The Economist Is Pre-Building a Second Web for AI Agents

The Economist is building two parallel versions of its web: one for human readers, one structured for AI agents, with Q&A-formatted text replacing carousels and feature art. VP of generative AI Josh Muncke told Digiday the publisher is preparing for "a world with two versions of the web," starting with marketing and B2B sales pages outside the paywall. If your brand isn't licensable as a structured feed, the agent track routes around it.

The mechanics, not the manifesto

The agent track is not a slimmed-down version of the same page. It's a different artifact with different formatting conventions, sitting at a different URL pattern, optimized for ingestion rather than reading. The Economist isn't dumbing things down for agents. It's recognizing the agent is a different reader with different needs, and building accordingly. Muncke's line that "agents want clear structure, questions and answers, ideally text" is the easy quote. The harder half is "ideally text," because that means letting go of a lot of design and editorial craft on the agent surface.

Underneath that content shift is an org shift. Muncke's team folded generative AI into product development cycles, restructured around 6 to 8 cross-functional pods, and shipped a CarPlay app five months ahead of schedule. "The main difference is that now everyone is a builder," he said. That part isn't marketing fluff. It's the staffing change required to actually run parallel content tracks at the velocity the agent market demands.

Why AEO 1.0 doesn't get you on the agent track

Most AEO work over the past year has been about appearing in human-facing AI surfaces: ChatGPT citations, Perplexity sources, AI Overviews snippets. The optimization layer has been schema, llms.txt files, chunking strategy, structured FAQ blocks.

Google's own first AEO guide, published a couple of weeks ago, more or less called most of that decorative. (We recapped it here.) Their version is: do good SEO, write for entities, ship clean schema. Which is fine, as far as it goes.

What The Economist is doing skips past that fight entirely. They're not optimizing the human-facing site for agents. They're building a separate surface, a separate URL pattern, a separate format, specifically for agent ingestion. That is a meaningfully different bet. It says the agent web isn't a layer on top of the human web. It's parallel infrastructure, and you have to ship into it directly.

If that bet is right, optimizing your existing site for agent visibility is the temporary phase. The end state is licensed feeds (structured, paid, distinct) feeding agent answer engines. Brands that don't ship those feeds aren't competing with worse versions of themselves. They just aren't on the list.

The Cloudflare layer most brand decks are missing

The plumbing is further along than most marketers know. Cloudflare introduced Markdown for Agents in March: crawlers send Accept: text/markdown, the edge converts the HTML, returns Markdown plus an x-markdown-tokens header showing estimated token cost. Same source URL, different artifact, served at the network edge. Paired with Content Signals declared in robots.txt comments (search, ai-input, ai-train), publishers can express which agent uses are permitted, even if those signals are preferences rather than hard enforcement.

Cloudflare also acquired Human Native to handle the licensing piece, turning publisher content into AI-ready data and ensuring creators get paid. AdMonsters called this the start of an AI licensing stack for publisher clients. None of it is hypothetical. Pay Per Crawl is already shipping, letting publishers choose allow, charge, or block on a per-crawler basis.

Where The Economist sits in that picture: they're building the agent-facing content artifact. Cloudflare is building the access, signaling, and payment layer underneath. Combined, you get the basic shape of a licensed agent web. The only piece missing in most brand stacks is the part where the brand actually has an agent-formatted feed to license.

Where this stops being a publisher problem

The Economist is a media company, so the agent track is content. For a brand, the agent track is product data, pricing, availability, support content, return policy, comparison material, and the specific Q&A surface an agent needs to answer "should I buy this." Most brand sites today have none of that as a structured artifact. They have PDPs designed for humans, with rich media, conditional copy, and JS-rendered specs that agents either parse poorly or pay tokens to digest.

The closest parallel I've seen is how Netflix ships separate metadata feeds to Roku, Apple TV, and connected TV platforms: same content, different ingestion surface, different format. The "web" Cloudflare is helping publishers build for agents is roughly that, except the agent is the device and the format is licensed text. Brands that spent a decade getting product feeds right for Google Shopping and Amazon are well-positioned to do this. Brands still treating their site as the canonical surface are not.

One more piece worth naming: the Model Context Protocol ecosystem crossed 9,400 public servers in 2026, and Anthropic donated MCP to the Linux Foundation's Agentic AI Foundation late last year. The agent stack already has vendor-neutral plumbing. The format is settling. The licensing rail is being laid. From what I've seen, most brand marketing teams treat this as an engineering problem and quietly punt it to a future quarter. That's probably the mistake.

The agent isn't reading your homepage. It's reading whatever artifact you decided to license.

Your minimum viable agent track (build it this week)

Stop trying to "win AEO" on your current site as a single layer. Start treating the agent track as a separate product surface. Three concrete moves this week:

1. Inventory your top 20 pages by either organic traffic or revenue. For each one, write the three to five questions an agent would actually ask to answer that page's job-to-be-done. That's your agent track scope.
2. Build one (one) Q&A-formatted agent-readable version of a single high-value page. Plain text, headings, no JS. Park it at a distinct URL pattern such as /agents/{slug}. Don't link it in human nav. The benchmark: a working, plain-text Q&A page parseable by an LLM without rendering, live by end of week.
3. Add Cloudflare Content Signals to your robots.txt with explicit ai-input and ai-train values that match your licensing posture, not whatever default came with your CDN setup.

Three small actions. Real artifact at the end. Roughly four working days if your team isn't fighting other fires.

One prediction with stakes

By Q3 2026, the publishers and brands with shipping agent tracks will have set the format conventions agents reward citation-wise. Everyone else gets ingested through whichever comparison aggregator, review site, or competitor cited them. The Economist is moving now because Muncke and team can see the citation graph forming, and they don't want to be a footnote on someone else's licensed feed.

I don't think there's a clean way to skip this phase. Either you have an agent-readable artifact on the open web by Q3, or your customers' next purchase-research conversation happens between an agent and a feed someone else owns.

Notice Me Senpai Editorial