Lesson 10 of 12 · Module 10 · Measure observations
Design repeatable AI visibility checks
What it is
An answer is an observation under particular conditions—not a universal ranking. Start with 5–10 genuine questions before considering the manuscript's advanced 100-prompt experiment. 1. Use actual customer wording and neutral prompts; naming your brand in every question inflates results. 2. Separate brand mention, recommendation, direct citation linking to your domain, third-party cited page that actually mentions you and competitor mention. 3. Record platform/model, date, locale/language, prompt, retrieval/browsing state and reset context. Follow platform terms. 4. Repetition is an optional method, not an SEO rule. 100 prompts × 10 runs = 1,000 observations per platform; four platforms = 4,000. Check budget, permitted access and current limits first. 5. Keep valid responses separate from errors/timeouts. Missing data is not brand absent. Verify sources actually support claims. 6. Rate = outcome count / valid runs × 100; no valid runs means unavailable. Six mentions in ten valid checks is 60% in that sample, not market share or 60% of shoppers. 7. Repeated outputs can be correlated; model, settings, date and personalization changes affect them. Record methodology changes. 8. Compare observations with referral, survey and sales data. Do not merge different platforms/modes into an unexplained AI rank.
Why it matters
Helpful resources • Internal: AI Visibility saves selected-client questions, provider responses and available sources. Sign-in/client selection and confirmed applicable credits required. Provider/API checks are not identical to consumer ChatGPT, Gemini, Perplexity or Claude experiences; direct responses may differ from web-enabled consumer search. Inspect mode and sources. • Free: a private manual-observation worksheet. Consumer free tiers have limits; a 100 × 10 batch scheduler is not included. • Source: OpenAI crawler purposes.
Free exercise workbooks
This lesson: AI Observations.
PDF: print and write by hand. Spreadsheet: edit in Excel or import into Google Sheets. Keep completed copies private; answers are not uploaded.
✓ Good example
A fictional retailer reports six mentions in ten valid checks with dates/settings and lists two failed requests separately.
✗ Common mistake
Calling a single answer a ranking or presenting API output as identical to personalized consumer search.
You pass when: You have prompts, outcome definitions, conditions and a budget-aware pilot plan. Report counts, valid totals, source checks and limitations.