Back to the research
The Agentic Webweb agentspage simplification

How Web Agents Simplify Your Page Before Acting

Research prototypes shrink HTML before an AI model sees it, which affects what your page can communicate.

Does a web agent read my whole page or a simplified version?

In two research agents, the model does not read the raw page. Both AutoWebGLM and Agent-E first reduce page code to what they judge essential. If you build for agents, assume the text and controls that survive that reduction are the ones that count. Do not assume every product works this way.

At a glance

  • Finding: these 2024 research prototypes simplify page markup before the model reasons over it.
  • Why it matters: content or controls lost in simplification may be invisible to the agent.
  • Limit: we reviewed only the papers' abstracts, and they describe research systems, not today's commercial agents.

What the research describes

The AutoWebGLM authors say web agents struggle partly because HTML is complex, pages offer many possible actions, and the open web makes tasks hard. Their response was an HTML simplification algorithm meant to keep vital information in compact form (Lai et al., 2024).

Agent-E takes a similar route. Its abstract describes "flexible DOM distillation and denoising." The DOM is the browser's structured representation of a page. It also describes "change observation," which helps the agent notice what its action changed (Abuelsaad et al., 2024).

An illustrative example, not from the papers: a checkout page has a visible "Place order" button, but the page also holds hundreds of lines of tracking and layout code. A distillation step might keep the button and drop the rest. If your button is an unlabeled, non-native element, a distiller could drop it or misdescribe it. That is an inference, not a tested result.

What the results do and do not show

Agent-E reports beating other state-of-the-art web agents on the WebVoyager benchmark in most categories by 10–30%. The authors also list distillation and de-noising of observations as a design principle. The abstract does not say how much of the gain comes from that step alone.

WebPilot, another 2024 system, focuses on planning and search rather than page representation. It reports a 93% relative success-rate increase over a concurrent tree-search method on WebArena (Zhang et al., 2024). That shows agent design matters beyond page markup. Benchmarks are controlled tasks, so these numbers do not predict results on your site.

Seeing a page is not permission to act

An agent that can find your button can also be misled by text on a page. The WASP benchmark found that top models could be deceived by simple, human-written prompt injections in realistic scenarios. Attacks partially succeeded in up to 86% of cases, yet agents often failed to complete the attacker's full goal (Evtimov et al., 2025). The authors call this "security by incompetence," which is not a safeguard to rely on.

For site owners, this separates two questions. What can the agent observe? What is it authorized to do? A clear page helps the first. Confirmation steps and server-side permission checks govern the second.

What remains uncertain

The papers are from 2024 and 2025, and we saw only abstracts. We cannot say which commercial agents use similar simplification, how they handle your markup, or whether cleaner structure leads to more completed tasks or sales.

Next steps

  1. Test your key flows, such as signup, booking or checkout, using native links, buttons and labeled form fields. Record whether an agent you can access completes them.
  2. Compare a simplified view of your page, such as its accessibility tree, with what you intended to show. Look for missing labels or controls. Treat this as a proxy, not how every agent sees the page.
  3. Require confirmation and server-side checks for consequential actions, and never let page text alone authorize them.
← Explore more research