A procurement manager at a series B company logged into a tab on a Tuesday morning and ran a prompt. By Wednesday she had a shortlist of four vendors, a comparison table built from public pricing pages, and a draft of the email she would send to the winner. She had not visited any of those vendors' websites herself. An agent had.
This is not a 2027 story. It is happening now, in pockets, on every CRM, payments, observability and analytics category we audit. The agents are clumsy. The agents are getting better every quarter. And the agents are reading our clients' websites very differently to the buyers we designed those websites for.
What an agent actually sees
An agent does not see your hero video. It does not see your founder quote, your aspirational tagline, or the gradient on your CTA button. It sees a flattened text representation of your DOM, then it parses that text for the answers to the questions its principal asked. "Find me a competitor to Vendor X with similar pricing and self-serve onboarding."
The pages that perform well for agents have three things in common. They expose structured data that a parser can lift cleanly. They publish pricing tables that survive being copied into a spreadsheet. They keep the substantive comparison information in the same paragraph, rather than spreading it across a tooltip, a modal, and a video the agent will not watch.
The pages that perform poorly are the ones we usually consider beautifully designed. Heavy hero animations. Pricing locked behind "talk to sales". Comparison information that lives only in an analyst report you have to fill in a form to download. To a human, this is polish. To an agent, this is a vendor that does not want to be evaluated.
Three shifts already visible in the data
We audited eight client sites in October for agent-readability. The findings were consistent enough to call them a pattern.
- Schema.org markup is becoming product copy. Product, Offer and FAQPage schema were previously a Google SEO concern. They are now the first thing several agent frameworks parse before reading any visible HTML. The companies that win in the next two quarters will treat their JSON-LD blocks like a second homepage, written by a copywriter, not a developer.
- Pricing pages are getting flatter. Plaid, Stripe, Linear and Vercel all moved towards published, copy-pasteable pricing in the last 18 months. The "contact sales" tier still exists, but it sits next to a transparent table instead of replacing one. Agents cannot evaluate a vendor whose price is a phone call. The vendors who hide their price are quietly disappearing from agent-shortlists.
- Documentation is now sales surface. When an agent is asked "does Vendor X support webhook retry with exponential backoff", it reads the docs. If the answer is buried in a forum post, the vendor is excluded. The companies treating their docs site as a marketing surface, with consistent voice, examples and search, are converting agents at a meaningfully higher rate than companies who treat docs as an afterthought.
The new buyer cannot be charmed. It cannot be retargeted. It does not respond to FOMO. It reads the page and decides.
What dies, what rises
Some things lose ground quickly. The personality-led founder homepage hero. The marketing persona built on age and job title. The lead-gen form that gates the white paper. None of these are bad ideas. They are simply illegible to the new third reader sitting on top of every browser session.
Some things rise. Open APIs for product data, so a competing platform's agent can integrate against you without permission. Trust signals that are machine-verifiable rather than emotional, like SOC2 attestation linked directly from your footer. Customer logos with hover-revealed case studies, where the case study is also a structured data document.
The internal version of this change is more disorienting. Marketing teams have spent twenty years optimizing for click-through rate and emotional resonance. They are now being asked to optimize for parsing accuracy and semantic coherence. The roles do not exist yet. The people who will hold them in 2027 are the ones learning, this quarter, how to write copy that performs well in both registers at once.
What we tell clients
Two practical moves to start with. Run your own homepage through an agent-readable test: open ChatGPT, paste your URL, and ask it to summarize what you offer, who it is for, and what the pricing model is. If the answer is wrong, your website is invisible to the new buyer. Then audit your top five competitor sites the same way. The gap between best and worst is currently very large, which means there is still time to be on the right side of it.
The funnel did not break. A new lobby was added in front of it, and only the websites legible to that lobby are getting through. The work to make a brand readable to humans was always worth doing. The work to make a brand readable to agents is the same work, finished properly, for the first time.