Define separate outcomes
| Observation | What it tells you |
|---|---|
| Search impression or click | A search platform reported exposure or a visit |
| AI brand mention | The business appeared in the sampled answer |
| AI citation | A source link pointed to your content |
| Recommendation | The answer explicitly suggested your business for the task |
| Referral visit | A recorded visitor arrived through an identifiable source |
| Qualified enquiry or sale | A business outcome occurred |
Do not combine these into a single unexplained “AI visibility” number. A citation of a technical article is useful exposure, but it is not necessarily a service recommendation.
Use the platform’s own reporting where available
Review Google Search Console for search performance and indexing. Google’s current generative AI optimisation guide also points to the Generative AI performance report for its Search and Discover AI features. Check the report’s definitions and available dimensions before comparing it with other reports.
Analytics can identify some referral visits from AI services, but missing or altered referral information limits attribution. Reconcile useful visits with enquiries and sales rather than treating every session as equally valuable.
Create a repeatable prompt sample
Build a small question set from actual buyer decisions: finding a supplier, comparing approaches, investigating a problem and researching a sector. Record exact wording, language, intended market, platform, model or mode where visible, date and whether web search was used.
Repeat the same sample over time and retain the answers and source URLs. Record mentions, citations and recommendations separately. Use more than one observation before deciding a change is meaningful; generated responses can vary without a change to your website.
Keep competitors and context in view
For the same question, record which businesses and sources appear and whether the answer is accurate. This can reveal missing service information, weak evidence or an unclear relationship between a person, company and product.
The sample is an observation panel, not a census of all user experiences. Location, conversation history, account settings and model changes may affect results. Do not present a small prompt test as a verified share of the entire AI-search market.
Connect measurement to a decision
Maintain a dated record of content and technical changes, then review the question set and relevant business outcomes over a suitable period. Identify whether new pages were discovered, cited or visited and whether the enquiries fit the intended service.
If a page is repeatedly cited but produces no relevant visits, examine the role it plays in the buyer journey. If service information is wrong, correct the underlying sources. The next action should follow the evidence rather than the desire to improve a dashboard score.
Common questions
Can I measure every ChatGPT or Claude recommendation?
No. Public sampling and referral reporting cover only part of the experience. State the boundaries of the dataset.
Are screenshots sufficient proof of improvement?
They document individual observations. Repeated, consistently recorded tests and business data support a more useful assessment.
Does a ranking tool know the platforms’ internal AI scores?
Do not assume that. Evaluate the tool’s data collection, definitions and limitations rather than treating a proprietary score as a platform metric.