Zapier gutted one of its own web pages on purpose. Deleted every useful word, kept the URL identical, then asked Claude to read it. The agent still described the page perfectly. If you sell anything online, how AI agents read websites just became your problem, because the way they browse is nothing like the way you assume.
The experiment
Zapier’s engineering team built a little contraption they call Agent Arm Runner. It runs Claude Code inside a Docker container with a proxy that can swap what’s behind a URL. Same prompt, same address, but the production page goes to the control group and an empty variant goes to the test group.
Their target was the Slack integration page, the one listing everything Zapier can do with Slack. For the variant, they stripped the body down to almost nothing. No triggers. No actions. No integration content.
The expectation was obvious. Empty page in, empty answer out.
Instead, Claude came back rattling off specifics: 25 triggers including new mentions and reactions, 27 actions like sending direct messages, 11 searches including find user by email. All of it, from a page that no longer contained any of it. So how AI agents read websites apparently has a workaround for empty pages. The logs explained everything.
Surprise 1: markdown, not your page
The proxy logs caught the culprit. The agent’s fetch tool requested the page with a header asking for text/markdown first, HTML second.
Turns out Zapier had been experimenting with markdown versions of its pages, cleaner copies that agents can digest without all the surrounding markup. The HTML variant was gutted. Its markdown twin was not. So the agent read the intact markdown and answered from that.
They even planted a canary string in the markdown only, a fake verification code with instructions to include it in any summary. The canary came back in the agent’s response. Case closed: the agent was never reading the page you’d recognize as “the page.”
Here’s the shift. For years, SEO was about one artifact: the HTML page. Now a site can have parallel versions, and understanding how AI agents read websites means auditing every version you publish, not just the pretty one. Whatever a fetcher can request, an agent can consume.
Surprise 2: a second model reads first
There was a second plot twist hiding in the logs. When the agent fetched a page, the content didn’t go straight to the main model. It went to a cheaper, separate model that summarized the page first, and only the summary reached Claude.
That middle layer has its own instructions. Quote-length limits. Paraphrasing rules. So the version of your page that the “reader” sees is already filtered, condensed, and reshaped before the conversation even starts.
Think about what that means. Your carefully worded landing page gets pre-chewed by a model you didn’t choose, under rules you can’t see, and the main agent reasons about the leftovers. Cloudflare is now offering whole browser infrastructure for agents, and the request path keeps growing more layers, not fewer.
What it means if you sell online
Agents already shop for people. ChatGPT recommends products, Perplexity summarizes services, and browsing agents fill forms and compare prices. If an agent can’t read your site, you’re invisible at the exact moment someone asks “which option should I buy?” The mechanics of how AI agents read websites decide whether your product ever enters that conversation.
Three consequences worth internalizing:
- Your content must exist in plain, server-rendered HTML. Most agents don’t run JavaScript. Pages that build themselves in the browser look empty to them.
- Structured data belongs in the initial HTML too. If prices and specs inject client-side, agents never see them.
- You have less control than you think. A summarizer model sits between your words and the answer, and it compresses ruthlessly.
How agents even find your site
Before any fetching happens, a crawler has to discover you. ClaudeBot, PerplexityBot, GPTBot and friends wander the web like Google’s spider always did, and site owners can block or allow each one in robots.txt. Plenty of sites blocked them in 2024 out of AI-training fears. That choice now has a side effect: block the crawler, and the shopping agent may never learn your store exists.
Worth a quick check. If your robots.txt bans every AI bot, you’ve opted out of agent commerce entirely, which is a strategy, but should be a deliberate one. The traffic shift is already visible elsewhere: Google itself admitted AI Mode traffic patterns aren’t bouncing back, and we’ve covered when AI search beats Google for different kinds of questions. Agents are the next layer of that same shift.
The practical stance for a small site: allow the crawlers, keep the content in server-rendered HTML, and let robots.txt do its thing deliberately instead of by accident.
The 10-minute self-audit
You don’t need Docker containers or proxies to check your own exposure. Zapier’s experiment sounds exotic, but the beginner version fits in a coffee break:
| Check | How | What good looks like |
|---|---|---|
| Raw HTML test | View-source your key page, search for your main offer | The offer text exists in the HTML itself |
| Agent Q&A test | Ask Claude or ChatGPT “what does [yoursite.com] sell?” | Answer matches reality, not a hallucination |
| JS blindness test | Disable JavaScript in dev tools, reload | Core content still visible |
| Title and headings | Scan the raw HTML heading order | One clear h1, logical h2s |
| Markdown mirror | Search your domain for /llms.txt or .md copies | You know what versions exist |
That last row matters more after Zapier’s canary test. If your platform auto-generates markdown or simplified copies of pages, they’re all in play. Audit the versions, not just the homepage. Ten minutes of checking beats months of guessing how AI agents read websites behind your back. And if your traffic graphs already look wrong, our piece on Google’s AI Mode traffic admission explains part of why.
And do the agent Q&A test on your competitors too. It’s competitive intelligence that costs nothing.
Don’t over-engineer it
Tempting response: rebuild everything, add markdown mirrors, chase every agent crawler name in your logs. Resist that for now.
The durable moves are boring ones. Server-render your important pages. Keep semantic HTML and clear headings, because agents navigate pages closer to how keyboard-only users do, stepping through elements in order. Put the facts (prices, specs, availability) directly in the markup. Answer real questions in plain language.
Beyond that, assume churn. Zapier itself closed by warning that today’s behavior won’t match tomorrow’s, even for the same tools. The teams winning right now aren’t optimizing for one agent’s quirks. They’re making their content legible to anything that fetches, then verifying with the Q&A test every few weeks.
Run that audit on your site today. Worst case, you learn your pages already read fine. Best case, you find the invisible problem before your next customer’s agent does.
Sources: What we learned by watching Claude read Zapier.com (Zapier Blog), Cloudflare Browser Run, How AI agents browse websites (Crazy Egg)