Nobody wants to read an AI-written Economist
As artificial intelligence reshapes the pathways between readers and information, The Economist is quietly constructing two parallel presences on the web — one crafted for human eyes, another stripped to its bones for machine intermediaries. This is not merely a technical adjustment but a reckoning with a deeper shift: discovery itself is migrating from homepages and search bars into the conversational interfaces of AI agents. The publisher's wager is that staying visible in this new landscape requires building for both audiences simultaneously, while ensuring that what subscribers ultimately pay for — human judgment, editorial voice, and earned trust — remains irreducibly at the center.
- AI agents are quietly displacing traditional search as the first point of contact between readers and publishers, making agent-readability a survival requirement rather than an innovation.
- The Economist is racing to build stripped-down, machine-legible versions of its content and sales pages before technical invisibility becomes a competitive liability.
- Inside the organization, small cross-functional pods armed with AI tools collapsed a years-delayed CarPlay app into a five-month early launch — proving speed is achievable but raising harder questions about where that saved time actually goes.
- The publisher is drawing careful lines: AI handles research, workflow, and utility, while editorial voice and authorship remain human — a balance that is easy to declare and difficult to sustain at scale.
- The real prize is not efficiency but new revenue categories — newsletters, data products, agentic subscriber tools — that only become viable when AI shortens the distance from idea to launched product.
The Economist is building two separate webs. One is the familiar, feature-rich experience designed for human readers. The other is a skeletal, machine-readable version — plain text, clean Q&A structures, no decorative layers — built for the AI agents that increasingly mediate how people find information. Josh Muncke, the publisher's vice president of generative AI, frames the shift plainly: discovery is changing, and being findable in agent-mediated searches is no longer optional.
The immediate pressure comes from B2B buyers, a growing share of whom now begin their research inside a large language model rather than a search engine. That means The Economist needs two parallel versions of its sales and marketing presence — one glossy and human-facing, another structured for machines. Experiments remain cautious and confined to content already outside the paywall, with internal sandboxes used to work out accuracy, tone, and voice before anything is released more widely.
Media consultant Alessandro de Zanche calls agent optimization a defensive baseline — something every quality publisher will eventually build, because the alternative is technical invisibility. But he draws a sharp distinction: discoverability is a technical problem, while retention is a business problem. Agents can surface content; they cannot manufacture the trust that subscriptions depend on.
Internally, the transformation runs deeper. Muncke reorganized his team into small cross-functional pods — designer, engineer, product manager, editorial staff — working together at what he calls AI speed. The CarPlay app, long stranded on the roadmap, became the test case: with AI handling tests, documentation, and boilerplate code, the pod shipped five months ahead of schedule and recorded roughly 8% efficiency gains in parts of the process.
That sprint is now a template across six to eight pods. Editorial staff are embedded in teams touching reader experience to ensure AI-powered features still sound like The Economist. A quiet vibe-coding culture has taken hold — editors on the science desk now write utilities that trawl academic journals, product teams generate automated performance reports, and a Chief of Staff agent drafts responses and surfaces daily priorities. Not everything worked: an automated copy checker built from the house style guide was paused, and AI companions for live subscriber events were shelved after testers found them distracting.
Muncke is unambiguous about the limits. Nobody wants to read an AI-written Economist. The technology is infrastructure — for research, workflow, and delivery — not authorship. The publisher has committed to clear labeling wherever AI is used. The real test ahead is whether it can spread these skills beyond early enthusiasts, surface the best internal experiments, and keep human judgment at the center of what subscribers are actually paying for.
The Economist is building two separate webs. One is the familiar thing readers see: glossy, feature-rich, designed for human eyes scrolling through a homepage or search results. The other is stripped down to its skeleton—clean Q&A structures, plain text, no carousels, no feature art. This second web exists for AI agents, the software intermediaries that increasingly stand between a reader and the content they want to find.
Josh Muncke, vice president of generative AI at The Economist Group, frames it plainly: the publisher is preparing for "a world with two versions of the web." The bet underneath is that discovery itself is changing. People won't start on homepages anymore. They'll ask ChatGPT, Gemini, or Claude a question, and an agent will fetch the answer on their behalf. For a publisher, that means being findable in those agent-mediated searches is no longer optional—it's a baseline requirement for staying visible at all.
The immediate pressure comes from B2B buyers. A growing share of them now start their research in a large language model rather than a search engine. The Economist's sales and marketing pages need to surface cleanly in those agent responses. That means building two parallel versions of the same pitch: one glossy and comparison-heavy for humans, another Q&A-style and structured for machines. It's not a side project anymore. It's part of the go-to-market plan.
Muncke is careful about what gets exposed. The Economist is a subscription publisher, which means every decision about what content sits outside the paywall carries real financial weight. Right now, the experiments are tentative and confined to content that already lives in front of the paywall. The publisher is using internal conversational search and agent-readable formats as sandboxes—places to work out the kinks in accuracy, performance, tone, and voice before anything gets released more widely.
Alessandro de Zanche, founder of media consultancy ADZ Strategies, calls agent optimization "a defensive baseline." Every quality publisher will build some version of it. The alternative is technical invisibility as search rebuilds itself around agents. But here's the harder part: discoverability is a technical problem. Retention is a business problem. Agents drive discovery, but they don't drive the trust and engagement that subscriptions and premium advertising depend on. Without that trust, the whole economics of the agent layer collapses. The publishers who survive will be the ones who figure out how to use agents to find readers, then convert that discovery into genuine subscriber loyalty.
Inside The Economist, something more fundamental is shifting. Over the past year, Muncke's team has reorganized around small, cross-functional pods—designer, engineer, product manager, editorial staff—all working together at what he calls "AI speed." The CarPlay app became a test case. It was a highly requested project that had been stuck on the roadmap for years, the kind of thing that would normally move through a long spec-and-handoff cycle. Instead, the pod got access to AI tools for writing tests, documentation, and boilerplate code. The app shipped five months earlier than planned. In certain parts of the development process, the team saw roughly 8% efficiency gains.
But efficiency is only part of the story. De Zanche points out that saved time is only valuable if it's used for something that matters—more products, faster experimentation that improves retention, or freeing creative people from repetitive work. If the capacity just gets absorbed into doing the same work faster, or if it masks headcount cuts, the return is an illusion. Abi Watson, head of publishing at Enders, frames the real opportunity differently: the medium-term play isn't about productivity. It's about what new product categories AI makes possible. When AI shortens the cycle from idea to a launched paid product—a new newsletter tier, a verticalized data product, an agentic research interface for subscribers—the upside is real because it's tied to subscription or enterprise revenue, not just internal efficiency.
The CarPlay sprint is now a template. The Economist has set up six to eight pods across its product stack. In areas that touch reader experience, editorial staff are embedded directly in those pods to make sure AI-powered features still sound like The Economist. And quietly, the publisher has been cultivating something it calls a vibe-coding culture: encouraging staff who have never touched a code editor to start building their own tools. Editors on the science desk are now writing utilities that trawl academic journals, pull relevant papers, and assess credibility for upcoming stories. Product teams are spinning up automated performance reports instead of waiting in queue with data teams. There's even a template "Chief of Staff" agent that plugs into inboxes and calendars, drafts responses, and surfaces daily priorities. Not every experiment works. The publisher paused an attempt to turn its 300-page house style guide into an automated copy checker, and it shelved AI companions for live subscriber events after testers found them more distracting than helpful. Muncke is clear about the lines The Economist won't cross. "Nobody wants to read an AI-written Economist," he said. AI's role is confined to research, workflow, and utility. The group has committed to clear labeling so readers aren't tricked about where AI is used. For all the experiments, The Economist still sees AI as infrastructure, not authorship—something to speed up research and product delivery, not a shortcut to churning out more copy. The real test over the next few months is whether Muncke can keep that balance: choosing the right use cases, spreading the skills beyond a handful of enthusiasts, and making sure the best internal hacks don't stay hidden on individual desks. If The Economist pulls it off, agents may do more of the fetching. But the thing subscribers are paying for will still be the human judgment in the middle.
Notable Quotes
Agent optimization is a defensive baseline. Every quality publisher will build some version: the alternative is technical invisibility as search rebuilds around agents.— Alessandro de Zanche, founder of ADZ Strategies
The main difference is that now everyone is a builder.— Josh Muncke, VP of generative AI at The Economist Group
The Hearth Conversation Another angle on the story
So The Economist is essentially saying that AI agents are now a distribution channel they have to optimize for, the way they once optimized for Google search?
Exactly that. Except it's more urgent because the intermediary is no longer a search algorithm—it's a conversational AI that a user is asking directly. The publisher doesn't control the interface anymore. They just have to make sure their content is legible when an agent goes looking for it.
And that's why they need two versions of everything? One for humans, one for machines?
Right. A human wants to see a beautiful layout, context, maybe some photography. An agent wants clean structure—questions, answers, plain text. You can't really do both in one design. So you build parallel versions.
That sounds expensive and complicated. Why not just let the agents figure it out?
Because agents don't figure it out well. They hallucinate, they miss nuance, they get confused by visual design. If The Economist wants its voice to come through clearly in an agent's response, it has to hand-structure the content that way. It's a control thing.
The CarPlay app shipped five months early. That's a big deal. But you mentioned that's not really the point?
It's a signal that the workflow is faster. But faster at what? If you're just shipping more of the same thing, it doesn't matter. The real value is if that speed lets you try new product ideas—new revenue streams—that you couldn't afford to experiment with before.
And the vibe-coding thing—editors writing their own tools. That seems like it could go wrong pretty easily.
It could. But it also means the people closest to the work—the ones who understand what's actually needed—can build solutions instead of waiting for a tech team. The risk is that you end up with a bunch of half-built tools scattered across desks. The opportunity is that you democratize problem-solving.