How to See What AI Actually Searches

Most AI search products show you the final answer, but hide the work that happened underneath it. That makes it hard to know whether the model searched the right thing, cited strong pages, or leaned on weak source material.

LLM Sleuth is built around a simple workflow:

1. Send the same question to multiple providers.
2. Capture the search queries each provider exposes.
3. Capture the source URLs the provider used or cited.
4. Fetch the source pages and extract headings, page text, HTML samples, and FAQ schema.
5. Compare the answer against the evidence.

Why Query Capture Matters

Two models can answer the same question while searching completely different phrases. One might search official docs. Another might search old listicles. Another might cite sources but not expose the URLs you expected.

When you can see the actual query strings, you can debug the answer instead of guessing.

What To Check First

Look at the provider search queries before reading the final answer. If the queries are too broad, old, or pointed at low-quality sources, the answer needs extra scrutiny.

Then inspect the source pages. The useful signals are usually:

• page title and snippet
• final URL and HTTP status
• headings
• extracted text
• FAQ schema
• source markdown from reader fallbacks

The Goal

The goal is not to prove one model is always better. The goal is to make AI-assisted research auditable. You should be able to answer: what did it search, what did it find, what did it cite, and what did it say?