thought it would be end of 2027, then early 2027, but agentic traffic growing so fast that bots have now passed human traffic online for the first time in the Internet’s history.”

Back in March, at SXSW, he’d predicted the crossover would happen in 2027. It happened in mid-2026. If you look at Cloudflare’s own Radar dashboard, bots are now around 57.5% of traffic, and humans are down to roughly 42.5%.

For the entire history of the web, the internet was a place humans went. As of this year, the majority of what moves across it isn’t us anymore.

The timing is the weird part

The strange part is the timing. The machines took the majority now, at a moment when most people still don’t use AI in any serious way. Walk into a typical company and you’ll find a handful of curious people quietly using ChatGPT, a leadership team that “knows it’s important,” and a large middle that has maybe tried it twice and decided it was overhyped. I’ve written before about why giving everyone a Copilot license isn’t a strategy — adoption among actual humans is still early.

And yet the agents are already here, at scale, reading everything.

That gap is the whole story. We’re picturing the AI shift as something that arrives when people finally adopt the tools. But the web already flipped underneath us, driven not by humans typing into chat boxes but by their agents — the research assistants, the shopping bots, the coding tools, the things quietly fetching pages on someone’s behalf. One person asking an AI to “compare the five best options and book me the cheapest” can fan out into thousands of page requests. Multiply that across millions of people who’ve delegated a single task, and the traffic math stops being about humans at all.

So the question I keep coming back to is uncomfortable: if more than half of what visits your website is now a machine, who exactly are you building that website for?

Almost everyone is still building for the 43%

Look around. Nearly every site on the internet is still designed, written, and optimized exclusively for the human minority. We obsess over hero images, scroll animations, the perfect call-to-action button. All of it aimed at a person with eyes, a cursor, and a few seconds of patience.

Meanwhile the actual majority of our visitors — the agents — don’t see any of it. They don’t render your animation. They don’t admire your font. They strip the page down to text, pull out what they can understand, and move on. And I have not seen many companies stop to ask whether they should be building something for that visitor: a clean interface, a machine-readable version, a way for an agent to do business with them directly.

This is the same blind spot I see everywhere with AI: we keep imagining the future as a slightly faster version of the present, and we miss the part where the rules underneath have already changed.

Machines don’t read the way Google ranks

There’s a second thing here that took me a while to understand, and it’s the part most people get wrong.

You cannot win with an AI the way you won with search engines.

For twenty years, the game was SEO: backlinks, keywords, page authority, climbing a ranked list of blue links so a human would click you instead of the result below you. That whole model assumes a list, and a human choosing from it.

An AI doesn’t give a list. It reads a pile of sources, synthesizes one answer, and quotes a few of them. There’s no page two. Either the model pulls your content into its answer and names you, or you simply don’t exist in that conversation. It’s less like ranking in a competition and more like being the one source a smart assistant decides to trust and repeat.

And what earns that trust is different from what earned a high ranking. The clearest evidence I’ve found comes from a Princeton research paper called “GEO: Generative Engine Optimization” (Aggarwal and colleagues, presented at KDD 2024). They tested what actually makes a generative engine quote a source. The moves that worked — adding clear citations, quoting credible sources, including concrete statistics — lifted a source’s visibility in AI answers by up to 40%. The classic SEO move of stuffing in keywords? Almost no effect.

Read that again, because it’s the practical heart of all this. The things that make an AI quote you are roughly the things that make a careful human editor trust you: state a clear claim, back it with a real number, name where it came from. Write like you actually know something, not like you’re trying to trick a ranking algorithm.

There’s a tidy irony in that. After two decades of SEO slowly teaching the web to write for robots in the worst sense — thin, keyword-padded, gamed — the machines that now dominate traffic reward the opposite. They reward writing that’s clear, sourced, and honest. Steter Tropfen höhlt den Stein: the web is being worn back toward substance, one agent request at a time.

What about all the “optimize for AI” advice?

There’s a lot of hype in this corner right now, and I’d rather you trust me than sell you something.

In 2024, Jeremy Howard of Answer.AI proposed a standard called llms.txt — a simple markdown file you put on your site that hands an AI a clean, structured map of your content, so it doesn’t have to fight through your navigation, ads, and JavaScript to understand you. It’s a genuinely good idea, and it points exactly where things are going: a web that publishes a human version and a machine version.

But the picture in 2026 is mixed. If you add llms.txt hoping it’ll make ChatGPT or Perplexity cite you more, the evidence so far is weak — the big consumer AI crawlers mostly don’t even request the file yet, and adoption across the web sits around 10%. Where it is quietly doing real work is the agentic layer: the coding assistants and dev tools (Cursor, Claude Code, and the like) genuinely fetch it when pointed at documentation. Google has started checking for it in its Lighthouse tooling. So it’s early, it’s real, and it is not yet a magic switch. Anyone telling you otherwise is guessing or selling.

I’d rather you take the durable lesson than the tactic-of-the-month. The tactics will churn. The direction won’t: more and more of your audience is software, and software needs content it can actually parse and a clean way to act on it.

How I’d think about it (not a checklist of tricks)

If you’ve read my other guides you know I don’t believe in memorizing tricks over building the right mental model. So I won’t hand you a list of ten hacks. Here’s the shift in thinking instead.

Write to be quoted, not just clicked. Make your claims explicit. Put the number in the sentence, not in a chart an agent can’t read. Say where a fact came from. This is good writing for humans too — it just happens to be exactly what machines reward now.

Assume a non-human reader exists, and don’t hide your substance from it. If the only way to understand what your company does is to watch a video or decode a clever animation, an agent will skip you. The plain-text version of your value has to be findable and clear.

For companies: think about the interface, not just the page. This is the bigger, slower shift. Right now there’s real movement toward agents that don’t just read your site but act on it — book the appointment, place the order, pull the policy. The plumbing for that is being standardized fast: Anthropic handed its Model Context Protocol to a Linux Foundation body at the end of 2025, and Google announced a commerce protocol for agents at the start of 2026. Most companies haven’t even asked whether an agent could do business with them. That question alone puts you ahead.

Don’t bet the strategy on any single file or standard. llms.txt might win. Something else might. The safe bet is the posture rather than the format: publish for machines on purpose, watch how agents actually use you, and adjust.

The one thing the machines can’t recycle

There’s a deeper reason I’m not worried about disappearing into a sea of machine-made content, and it might be the most hopeful part of this whole shift.

Left to their own devices, AI systems converge. They pull the whole internet into one averaged-out distribution and hand back the middle of it. This isn’t a hunch — it’s measured. A 2025 study in PNAS (“Echoes in AI”) found that large language models produce strikingly less variety than humans when asked to write stories, and researchers testing whether writing with AI reduces content diversity found exactly that. Worse, when you train AI on the output of other AI, it doesn’t get smarter — it degrades, a failure the Nature paper that named it called “model collapse”. The systems need a steady supply of fresh, genuinely human signal or they drift toward mush.

So original thinking isn’t getting cheaper. It’s getting rarer, and worth more. An AI, by default, recycles what already exists. The genuinely new thought — the argument nobody’s made yet, the lesson from something you actually did, the number from your own data — still comes from a person. Or from an AI that a sharp human pointed somewhere it wouldn’t have gone on its own.

And that lines up with how you get quoted in the first place. Remember the Princeton finding: original statistics and first-hand sources are exactly what generative engines pull into their answers. The machines are hungry for the one thing they can’t produce themselves. So the goal was never to out-publish the bots. It’s to write the thing that isn’t already out there — which, funnily enough, is the same advice I’d have given before any of this started.

Why I think this matters more than it looks

It would be easy to file this under “SEO has a new acronym, ignore until necessary.” I think that’s a mistake, and it’s the same mistake people made when they decided AI was just autocomplete.

The composition of your audience has changed in a way it never has before. The majority of the visitors to the open web are no longer human, and the way you earn the attention of the new majority is genuinely different from everything the last twenty years taught us. Most people haven’t noticed, because the agents are quiet and the dashboards are hidden. That’s exactly why there’s an opening.

This is the kind of advantage that compounds, and that you can’t buy your way into later. The people who start writing for both audiences now — and the companies that start building a door for agents now — get a year or two of head start while everyone else is still polishing a homepage that more than half their traffic can’t even see.

I’m not gloomy about it. If anything, I find it oddly hopeful: the machines that now run the web reward clarity, honesty, and real substance over manipulation. After twenty years of the SEO arms race, that’s not the worst trade.

Your next reader is a machine. It’s reading right now. The only real question is whether you’ve written anything it can understand.