Measuring AI search visibility: follow the lead, not just the mention.

Start with the outcome your business can use

To measure local AI search visibility, connect three views: a small record of how your business appears in answers, identifiable visits to your website, and the quality of resulting enquiries. Keep those views separate before comparing them. A mention is not a visit, a visit is not a lead, and a submitted form is not automatically a qualified opportunity.

I would agree on a qualified lead before looking at a dashboard. For a local installer, that might mean a real project in the service area, a relevant installation need and permission to discuss timing. A job applicant or supplier enquiry should not count simply because the same form was used.

That definition makes the work useful. The question becomes whether discovery is producing conversations the team can serve, rather than whether a monthly report contains more screenshots.

Build a small, repeatable prompt panel

I would begin with twelve realistic questions: three buying intents, two priority services and two genuine local areas. The intents could be finding a provider, comparing an option and checking a practical constraint. Examples include “Who installs heat pumps in this town?” and “What should I ask before replacing a furnace in an older house?” Adapt them to the business’s actual customers.

Run the panel on the selected platforms on a recorded date, using a consistent account and session approach. Save the exact prompt, stated location, product or mode, whether web search was used, the answer and its cited pages. Separate a linked citation from an unlinked mention. Check whether the cited page actually supports the statement.

Repeat after a meaningful interval and record errors such as an old address or a service you do not offer. This is a modest diagnostic sample. It is not a universal ranking, a measure of total market exposure or proof that every customer sees the same answer. Keep the panel stable enough to compare, then add new questions from real sales conversations.

I would use missing or inaccurate information to guide a repair, not create dozens of pages solely to improve a score. The broader rationale is in why local businesses need a clearer AI search presence.

Read referral traffic without pretending it is complete

Google’s GA4 campaign URL guidance explains how UTM parameters identify campaigns that refer traffic. Use consistent source, medium and campaign labels on external links you control, and confirm that redirects preserve them. You cannot assume an AI platform will use the tracking labels you would prefer.

Review identifiable AI-related sources and landing pages in your analytics. Keep platform sources separate long enough to understand them, and document the rules used to group them. I would never label all referral traffic as AI traffic, or assume that a branded Google search proves an earlier AI interaction.

GA4’s guidance on direct traffic explains that direct/none can include visits without clear referral information. A person might also read an answer, remember your name and contact you later. Those possibilities are reasons to acknowledge uncertainty, not to reassign every unexplained visit to AI.

Test enquiry tracking against successful submissions. A click on the submit button, an abandoned form and a completed request should not all count as the same outcome. Reconcile the report with enquiries actually received, and remove tests, spam and duplicates.

Keep first discovery, the final visit and self-report distinct

Google’s traffic-source scope documentation distinguishes first-user, session and event-level attribution. These answer different questions. First-user data reflects recorded acquisition; session data describes visits; event attribution assigns credit under the selected model. None is a complete recording of everything a customer encountered.

My lead record would retain the first known source, the source associated with the enquiry visit where available, the landing page and the customer’s own account of discovery. Do not overwrite the first source every time someone returns. Equally, do not call first-user data the person’s first exposure to the business; it records the first acquisition identified for that user in that analytics setup.

Ask a neutral, optional question: “How did you first hear about us?” Provide an open response or choices that include search, AI assistant, recommendation, other and unsure. If someone says ChatGPT but analytics records a later Google visit, preserve both facts. Self-report is useful context, not a perfect attribution system.

Use the CRM or controlled lead log for contact details and source notes. Keep names, email addresses and free-text project descriptions out of analytics URLs and event labels. The purpose is to understand the journey without turning every conversation into a tracking exercise.

Work through the numbers before scaling the budget

Consider a hypothetical monthly cohort: 200 identifiable AI-referral sessions produce 10 unique valid enquiries, of which four meet the agreed qualification criteria. Assume the lead log can connect those enquiries to the recorded sessions. That is five enquiries per 100 sessions, and a 40% qualification rate among the ten enquiries.

If the business allocated $600 of content and measurement cost to that cohort, the calculation is $600 divided by four, or $150 per observed qualified lead. It is not customer acquisition cost, because none of the four has yet been established as a new paying customer. It is also not proof that the content caused four additional leads.

The allocation needs a note: content may support other channels and future months, so this is a management convention for this example. Compare consistently, show the raw counts and avoid treating a small month-to-month change as a reliable trend. Four qualified leads becoming three does not, by itself, explain what went wrong.

Follow the cohort through response, appointment, proposal and sale where those stages apply. A modest number of suitable enquiries may be more valuable than a large increase in irrelevant traffic. If leads stall because nobody responds, publishing more content will not resolve the operational problem.

Use a monthly decision meeting, not a visibility celebration

I would keep the review short and end it with an owner for each action:

• Accuracy: what factual mistakes appeared in the prompt panel, and which source needs correcting?
• Discovery: which identifiable sources and landing pages brought relevant visitors?
• Conversion: did enquiry delivery work, and what prevented useful conversations?
• Quality: how many valid leads qualified, and why were the others unsuitable?
• Commercial progress: what happened to earlier qualified leads after the first response?

Use the technical readiness checklist when access or indexing is the problem, and profile and review consistency when basic business facts conflict. If you need a measurement plan tied to real enquiries, talk with AIM about your visibility goals.