How to make AI recommendations for your shop measurable – without clean click tracking

Why your click tracking fails with AI recommendations

Imagine someone asks ChatGPT: "Where do I get sustainable running shoes under 120 euros?" and your brand is named. The user then types your shop name directly into the browser or clicks a link without UTM. In your analytics this shows up as "Direct" or "Organic". The revenue is there, but the source is invisible. This is exactly where classic e-commerce tracking, which for years relied on clean referrers and campaign parameters, breaks down.

The problem is structural, not just technical. Many AI answers happen completely without a click: the user reads that your shop carries three models, compares in their head and decides later. Between recommendation and purchase lie days, a different device and often a direct visit. No cookie, no referrer, no session chain. For an online shop this means: the channel that is currently growing fastest is the most poorly measured one.

The consequence is dangerous. If you do not measure AI visibility, it looks like nothing in the report – and you cut budget exactly where new customers come from. So before you look for a perfect tracking that technically does not exist, you need replacement signals: indirect measurements that together yield a robust picture.

Step 1: Measure the recommendation itself, not just the click

The first measurable metric is not traffic, but the mention. Before anyone clicks, the AI has to name your shop at all. That is exactly what you can test systematically. Build yourself a list of 30 to 50 purchase-intent prompts that your real customers would ask: "best cable management for desk", "buy vegan protein powder without sweetener", "gift for coffee nerds under 40 euros". Ask these prompts weekly in ChatGPT, Perplexity, Google AI Overviews and Gemini.

Document three things per prompt: Are you named? At what position? And in what context – as a recommendation, as an example or only as a footnote? From this you build a simple share-of-voice metric: in what percentage of your purchase-intent prompts do you appear, compared to your three biggest competitors. This number is stably measurable, even if not a single click is tracked.

Important for shops: separate generic prompts from category prompts. "Where do I buy sneakers" is hard to win, "sustainable barefoot shoes for wide feet" is not. Your niche is where you can realistically appear in AI answers. Measure the niche first, then work your way up.

Step 2: Server logs beat JavaScript tracking

AI systems leave traces before they recommend you: their crawlers fetch your product pages. GPTBot, OAI-SearchBot, PerplexityBot and Google-Extended access your shop. These accesses are in your server logs, not in Google Analytics, because bots do not execute JavaScript. If you want to know whether AI systems even know your assortment pages, the log is your most honest data source.

Filter your logs by these user agents and look at which pages are fetched how often. A typical pattern in e-commerce: the homepage and bestsellers are crawled frequently, deep category and filter pages hardly at all. That then explains why the AI knows your main products but never recommends your longtail niches. This gap turns into concrete work: internal linking, a clean sitemap, less parameter chaos in the URLs.

It is also worth a second look at Perplexity and Bing-based systems, which sometimes do live fetches. If a Perplexity access appears in your logs shortly after a user prompt, and a purchase without a referrer follows, you have a strong indirect signal – with AI traffic you currently rarely get closer to genuine attribution proof than this.

Step 3: Ask your buyers directly – post-purchase survey

The most underrated measurement method in e-commerce is the simplest: a single question after the purchase. Show "How did you find us?" on the thank-you page or in the first transactional email and include "ChatGPT / AI assistant" explicitly as an answer option. Many shops discover this way for the first time that 4 to 12 percent of new customers came via an AI recommendation – numbers that stood in no analytics interface.

This self-reported attribution is not perfect, but it closes exactly the gap that cookies and referrers leave open. What matters is that you link the answers with the actual order value. Then you see not only that AI customers come, but also whether they spend more or less, which products they buy and whether they come back. For the internal budget discussion this is often more convincing than any dashboard.

Keep the question short and optional so you do not disrupt your conversion. A dropdown with five options is enough. Evaluate monthly and watch the trend: a rising AI share is your best proof that work on AI visibility pays off.

Step 4: Read cohorts and direct-traffic patterns

Even without clean tracking, AI traffic leaves patterns in your aggregated data. Watch for a growing share of "Direct" sessions with strikingly long, specific dwell times on exactly the product pages that match your AI recommendations. If your barefoot-shoe page suddenly gets more direct visits while you appear there in Perplexity answers, that is not coincidence but correlation you should document.

Build cohorts by landing page. New direct visitors who enter on a deep product or guide page instead of the homepage often behave like AI-recommended users: they arrive pre-informed, bounce less and convert faster. Compare this cohort over time with your prompt-test results from step 1. If both curves rise together, you have largely solved your attribution puzzle.

Supplement this with brand-search data from the Google Search Console. An increase in searches for your shop name plus product category is a classic side effect of AI recommendations: the AI names you, the user googles you later. This "assisted brand lift" is measurable, even when the original AI contact is not.

Step 5: Optimize specifically what the AI pulls from your shop

Measuring is half the battle; the other half is giving the AI citable material at all. AI systems recommend shops whose product information is unambiguous, structured and fact-rich. For you this means: product descriptions with concrete facts instead of marketing prose, real dimensions, materials, use cases. "Waterproof to 10 meters, 42 grams, for wrists from 14 to 20 cm" gets cited. "The perfect companion for your adventure" does not.

Implement structured data cleanly: Product schema with price, availability, reviews and GTIN. This helps not only classic SEO but gives AI crawlers machine-readable facts they can adopt directly into answers. Add real comparison and guide content – "Model A vs. B for beginners" – because it is exactly such pages that assistants cite when users seek purchase advice.

Measure the effect of this work again via step 1. If you factually tidy up a category and appear in more prompts three weeks later, you have a clean before-and-after comparison. That way GEO for your shop becomes iterative and verifiable instead of a blind flight.

Step 6: An AI-visibility dashboard that works without a pixel

Bring the individual signals into a common picture. A pragmatic AI dashboard for online shops has four rows: first, share of voice from the weekly prompt tests; second, crawler accesses of the AI bots from the server logs; third, the self-reported AI share from the post-purchase survey; fourth, brand search plus direct-traffic cohorts. None of these numbers is proof on its own, but together they yield a robust direction.

Update the dashboard monthly and link it to your revenue figures. The decisive sentence for your team is not "We had 800 AI clicks", but "Our AI share of voice rose from 18 to 31 percent, the self-reported AI share of new customers from 5 to 9 percent, at a stable order value". That is the language with which AI visibility in the shop becomes budget-worthy.

Keep the method honest: mark clearly what is measured and what is estimated. Precisely because clean click tracking is missing, you gain trust by naming the uncertainty instead of feigning a precision the data does not support.

Step 7: Link AI visibility with your cart value

Pure visibility helps you little if you do not know what it moves in the cart. So take a second look at the order value of purchases that land with you without a classic click source. If in step 4 you formed direct-traffic cohorts, attach to each cohort the average cart, the return rate and the number of line items per order. That way you see whether AI-recommended buyers shop for higher-value items or only reach for entry-level products.

In practice, a clear pattern often shows: whoever comes via an AI recommendation already knows your product by name and buys more purposefully. At a kitchen shop, the cart of this direct cohort was around 18 percent above average, because the AI recommended complete sets instead of individual parts. You make exactly such relationships visible by segmenting the order value by cohort – entirely without a pixel, only with your shop and order data.

Limits: What you cannot measure with this method

Be honest with yourself about the gaps. Here you measure approximations, not clean attribution chains. A direct visit can come from an AI recommendation, but just as well from a saved bookmark page, a newsletter or an offline conversation. So treat your numbers as a trend, not as proof. A rise in brand-related direct visits parallel to more mentions in AI answers is a strong signal – but not a court verdict.

Second, with your optimizations you are changing a moving target. AI models are retrained, answers fluctuate, and what is recommended today may look different in two months. So plan fixed measurement points, for example the same test questions to the assistants every four weeks. That way you separate real improvements from random noise and avoid rebuilding your entire assortment on the basis of a one-off outlier.

Frequent questions from shop practice

Is it worth it for small shops? Yes, especially then. You do not need an expensive analytics setup, but your server logs, your order data and ten honest test questions per month. The effort is more like one to two hours per week than a dedicated tool budget. With small assortments the lever is even greater, because individual recommended products come through more strongly in percentage terms.

How often should I measure? Set a fixed rhythm: check the logs and cohorts weekly, repeat the AI test questions monthly, reconcile your dashboard with revenue quarterly. And what if the AI does not name your product at all? Then that is your most important finding. Go back to step 5 and check whether your product texts, data sheets and FAQ really answer the questions buyers ask the assistants.