# Check a web extract before an agent uses it

1. Choose three public pages you can independently verify.
2. Before running, write down the exact facts and usable destination links the task requires.
3. Record the URL, run date, connector name, parameters and the complete returned output.
4. Open each original page independently. Compare price AND scope, dates AND time zones, and link destinations.
5. Mark facts as present, absent, contradicted or unclear. An omitted heading can make a correct number ambiguous.
6. Keep claims about observed output separate from theories about caches, models or retrieval internals.
7. Resolve important discrepancies at the primary source before publishing a recommendation or inviting someone to pay.
8. Repeat on another day and on pages outside your own site before selecting a tool for a continuing workflow.

Record for each case: source URL / source check date / required fact / exact supporting excerpt / destination URL / result / correction needed.

Worked test: https://aisuperpower.cc/guides/exa-vs-tinyfish-web-extraction-test
Evidence: https://aisuperpower.cc/downloads/web-extraction-trial-2026-11-14.json
