What matters
- Keep non-brand, brand, comparison, local, and factual prompts separate.
- Record absence and incorrect answers as carefully as citation wins.
- Pair AI tests with search, bot, visit, and lead baselines.
A useful small baseline
Start with 30 to 60 real search and buyer questions. Sort them before the test. Write short score rules for mentions, citations, correct facts, and advice. Run each prompt more than once and save every answer.
The result may be zero citations to your site. That is useful. A baseline is a starting point, not a sales asset.
The public AI visibility run log includes three sample prompt rows so you can see the panel fields before editing it.
What to freeze
- Exact prompt wording and ID.
- Included products and modes.
- Target place and account state.
- Number and schedule of repeated runs.
- Brand mention and citation score rules.
- The main fact list used to check answers.
- The change window and the next measurement date.
Do not change prompts in the middle because other words work better. Put new questions in the next set.
Use the full AI visibility guide to build your test log.
Sources
- Towards a Measurement-Based Audit of Generative AI Citation Behavior, Proceedings of Machine Learning Research