How to Measure Whether AI Can Recommend Your Business
Businesses want a simple answer to a reasonable question: when a customer asks an AI assistant for a recommendation, do we appear?
The temptation is to run one prompt, take a screenshot, and declare success or failure. That is not a reliable measurement system. AI answers vary with provider, model, location context, wording, time, available web evidence, and product surface. A business can appear in one response and be absent from the next without any underlying change.
Useful AEO measurement accepts that variability and records observations honestly.
Define the questions before testing
The question set should reflect real customer intent, not prompts designed to force the company name into the answer.
A local business usually needs a compact set covering several angles:
- Broad category intent, such as the best options in a location.
- A high-value service or product.
- A supported specialty or differentiator.
- Geographic intent at the city or neighborhood level.
- A direct best or recommended request.
Five to ten questions are usually enough for an initial operating set. The owner should review them so the measurement reflects the customers the business actually wants to serve.
Preserve the exact context
Every observation should record the provider, model surface when available, exact question, location context, and time. Without that information, later comparisons become anecdotes.
Location deserves particular care. “Best pizza near me” is not reproducible unless the system records what “near me” meant. A service-area question should state the intended city or region. A multi-location business should measure locations separately rather than treating one answer as the brand’s universal result.
Use honest result states
The system should not infer a precise rank unless the provider actually returns one. A practical classification is simpler:
- Absent: the business did not appear in the observed answer.
- Mentioned: the business appeared, but not as a clear recommendation.
- Recommended: the answer explicitly presented the business as a suitable choice.
- Cited: the business or its supporting evidence appeared in the provider’s citations.
- Unavailable: the provider failed, timed out, or returned insufficient evidence.
Unavailable is not a negative score. It is a missing observation.
Keep providers separate
A result from one answer engine is not proof of another. A Gemini response grounded in Google Search or Maps measures that environment. A ChatGPT response with web search measures a different environment. Perplexity, Claude, Copilot, and other systems have their own retrieval and presentation behavior.
Combining all providers into one opaque number hides the most useful information. A business needs to know where it appears, where it does not, and which sources each provider relied on.
Provider-specific reporting also prevents a dangerous overclaim: no API connection to one model can cause a different model to recommend the client.
Capture sources and competitors
When the provider supplies citations, store them. They help explain why a business was selected and which public sources are shaping the answer.
Competitor appearances are also useful when handled carefully. The objective is not to copy a competitor’s claims. It is to understand the evidence pattern. Does the selected business have clearer location data? More relevant reviews? A stronger specialty? Better category alignment? A trusted third-party source that the client lacks?
The answer should lead to a specific, legitimate improvement rather than generic content production.
Compare periods, not isolated prompts
An observation history becomes more useful after a meaningful change. The business may correct a location conflict, publish an approved specialty, improve crawler access, align important profiles, or earn a new credible citation. The question set can then be run again.
The comparison should avoid deterministic language. One improved response does not prove a permanent ranking increase. Repeated improvement across several relevant questions and observation dates is stronger evidence that the public field is becoming clearer.
Turn measurement into action
Measurement creates value only when it produces a prioritized next step. A good report should answer:
- What did the provider show?
- What evidence did it cite?
- Which customer intent was not served?
- What legitimate business fact or authority signal is missing?
- What is the single best action to take next?
The purpose is not to promise control over an external model. It is to create a disciplined improvement loop: observe, diagnose, make an evidence-backed change, publish it safely, and observe again.
That is how AEO measurement becomes more than screenshots. It becomes an operating system for reducing uncertainty over time.