Can an agent use your website? How Ora finds out with eve

Highlights

• AI agents grew by over 8,000% in the past year and have crossed human traffic on the web for the first time.

• Only 4% of the web is barely usable by AI agents, with top companies scoring an average of 41 out of 100.

• Agents reach `llms.txt` files in only 32% of cases, and 86% of those successes happen because the site explicitly links to the file.

• Agents guess specific URLs such as `/pricing`, `/docs`, and `/api` in 35% of cases before performing a web search.

• Open-source models outperform closed-source models by 2% on web tasks while operating at a significantly lower cost.

Timeline

1. 00:00 - 02:23

   Rise of Agentic Web and the Agent Front

   Asaf Elovik introduces the concept of the agent front, defining a new stack of protocols and front-end optimizations tailored for how AI agents see, think, and interact with websites differently than humans.

   • AI agents grew by over 8,000% last year and are projected to grow 1,000x over the next five years.
   • Agent traffic has crossed human traffic on the web for the first time in history.
   • Major infrastructure providers like Salesforce and Sentry now expose backend systems for agent interaction.
2. 02:23 - 06:44

   Website Usability Audit and Agent Behavior Research

   Research reveals that agents frequently fail on standard websites, often resorting to guessing URLs or falling into competitor sites when documentation or pricing pages are missing or obscured.

   • Aura analyzed thousands of intents across top agents on 120,000 sites and found only 4% of the web is barely usable.
   • The agentic.com tool built with Vercel allows users to scan any URL for agent usability metrics.
   • Top-tier companies score an average of 41 in agent usability.
3. 06:44 - 14:10

   Empirical Findings on How Agents Navigate Websites

   Detailed truths about agent behavior show that modern models rely less on static memory and more on real-time web access, while open-source models achieve slightly higher success rates than closed models on web tasks at a fraction of the cost.

   • Agents reach `llms.txt` files in 32% of cases, primarily when sites explicitly link to them.
   • Agents process HTML rather than visual components, and web fetch tools often trim content.
   • Spinning up a browser requires 45 times more tokens and 23 times more time than processing HTML.
   • Agents guess URLs like `/pricing`, `/docs`, and `/api` in 35% of cases before executing web searches.
4. 14:10 - 19:13

   Optimizing Infrastructure with EVE and Actionable Fixes

   To prepare for the agentic web, site owners must ensure accessibility, provide explicit routing for guessed URLs, and integrate tool-use protocols so agents can interact with services directly.

   • Aura utilizes Vercel infrastructure and the EVE harness to orchestrate agent tasks efficiently.
   • Developers must link `llms.txt` from home and docs pages, placing critical content at the top of HTML.
   • Creating markdown twins and adopting protocols like WebMCP enables agents to use sites effectively.