What should you count first?
Choose the result that matters to your business. For a shop it may be a completed, retained order. For a service it may begin with a suitable enquiry, followed later by an agreed job and payment. Write down what makes an enquiry suitable: the right need, location, budget or delivery date, for example.
Keep discovery separate from that result. An assistant mentioning your business, someone visiting a page and someone buying are useful but different observations. A click on “shop” or “contact” does not confirm an order or even a sent enquiry.
Use your existing records: an order system, a simple enquiry log and whichever website analytics you already use. Record the date you change the site, tracking, prices or promotion. This prevents a later improvement from being casually credited to whichever change you remember best.
Where can you find AI-search evidence?
If you use GA4, look for the AI Assistant channel in acquisition reporting, then inspect the landing pages and relevant actions. Google added this classification for recognised assistant referrers in May 2026. It captures identifiable visits, not everyone who consulted an assistant before buying. Google Analytics' product update
In Reports, open Acquisition and Traffic acquisition, then use the Session default channel group to find AI Assistant. Google's AI Overviews and AI Mode are excluded from that channel; their non-ad search visits belong to Organic Search. Keep those routes separate when reading the reports. GA4's current channel definitions
Google Search Console's generative-AI reports provide another view: exposure of your pages within Google's AI features. That exposure is already part of overall performance, so do not add the two totals. Google's reporting guide
Bing Webmaster Tools can show citations on its supported AI surfaces. A citation means your page was referenced there; it does not mean the person visited or became a customer. Bing's explanation
If you use another analytics tool, inspect its referring sources and landing pages for identifiable assistant visits. You do not need every report to begin. Keep the available sources and their gaps visible; a visit without an identifiable source stays unknown.
Before relying on the numbers, submit a clearly labelled internal enquiry or appropriate test booking and check that it appears once in the intended record. Exclude it from customer totals. If booking happens on another domain, ask the provider to check whether source information and the completed outcome survive that move. Keep a monthly line for the period, source evidence, distinct customers, retained contribution, cash spending and time.
What can you ask a real customer?
Add one optional question to an appropriate enquiry or follow-up: “Where did you first hear about us?” Allow a free-text answer and an honest “can't remember”. If the person says an AI assistant, ask what they were trying to find and whether anything about your business was unclear.
Avoid asking customers to provide private conversations. A short voluntary description is enough for ordinary business learning. Keep it with the existing enquiry or order, under your usual data practices.
The question concerns remembered discovery, while analytics concerns an observed website visit. A customer may discover you in an assistant, return through Google and pay later. Both records can be true. Neither should create a second customer in your count.
How do you read a small, messy result?
Here is an invented example for a local workshop business. Over four weeks, analytics identifies 18 unpaid assistant visits. The booking system records three linked bookings: one is the owner's test and two are paid customer bookings. Both customers attend. Their tickets are £50 each and identified variable costs are £20 each, leaving £30 contribution per place before the discovery work and other fixed costs.
Three people answer the discovery question. One of the two already-counted buyers says ChatGPT; another paying customer, whose website source is unknown, also says ChatGPT and attends under the same price and cost conditions without a refund; the third respondent has only enquired.
The evidence is two retained buyers linked through analytics and one additional buyer who reports AI discovery. That is three distinct buyers with different kinds of source evidence, not five buyers from “two tracked plus three survey responses”. The enquiry remains an enquiry. The 18 visits are sessions, not 18 verified prospective customers.
Their combined contribution is £90 if each bought one place under the stated costs. Two bookings provide £60 of analytics-linked contribution; the additional £30 depends on the customer's account of discovery. Preserve that distinction.
Suppose the owner spent £40 in additional cash and five hours preparing the pages. The £90 associated contribution exceeds that cash by £50 before other costs. A modest next test is worth considering within an affordable cash and time cap. This cost-coverage calculation cannot establish £50 of additional profit caused by AI search or support a confident growth forecast, because some buyers may have come anyway, the time has value and other fixed costs remain.
What is worth checking each month?
Use complete, comparable date windows and a consistent definition of a completed result. Check the same season last year where the business is seasonal. Note refunds, cancellations, internal tests and the normal delay from enquiry to sale. Keep paid advertising on assistants separate from unpaid referrals.
For the pages you changed, ask whether relevant people arrived, understood the offer and reached the intended next step. Check a small, stable set of genuine customer questions in a search-enabled assistant. Save the question, date, service and cited page. Look for errors in your name, price or service area. This is a sample of answers, not your share of the entire market.
Use the finding to choose the next job: correct an inaccurate price or repair a broken booking form when relevant visits reach it. If citations concern questions outside your market, reconsider which content you are producing.
When should you do more?
Continue a small amount of work when it attracts suitable demand at a cost you can afford, or resolves a clear obstacle for customers. Set a spending and time limit for the next period. At £30 contribution per extra retained booking, two genuinely additional bookings would exceed a £40 cash budget; that arithmetic does not predict those bookings or value your time.
When the numbers are too small to interpret, keep observing and talk to buyers instead of publishing dozens more pages. Use the broader weekly launch review to decide how much attention AI search’s likely business contribution deserves alongside the channels customers already use.