Agent Simulation Testing — Test Your Site With a Real LLM
Static checks tell you if your site can be read by agents. Simulation testing tells you if agents can actually do things on your site. This is the difference between passing a compliance checklist and passing a real-world test — and it's a test you have to run yourself, today, with a real LLM.
What is Agent Simulation Testing?
Agent simulation testing is the practice of sending a real LLM (like Claude or GPT-4o) to interact with your website and measuring whether it can complete specific tasks. Instead of checking for the presence of files or the validity of schemas, it answers a more fundamental question: can an AI agent actually use your site?
The process works like a human QA tester, but the tester is an AI. You define tasks — "Find the pricing for the Pro plan," "Start a trial," "Locate the API documentation" — and the agent attempts each one. The output is a pass/fail result with a detailed trace of what the agent tried, where it got stuck, and why it succeeded or failed.
This approach catches problems that static analysis misses. Your llms.txt might be perfect, your schemas valid, your robots.txt open — but if your pricing page is a client-rendered React component with no server-side content, the agent sees an empty page and cannot answer the most basic question about your product.
How to Run This Yourself, Step by Step
You do not need any special tooling for this — just an LLM with web access. Here is the structured version of the process:
Context Gathering
Have the agent read your llms.txt, homepage, and sitemap first to build an understanding of your site structure — just like a real agent would in the wild.
Task Assignment
Give it tasks representing common agent use cases: finding pricing, locating documentation, identifying contact methods, starting a trial, and comparing features.
Autonomous Navigation
Let it navigate your site following links and reading content to extract the information needed for each task, the same way a real browsing agent would.
Completion Assessment
Judge each task as completed, partially completed, or failed, based on whether the agent actually produced the right answer or got stuck.
Trace Review
Ask the agent to show its reasoning — which pages it visited, what it extracted, and where it got stuck — so you know exactly what to fix.
What Simulation Testing Catches
Simulation testing reveals problems that are invisible to static analysis:
Client-side rendering gaps
Pages that look fine in a browser but return empty HTML to agents. Common with SPAs that defer all rendering to JavaScript.
Broken navigation paths
Links that work for humans (via client-side routing) but do not resolve when an agent follows the raw href attribute.
Missing or ambiguous pricing
Pricing pages that use interactive sliders, toggles, or custom calculators that agents cannot operate. The agent sees the page but cannot extract a price.
Gated content without signals
Content behind login walls without any indication in the HTML that authentication is required. Agents see a blank page or redirect loop.
Conflicting information
Pages that say one thing in visible text and another in JSON-LD. Agents that find contradictions lose confidence in your site.
Dead-end pages
Pages with no internal links, no schema, and no clear next step. Agents arrive and cannot figure out what to do next.
How to Test Your Site Manually
You can run a basic agent simulation yourself using any LLM with web browsing capabilities. Open ChatGPT, Claude, or Perplexity and ask it to complete tasks on your site:
- "Go to [your-site.com] and tell me how much the Pro plan costs."
- "Find the API documentation for [your-site.com] and show me how to authenticate."
- "What does [your-company] do? Find the answer from their website, not your training data."
- "Walk me through starting a trial on [your-site.com]."
If the agent struggles, hallucinates, or gives wrong answers — those are the same failures real users will experience when they ask AI assistants about your product. The difference: you will never see those failures in your analytics because the agent never clicks through to your site. The user just gets a wrong answer and moves on.
What the AX Audit Checks Instead — and Why You Still Need to Test Manually
To be direct about what the free AX Audit does and does not do: it runs 17 static, structural checks — it does not currently send a live LLM to attempt tasks on your site and grade the outcome. There is no "Task Completion" line item in the score.
What the audit does check is the scaffolding that task completion depends on: AR-PROV (can an agent open an account without a human), Authentication Standards (OAuth/OIDC, Web Bot Auth), SSR Detection (is content present in the initial HTML, or only after client-side JavaScript runs), and API Documentation. A site that fails these checks is very likely to also fail the kind of manual simulation described above — a client-rendered pricing page with no SSR is exactly the failure mode that both a low SSR Detection score and a failed "find the price" test point to.
But passing every static check is necessary, not sufficient. The only way to know for certain whether an agent can actually complete a task end-to-end on your site is to run the manual simulation above yourself, today, with a real LLM.
Check the Scaffolding Task Completion Depends On
The AX Audit checks whether the technical prerequisites for agent task completion are in place — SSR, authentication standards, provisioning, API docs. Then use the manual simulation above to verify real task completion yourself.
Run a free AX audit→Free audit. No signup required.