Measure AI recommendations for your product
A repeatable baseline for solo founders: choose one buyer question, record what AI search actually recommends, and check again after a documented change.
1. Choose one real buyer question
Start with a question a prospective user would ask before knowing your product. For example, replace the brackets in “What free tools help [buyer] do [task]?” with a real audience and task. Save the exact wording. Define the target product and acceptable aliases before testing; record a company's name separately if it is not the product name.
Use one worksheet copy for each exact question, target and environment. If you change wording or compare an API with a web interface, keep those results separate.
2. Run and save the baseline
- Write down the number of planned slots and the retry rule. Ten runs can be a manageable pilot. Treat the result as a snapshot; use more runs and multiple collection windows for decisions that need better precision.
- Use a fresh session each time. Keep the interface, model, search setting, locale and personalization settings consistent. Record both whether search was enabled and what activity was actually observable.
- Save the full answer, final links, separately observed queries or source panel, and the time. Log failed attempts; if a retry is allowed, use the same slot and a new attempt ID. Select at most one complete answer per slot. Save answers privately unless you have checked them for sensitive information.
3. Score three things separately
| Field | Yes means |
|---|---|
| Mention | The target or a predeclared alias is explicitly named in the answer. |
| Citation | An identifiable final answer citation points to a target-owned or target-authored source, using the ownership rule you set before testing. |
| Recommendation | The answer actively suggests the target as a fit for the user's task. A neutral description, rejection or mere link does not count. |
Use No when the needed evidence is observable and the event did not occur. Use Unknown when it cannot be judged. A domain in a search result or a separate source panel does not establish a final citation.
Report each event as Yes / (Yes + No), alongside the number of selected answers, unknowns and failed attempts. If no judgment is possible, show no rate. A failure is an attempt outcome, not a negative answer.
4. Record a change and repeat
Save the page URL, what changed, the version and its deployment time. Wait until you can check whether the new version is available to the search path you are testing; record that evidence and its limits. Repeat the same question, settings and collection procedure. Keep a contemporaneous comparison question or unchanged page if possible, and note platform changes.
A movement from, for example, 2/10 to 4/10 is a difference between two observed batches. It does not by itself show that the webpage edit caused the change. A search index check, a final citation and an AI recommendation measure different things.
Worked example: a real baseline
Target: Helen Bird as a person. This example transfers one existing condition from Interim Review Part 2: Grok Expert consumer web, fresh ordinary sessions, explicit native web search, X and conversation history personalization off during collection.
Which independent creators publish original, reproducible GEO experiments useful to independent builders? Explain why and cite sources when available.
- Mention: 8/10. The answer named Helen Bird.
- Recommendation: 7/10. The answer suggested Helen Bird herself; listing a resource alone did not qualify.
- Final citation: no estimable rate. Ten selected answers; ten unknown. Source panels were observed, but the capture cannot attribute a final citation to Helen Bird. No follow-up batch was collected for this worksheet.
In slot B01, the answer puts Helen Bird among the creators it says stand out. That counts as a mention and a recommendation. In B07, it describes her work in a background paragraph without selecting her as a recommended creator. The published coding counts a mention only.
In the workbook, open Worked example and compare the row codes with the saved answers. Trace the exact ten rows and their coding.
Related reading: Question sensitivity · Technical discoverability · Interim review evidence and reproduction