Why traffic is the wrong number
When a model cites you, the user often never clicks. Counting visits therefore measures the opposite of what happened. The signal you actually control is whether you appear in the answer at all.
Build the question set first
Pick 30 to 60 questions a real buyer would ask. Include comparison, pricing, suitability and risk questions, not just definitional ones.
- Comparison: 'X vs Y for small teams'
- Pricing: 'how much does X cost'
- Suitability: 'is X appropriate for enterprise'
- Risk: 'what are the limitations of X'
Record four fields per run
- Mentioned: yes or no
- Position: first, middle or tail of the answer
- Cited source: which URL the model relied on
- Accuracy: whether the description matches your facts
Separate change from noise
Model answers drift. A single snapshot proves nothing. Run the same set on a fixed schedule and judge a trend, not a day. If mention rate moves and stays moved, that is signal; if it oscillates, that is the platform.
Then close the loop
Once you know which questions you win and which you lose, the next step is a retest: change the page, rescan, and confirm the mention rate moved. Without that step you have a dashboard, not a practice.