The practical AI Visibility guide
How to measure AI visibility without mixing up the signals.
Start with a question set and a measurement definition. Keep model recommendations, search citations, and customer outcomes separate so each result tells you what to do next.
Put it into practice
Define the experiment
Name the audience, questions, and system.
Measure AI visibility by observing a fixed set of relevant questions in a defined system, recording the answers, and calculating clearly named signals from completed observations. Start with one audience and one buying situation. For example: five-person agencies choosing a way to collect design feedback and final sign-off.
Choose which questions matter before looking at scores. Separate a model-knowledge baseline from a web-search check. Write down the exact model, language, geographic constraints where relevant, and whether the business is named. A result from one API model is not automatically a result from a consumer chat product with search and personalization enabled.
Check the prerequisite
Establish business clarity before recommendation fit.
In AEOchamp, add your domain and run Domain Clarity. Review recognition, category, audience, job, and confidence. Save the answer that explains a mismatch. If the model misunderstands the business, recommendation scores alone will not tell you what needs clarifying.
Next, review the five baseline questions in User Intent. Choose the available baseline model and let both analyses complete. Organic tests the buyer question without introducing your brand. Mentioned supplies the domain so you can inspect assessed fit. Champion lets you compare baseline intents across all eligible models and add custom questions.
Recognition
2/ 3 modelsTwo models recognize the domain.
Category agreement
2different categoriesRecognizing a name is only the first step.
Live comparison
3website checksCompare existing knowledge with the current site.
What each model thinks your business does
Select a model to read its answer and compare the live website check.
| Model | Category | Audience | Purpose | Confidence | Status | ||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
61 |
|||||||||||
|
Training answer · OpenAINo website visit I associate example.com with a project management consultancy that advises large businesses on team processes. Same model · live website checkWebsite inspected
The current website gives this model a clearer description of the business. |
|||||||||||
88 |
|||||||||||
|
Training answer · AnthropicNo website visit I recognize example.com as client approval software for creative agencies, helping teams collect feedback and approve designs. Same model · live website checkWebsite inspected
The live check adds a more specific audience to an already relevant description. |
|||||||||||
| — | |||||||||||
|
Training answer · GoogleNo website visit I do not have enough existing knowledge of example.com to identify its business reliably. Same model · live website checkWebsite inspected
The current website gives this model a clearer description of the business. |
|||||||||||
This preview uses illustrative data. In your actual analysis, note which results have completed and which are still pending, failed, or locked. An unfinished result is missing evidence, not a zero recommendation score.
Keep a small evidence log
Record enough to reconstruct the comparison.
| Field | Example | Why it matters |
|---|---|---|
| Question and intent version | “Which tools help agencies collect design feedback?” / v1 | A wording change can alter the answer. |
| Model and mode | Selected model / training knowledge / Organic | Do not pool search-grounded and ungrounded results. |
| Date and status | 7 October 2026 / completed | Keep failures outside the completed-results denominator. |
| Result | Organic 32; Mentioned 84; matched gap +52 | A recommendation score is not a mention frequency. |
| Evidence | Answer text and the relevant competitor comparison | Keep enough context to challenge the diagnosis. |
| Action and page | Clarify approval records on the workflow page | Connect measurement to a reviewable task. |
If you separately count mentions, define mention rate as qualifying completed answers containing your brand divided by qualifying completed answers reviewed. If you count citations, decide whether the link must point to your domain or can point to a third-party page about you. Use separate columns. The mentions and citations guide shows how to classify the cases.
Compare like with like
Freeze the question set before calling a change an improvement.
Keep a stable core of questions. When you add a new use case, report it as a new segment rather than silently changing the overall average. Keep the same model subset when comparing periods, or show the subset change beside the result. Otherwise, a newly included model can move an average even if the original models are unchanged.
Model outputs can vary for the same input. OpenAI’s evaluation guidance recommends task-specific evaluation and human review. For a decision that matters, inspect repeated comparable observations where available rather than treating one changed answer as proof that a page edit caused it.
Choose the follow-up
Use the finding to change a page, not just a report.
| Finding | First action | Useful follow-up |
|---|---|---|
| Wrong category or audience | Clarify the homepage and About description. | Run a live Domain Clarity check after publishing. |
| Low Organic, strong Mentioned | Read the question’s model answers and relevant competing use cases. | Improve the page for that buying situation; keep the baseline question stable. |
| Low Organic and Mentioned | Check genuine product fit and missing workflow detail. | Explain capabilities and limits before chasing mentions. |
| Citations but few qualified enquiries | Review the cited page’s intent and customer path. | Inspect analytics and enquiry quality separately from model scores. |
For Google’s own generative AI search performance, Google directs site owners to Search Console reporting. Keep that evidence alongside your AEOchamp diagnosis rather than treating model recommendation scores as Google impressions. Finish with an owner, a changed URL, and a review date.
Your questions, answered
What should an AI visibility baseline include?
Record the buyer questions, exact model or search experience, browsing mode, analysis date, completed results, underlying answers, and the definition of each metric.
How often should I measure AI visibility?
Choose a review cadence you can sustain. Recheck current website interpretation after a meaningful page change. Compare model-knowledge results when the question, model, and analysis context are comparable.
Can I measure Google AI visibility with AEOchamp?
AEOchamp measures model knowledge and live website interpretation. It does not directly measure Google AI Overview rankings, impressions, clicks, or citations. Use Google’s reporting and actual search observations for those signals.
From advice to a clear next step
Make your first baseline useful.
Use Domain Clarity and five baseline User Intents to document what the models understand and which buyer questions need attention.
Start your free AEOchamp check