Define the evidence

Choose prompts that reflect real reader decisions, such as evaluating document-sharing practices or comparing relevant operational approaches. Keep the questions neutral rather than inserting the publication’s name to force a mention. Record why each prompt belongs in the panel and the audience it represents. A small stable panel can help observe changes over time, but it should not be presented as a representative sample of every question users ask.

Make it operational

Run checks under documented conditions and record all results, including no citation, an unrelated citation or an inaccurate response. Do not keep rerunning a question until the desired source appears and then report only that attempt. If prompts change, version the panel so comparisons remain understandable. Review the cited pages and the accuracy of their use. The practical outcome is a list of evidence gaps and useful content improvements, not a promise that a wording trick can control an AI system’s answers.

Imagine this

Hypothetical example

A journal tracks a fixed set of operational questions monthly. Some responses cite its guides and others do not. The team records every check and improves an unclear definition discovered during review, rather than selecting only the most flattering screenshots for its performance report.

Make it practical.

Use this checklist as you read. Checks are not stored.

References & further reading

Further reading: Google AI features and websites ↗

Published by Tap Tap Go. Product availability and eligibility vary by programme and market. About our stories.