Same ecommerce site. Same product. Same price. One version scored 49.3% task success with AI browser agents. The other scored 89.3%. The only difference: how readable the page was to an agent, not to a human. Here's what actually changed. 🧵

The 20-minute agent-view test: → View Source: are price, billing, refund and requirements in the HTML? → Tab through: does every button get focus and a clear name? → Ask a browsing AI 6 buying questions and have it label each answer EXPLICIT / INFERRED / UNKNOWN

Countdown timers and "only 3 left" pop-ups Agents notice them and get steered. Bigger models that reason longer were MORE susceptible, not less. Tempting? The researchers call it a safety flaw, and flaws get patched. Real deadlines in plain text survive the patch.

A cookie banner on top of your main button In the DECEPTICON benchmark, "obstruction" (extra steps in the path) was the most effective dark pattern against agents: 97% in one setup, 89% in another. Keep banners small, off the CTA, with a clear Accept and Reject.

A schema price that doesn't match the page Your JSON-LD says $17. The page says $27 since last month's price change. Humans never see the schema. Machines see both. Google already requires markup to truly represent the visible page. Update both in the same edit.

"Thousands of happy customers" An agent asked to compare carefully can't verify that, so it reads as noise. Swap one vague claim for one checkable one: a dated number, a named case study, reviews on a third-party site.

A CTA that just says "Continue" Continue to what? A trial? A checkout? A charge? Agents are built to pause before consequential actions, and a vague button is a reason to pause. Say what happens: "Buy now for $27 (one-time)."

Form fields labeled only by placeholder text "Enter your email" in grey, gone once you type. WebAIM: 33.1% of form inputs on the top 1M home pages weren't properly labeled. For an agent filling in your opt-in, that's a guessing game. Add real labels.

A styled div that looks like a button To you, it's a button. To the accessibility tree, it's a box of text. WebAIM found empty buttons on 30.6% of the top 1M home pages. Use real <button> and <a> elements with descriptive text.

Content your JavaScript builds after load Vercel/MERJ measured AI crawler traffic: none of the major AI crawlers rendered JavaScript. GPTBot alone made 569M requests in one month. Check your page with View Source, not DevTools.

Terms hidden in a modal that loads on click Refund terms in a pop-up. Pricing inside an accordion that fetches its content when opened. If it's not in the initial HTML, crawlers never see it. Collapsed is fine. Loaded-on-click is not.

Requirements that only live in your video "Works with WordPress." "Windows only." "Needs a paid OpenAI account." That's exactly what a comparing agent needs, and it's usually said once, at minute 7. Agents won't watch it. Put every requirement in visible text.

The price that only exists in an image Crawlers reading raw HTML see nothing. In one study, turning button text into images of text made it harder for agents to understand what buttons do. Fix: price, currency and billing as real text.

Quick test you can run right now: open your best-converting page and try to answer, using ONLY what's on the page - no guessing: "What am I getting, what does it cost, and what happens when I click?" If you hesitate, an agent hesitates too. It just doesn't ask.