A competitor appears in AI. What should you compare?
AI Recommendation Comparison Worksheet
A higher Google position or more reviews does not establish how an AI answer will recommend companies for a particular question. Record repeated answers and their sources before drawing conclusions. Differences can generate hypotheses, but an AI explanation is not access to its ranking logic.
Create one row per prompt, platform and independent run. Keep repeat appearances across different prompts. Use a separate fresh chat for each run and keep model settings, location wording and prompt wording consistent. Do not include private customer information.
1. Run record
Date, time zone, platform, displayed model, chat mode, exact prompt and intended market. Record unavailable settings as unknown.
Your evidence / decision / next action:
2. Recommendation evidence
Save the response. List each recommended company in appearance order, distinguishing grouped categories from an explicit ranking. Record a missing company as absent only for that response.
Your evidence / decision / next action:
3. Citation evidence
Record linked pages separately from recommended businesses. A source citation does not mean the business was recommended. Deduplicate repeated links within the run.
Your evidence / decision / next action:
4. Matched Google observation
Record query, date, location, device and observed organic positions separately from map results and ads. Not found in inspected results is not proof of no ranking.
Your evidence / decision / next action:
5. Comparable page evidence
Note the service, location, credentials, original proof and third-party references actually visible on the recommended sources. Do not turn correlation into a cause.
Your evidence / decision / next action:
6. Change and retest
Document one proposed change and the reason. Repeat the same prompt panel later, preserving misses and unchanged results. Avoid claiming causation from a before-and-after screenshot.
Your evidence / decision / next action:
Calculate appearance rate
For each platform and prompt, divide runs where the company was recommended by completed valid runs. Report the numerator and denominator. Track errors and incomplete runs separately, not as absences.
Keep discovery separate from validation
Select the prompt panel before collecting the evaluation sample. Branded follow-up questions such as "Why not our company?" are useful exploration, but do not count them as unaided recommendations.
What we can responsibly conclude
A company appeared in a stated share of the observed answers under the recorded conditions. We cannot infer customer search volume, a permanent AI rank or the causal importance of a page feature from that alone.
Where to go next
Check whether the cited pages answer the service and location question better, whether your information is accessible and consistent, and whether there is independently verifiable proof. Choose a testable improvement rather than copying every competitor feature.
Want to discuss your findings?
Bring your completed worksheet to a 15-minute call. The worksheet does not send your answers or create a lead automatically.
Book a 15-Minute CallRelated: Our AI-search measurement method