An AI agent fills out your form instead of the visitor: what that means for your landing page

Until recently, a landing page had only one kind of visitor to convince: a person who reads, hesitates, and clicks. Since the rise of so-called "agentic" browsers — Perplexity Comet, or the agent mode built into ChatGPT after its standalone Atlas browser was shut down on August 9, 2026 — part of that traffic no longer reads your page at all: a software agent scans it, locates the form, and fills it out on behalf of a human who simply gave it an instruction in plain language. This is far from anecdotal: according to the security firm HUMAN, agent traffic directed at e-commerce sites jumped 144.7% during the Black Friday–Cyber Monday 2025 weekend compared to the days before. A landing page built only for the human eye is already leaving conversions on the table.

What the research says about how reliable these agents are

Before changing anything, it's worth having a realistic picture of what these agents can actually do — and their limits, well documented by academic research even before the consumer products arrived. WebShop, a benchmark published in 2022 by Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan (NeurIPS), tasked agents with buying a specific product on a simulated e-commerce site of 1.18 million listings: the best trained agents reached a 28.7% success rate, against 59.6% for humans on the same tasks. A year later, Mind2Web (Xiang Deng et al., NeurIPS 2023), tested on over 2,000 real tasks across 137 websites, showed that GPT-4 picks the correct page element only 35.8% of the time on a site it has never seen before. The common thread in both studies: an agent does markedly better against a standard, predictable interface it recognizes, and fails more often the moment a page's structure departs from common web conventions.

In other words, an AI agent isn't an infallible visitor that "reads" your page the way an experienced human would — it's a visitor that succeeds at its task provided your form looks like what it has learned to recognize elsewhere on the web. The same flaws that lose a screen-reader user or a visitor navigating by keyboard — covered in our guide to RGAA/WCAG accessibility and our article on keyboard navigation — are exactly the ones that trip up an agent: non-standard components, unpredictable structure, missing semantic cues.

What actually helps an agent fill out your form

• Native elements, not visual reconstructions — a real <select>, a real <input type="email"> or type="tel" rather than a styled div pretending to be a field: an agent relies on DOM structure and declared type, not on appearance.
• A <label> tied to each field (via for/id), not a placeholder alone that vanishes the moment typing starts — a placeholder used as the only cue trips up an agent just as much as a human, as our article on form autocomplete details.
• Standard autocomplete attributes (name, email, tel, address-line1, cc-number…) — the same signal that lets a browser offer reliable autofill also gives an agent an explicit hint about the expected data type, far more reliable than a field label it has to interpret.
• Error messages tied to the field they belong to (via aria-describedby), not a generic banner at the top of the page — an agent, like a screen reader, needs to connect the error to the offending field without guessing; our article on form microcopy covers how to phrase these messages.
• A linear flow rather than a form scattered across a carousel, drag-and-drop, or steps revealed on mouse hover — a multi-step form still works fine if it uses conventional navigation (see our guide to conditional forms); an embedded widget like Typeform or Tally, generally well-structured in native HTML, actually fares better than a poorly marked-up custom component (comparison in our article on Typeform vs. native forms).

Should you block AI agents like spam?

The temptation exists: non-human traffic filling out forms looks, on paper, like exactly what you'd normally want to filter out. That's a confusion worth avoiding. An agent acting on the explicit instruction of a consenting visitor — "sign me up," "book an appointment," "add this template to the cart" — isn't a malicious bot; it's an intermediary for a genuine purchase intent, much like a human assistant filling out a form on their boss's behalf. Our article on CAPTCHAs already recommended reserving visible defenses for cases where they're truly necessary rather than as a default reflex; the rise of agents is one more reason to favor invisible defenses (honeypots, rate limiting, server-side validation) that filter out spam without arbitrarily telling a human apart from a legitimate agent acting on their behalf.

What doesn't change

None of this requires building a version of your page "for AI agents," and that wouldn't be advisable anyway: special markup or content artificially chunked for bots does nothing for a human visitor and has no guarantee of working better for an agent, whose exact behavior remains proprietary and shifts with every update. The good news is that the useful work is the same work that has always served accessibility and classic SEO: semantic HTML, native form elements, explicit labels, errors clearly tied to their field. A form built to work for a screen-reader user or a visitor who never touches a mouse is, as a side effect, already ready for an AI agent — without a single line of code written specifically for it.

The 10 LanderKit templates (€89 per template, €229 for the full pack) ship their forms in native semantic HTML — labels correctly tied to fields, correct input types, autocomplete attributes filled in — source code included, verifiable directly on the single-product e-commerce template demo or the SaaS waitlist template demo.