Measuring AI recommendations for a shop without clean click tracking

What the clicks miss when an answer names your shop

An answer that names your product can end in several ways that leave no referral. The buyer searches your shop's name a day later, types your address from memory, buys the same product on a marketplace, or sends the link to someone who orders on another device. Each of those reaches your reports as direct traffic, a branded search or a sale somewhere else. That is zero-click search as a shop meets it.

Google's own AI features add a blind spot of their own. Google's page on AI features says clicks from AI Overviews and AI Mode are counted in overall search traffic in Search Console, in the Performance report under the Web search type, so the click counts there do not separate them from other Google results. Since August 31, 2026, a generative AI performance report has separated one thing: how often those two features showed a link to your site, page by page. Whatever analytics counts as assistant visits is therefore a floor, and a shop needs readings that reach above it.

Reading one: the panel, by product question

The panel measures what no other reading can: whether you are named at all. Put the same product questions, written by need and condition, to ChatGPT, Claude, Gemini and Perplexity each month, three runs each, and record whether the shop, the brand or the product was named, where the answer pointed, and whether the facts it stated were right. AI visibility for online shops covers how to write the questions, and how to measure AI visibility covers repeated runs and noise.

Report it per product line rather than as one number for the shop. A line whose pages changed and a line whose pages did not are the comparison the fourth reading needs, and a single shop-wide visibility score hides it. The first month's reading is the baseline every later one is read against.

Reading two: assistant referrals, counted as a floor

Group the assistants' referral hosts into one channel, as spotting AI traffic in Google Analytics describes, then read that channel the way a shop needs to: landing pages by product line, sessions that reached the cart, and orders and revenue from those sessions, with the purchase set up as a key event. The conversion rate of those sessions, beside the rate for the rest of your traffic, says whether the answers send buyers who were ready.

Treat any tag or parameter on an incoming link as a clue about that link, never as a count. Whether an assistant marks its links, and how, is its operator's decision and can change without notice, so a report built on a tag measures the tag. The referral host, where the browser passes it on, is the steadier signal, and it is still only a floor.

Reading three: a question on every order

Ask every buyer, once, how they found you, and keep the answer with the order. It is the one reading that reaches the buyer who read your name in an answer and came back another way. Put the question on the order confirmation page or in the checkout, make it optional, and never pre-select an answer. Offer a short fixed list, with a box for anything else:

• An AI assistant (ChatGPT, Claude, Gemini, Perplexity or another)
• A search engine
• A marketplace
• Social media
• A friend or family member
• Something else, with a box to say what

Keep the wording and the list fixed, because a changed question breaks the series the same way a changed panel does. Store the answer with the order number and the product line, and nothing about the buyer that the order does not already hold. Then read it monthly: the share of orders answering "an AI assistant", by product line, beside the number of orders that answered at all.

Know its limits. Buyers who skip the question are not counted. A buyer may remember the last thing that sent them rather than every step, and one who asked an assistant and then searched may tick either. Read the share as a direction over months, never as a precise count.

Reading four: cohorts before and after a change

When you change something, compare like with like. Write the change and its date in a log (the product pages of one line rewritten on [date], say), choose a line you left alone as the control, and compare the same number of weeks before and after for both lines: panel readings, assistant referral sessions and orders, and order answers naming an assistant.

Read the difference between the lines, not the change in one of them. If both lines rose, the season, a promotion or the assistants moved them, and your pages cannot claim it. Change one thing per line per period, keep promotions out of the comparison weeks where you can, and for a seasonal product compare the same weeks of the year.

The four readings side by side

When the readings agree, the direction is probably real. When the panel rises and the order answers do not, the answers may be naming you for questions your buyers do not ask, or naming you with a wrong fact; the accuracy column will show which. When referrals rise and orders do not, look at the pages those visits land on, and at the price, shipping and returns they find there.

What none of the readings can show

The buyer who read your name, never said so and ordered anyway is in none of the four. Nor is the one who bought your product on a marketplace, or the gift buyer who asked on someone else's behalf. Report the four readings as a direction with its range, not as revenue credited to AI; the one-page report shows the shape, and the break-even sum turns an order count into a decision.

If you would rather have a panel run for you, monitoring reads twelve fixed questions with one assistant, Claude, each month for $290 a month, which you can cancel in any month. The other three readings come from your own analytics and your own orders, and the overview for online shops shows the work they are measuring.