 (1).png)
The web is splitting into two layers. One is built for humans: rich design, branding, narrative, conversion. The other is built for AI agents: plain text, clear structure, explicit questions and answers, and facts that can be extracted cleanly. As AI assistants become a starting point for discovery, publishers increasingly need both surfaces. And the moment you build an agent-readable layer, you need a way to know whether agents are actually reading it.
For most of its history, the web had one audience. Everything we learned about building websites (layout, hierarchy, hero images, calls to action, the craft of holding attention) was learned in service of a human looking at a screen. That assumption held for thirty years.
It's quietly ending. A second reader has arrived: the AI agent that fetches a page not to look at it, but to parse it, summarize it, and answer a question on someone else's behalf. It doesn't care about your hero image. It wants the answer, structured so it can be lifted with confidence. That's a different job, and it's starting to require a different kind of page.
The web is no longer read only by humans
The shift is concrete enough that serious publishers are already acting on it. According to Digiday's reporting, The Economist is experimenting with agent-readable versions of content, beginning with marketing copy and B2B sales material that already sits outside its paywall, explicitly preparing for what Josh Muncke, who leads generative AI at The Economist Group, described as "a world with two versions of the web."
The reasoning is straightforward. If discovery increasingly begins inside an AI assistant rather than on a homepage or in a search box, then the version of your content that the assistant can read becomes the version that determines whether you show up at all. The agent is the new front door, and it reads differently than a person does.
This is no longer a thought experiment. By some industry estimates, more than half of B2B buyers now begin product research with an AI assistant (ChatGPT, Gemini, Claude, or Perplexity) rather than a traditional search engine. Marketing and comparison pages that don't render cleanly to those systems risk being absent from the answer entirely. For publishers, editorial content raises a harder set of questions around paywalls, licensing, and value protection. But the underlying direction is the same: agent-readable content is moving from SEO experiment to part of the go-to-market plan.
Human pages and agent pages have different needs
It helps to be precise about why one page can't always serve both readers well.
A human-facing page is optimized for attention. It uses design, brand, imagery, pacing, and persuasion to hold someone's interest and move them toward an action. Carousels, feature art, layered layouts, and a strong narrative arc all earn their place because they work on people.
An agent-facing page is optimized for retrieval. It rewards clear structure, plain language, explicit entities and facts, defined terms, comparison points, and question-and-answer formats that a machine can extract and reuse with minimal friction. The things that make a page persuasive to a person (visual richness, implication, atmosphere) are often noise to a system trying to pull a reliable answer.
The two overlap, but not completely. A beautiful landing page can be genuinely hard for an agent to parse, with the key facts buried in imagery or scattered across interactive elements. A clean, Q&A-style machine-readable page can be flat and unengaging for a human while being exactly what an AI system needs. Recognizing that these are different jobs, rather than assuming the ideal human page is automatically the ideal source for a machine, is the whole insight behind the two-track approach.
Why publishers are experimenting with agent-readable surfaces
The practical unit of competition is shifting. For years it was the article, the ranking page, or the paywall. Increasingly it's the structured information layer that AI systems can parse and cite. (Honeylog's media and publishers use case is built around exactly this change.)
That's why agent-readable surfaces are showing up first where the stakes are clearest and the risk is lowest: marketing copy, product and pricing explainers, comparison pages, and B2B sales material that already lives outside the paywall. These are pages a publisher wants discovered. Making them legible to agents is close to pure upside. The same logic applies to any content-driven business. B2B and corporate sites face the identical question about whether their sales and product pages show up cleanly in AI answers.
Editorial content is the harder frontier, because every decision to expose it cleanly to machines is also a decision about how much of your value you're giving away. Which brings us to the part of this conversation that gets skipped.
The missing analytics layer
Here's the gap. Publishers are being told, correctly, to build agent-readable content. Almost no one is talking about how to measure whether it works.
This matters because the standard analytics stack can't see the audience in question. Tools like GA4 are built around a JavaScript tag that fires in a human's browser. AI agents and crawlers typically fetch your HTML directly, without running that script, so their visits don't register as visits at all. You can build a pristine agent-readable page, publish it, and watch your dashboard report essentially nothing. Not because nothing happened, but because the dashboard was never designed to detect this reader. (This is the core difference behind any honest Honeylog vs GA4 comparison: one waits for a browser, the other reads the server.)
So a publisher investing in a two-track strategy is, by default, optimizing one of the two tracks completely blind. They can A/B test a human page all day. They have no native way to answer the questions that actually determine whether the agent-readable effort paid off:
- Are AI agents actually visiting the machine-facing pages we built for them?
- Which bots are accessing which surfaces, and how often?
- How does crawl activity on agent-facing pages compare to our human-facing equivalents?
- Are our B2B pages, explainers, teasers, documentation, and open content being read by the systems we're trying to reach?
Without that evidence, "optimizing for agents" is an act of faith.
The paywall and discoverability tradeoff
For subscription publishers, the two-track question is inseparable from the paywall. They want discovery (invisibility in AI answers is its own kind of death) but they can't give everything away without eroding the reason anyone pays. The Economist's own experiments are deliberately careful for this reason: the opportunity and the risk sit in the same place.
The usual framing of this debate is "what should we put outside the paywall?" That's the wrong first question, or at least an incomplete one. The more useful question is "who is actually accessing what we've already made open, and what are they doing with it?" You can't make a sound decision about exposure without knowing the current behavior. Are agents crawling your open teasers heavily and ignoring the rest? Are particular bots hammering specific high-value explainers? That's the kind of evidence that turns a philosophical argument in a strategy meeting into a data-driven decision.
How Honeylog helps measure agent access
This is the layer Honeylog is built for. It works at the server level, analyzing traffic and logs rather than waiting for a browser to run a tag. That lets it identify AI bots, crawlers, and LLM-related activity, and show what content they accessed, on which paths, and how frequently. The detection approach is on the features page.
For a two-track strategy, that turns the unanswerable questions above into reportable ones. You can see whether agents are reaching your agent-readable pages. You can compare crawl activity across human-facing and machine-facing surfaces. You can watch which bots concentrate on which content, and whether your open marketing and B2B pages are being read by the systems you built them for.
It's worth being precise about scope, because this space is full of overclaiming. Honeylog does not prove what a model trained on, and it doesn't reveal what an AI assistant "knows" about you. Anyone promising that is selling a guess. What it shows is concrete and server-verified: access, crawling, and traffic patterns from AI agents and bots. That's a narrower claim than the marketing around this category usually makes, and a far more defensible one, because it's grounded in evidence you can point to. It complements the AI visibility tools that track external citations; it doesn't replace them.
Why this matters before publishers scale AI-agent strategies
Agent optimization is on its way to becoming a baseline rather than a differentiator (which is something every serious publisher does) the way responsive design or basic SEO became table stakes. The danger is treating the building as the whole job.
If agent-readable content is going to be a standing investment, then visibility into agent behavior has to be part of that baseline too. Otherwise publishers scale a strategy they can't observe: producing parallel versions of pages, exposing content past the paywall, restructuring marketing material, all on the assumption that agents are reading it, with no instrument confirming they are. Build the second track, by all means. But build the gauge that tells you it's carrying traffic before you pour more into it.
A two-track web needs two-track analytics: one layer for human behavior, and one layer for the machines now reading the web on our behalf.