You don’t have to wait to measure AI shopping
How Fabric measures AI shopping today, from simulating shopper questions to testing whether better product data drives more traffic and sales.

Nobody denies that shoppers use AI to discover products. The objection we hear is that it can’t be measured: no clicks, no keywords, no clean referral data. Marketers used to say the same about podcast ads, which have no link to click. But the industry figured out a way to measure them anyway, with pixels that match a listener’s download to a later visit to the brand’s site. AI shopping can be measured too. The rails aren’t all built yet, but we’ve figured out some of it already. For AI shopping, the stand-ins for the pixel today are simulated shopper questions and controlled before-and-after tests. The brands we work with keep asking the same questions:
- Which AI engines recommend us, and at what point in the shopping journey do they stop?
- Do AI engines send traffic to our website?
- When an AI agent buys for a shopper, are our products found?
- When we improve our product data, how do we know it made a difference?
Most measurement tools were built for shoppers who type keywords and click links, so they answer few of these well. Here’s how Fabric answers each one today, and what still needs to be built.
Which AI engines recommend us, and when do they stop?
Keyword rankings don’t translate to AI. An AI engine writes one answer instead of a ranked list, and that answer changes with the question, the model and the wording. AEO tools track how often a brand is mentioned, but their traffic numbers usually count only clicks on a link inside an AI answer.
What a brand needs to see is what the AI actually says about its products, and at which step of the purchase. Since brands can’t see real conversations, we simulate them. We ask ChatGPT, Gemini, Perplexity and other assistants the questions a shopper would ask at each stage:
- Category education: “What should I know before buying a robot vacuum?” Does the answer draw on the brand’s content?
- Discovery: “What’s the best robot vacuum under $500 for a small apartment?” Is the brand suggested?
- Comparison: We ask the assistant to compare its suggestions. How often does the brand win?
- Validation: We narrow the list (“it needs to handle pet hair”) or widen it by raising the budget. Does the brand stay on the shortlist, or come back?
- Conversion: We ask where to buy. Does the assistant name a place to buy the brand’s product? For retailers, this is often the only step they appear in.
The point is to find the step where a brand drops out, because each one calls for a different fix. A brand that’s never suggested needs different work from one that’s suggested but loses comparisons. The questions never name a brand, so we see whether the brand shows up and which other brands appear as its competitors. We ask a panel of them on each engine separately and repeat them over time, which shows how strong each brand’s presence is at every step and on every engine.
For multi-brand retailers, the funnel looks different. In our simulations, retailers rarely show up while a shopper is learning about a category or comparing products. They compete only at the last step, when the shopper asks where to buy, and usually win it on something specific: a local store, fast delivery, fitting or repair services. As agents take on more of the shopping, we expect more of that last step to go straight to the brand’s own site. Retailers that aren’t measuring where they appear, and acting quickly when they drop out, risk being left out of AI shopping altogether.
Do AI engines send traffic to our website?
Yes, even when there’s no click to track. Shoppers tend to research in an AI engine, then search Google for the brand it recommended or type in the website address. Many AI apps also don’t identify themselves when they send someone to a site, so those visits land in Direct traffic. Some analysts call this dark AI traffic.
To account for it, we look at the traffic product data can actually affect: visits from AI assistants, which are easy to spot, and organic search. We leave out paid, affiliate and email traffic, since enrichment doesn’t change those.
When an AI agent buys, are our products found?
Google and Shopify’s Universal Commerce Protocol (UCP), introduced in January, lets AI engines and agents search a brand’s catalog and buy directly, without navigating product pages. Shopify has UCP built in, and brands on other platforms build the connection themselves.
The catch is that when an agent searches a catalog through UCP, a product only comes back if its data matches what the shopper asked for. So we run a set of shopper questions against a retailer’s catalog through UCP and count how many return the right products, before and after enrichment. Enriched catalogs are often returned for a wider range of questions in their category. For purchases agents complete through UCP on Google, Merchant Center shows how those checkouts perform.
Did our product data changes make a difference?
Raw traffic can mislead, especially in Q4. Many retailers see visits rise 200% to 400% over the holidays, and most of that is seasonal. To isolate the effect of enrichment, we use a method called difference-in-differences. We compare enriched products with similar products that weren’t enriched, over the same period. The season lifts both groups, so whatever the enriched products gained beyond the others is the effect of the enrichment.
Say traffic to unenriched products doubles from October to December, and traffic to enriched products rises 2.3 times. The season explains the doubling, so the enriched products finished 15% ahead of where the season alone would have put them. That 15% is the enrichment. For one pet retailer we work with, enriching products with supplier data increased total sessions by 17%, measured with this same method.
This works whether an AI assistant picks up new data through its crawler over time or reads the product page live during a conversation. We share these results with customers in dashboards they can track in real time, along with early indicators in the first weeks after an enrichment.
Where the rails aren’t built yet
Simulations show what AI assistants say, not what real shoppers do. They can’t tell us how many people asked a given question, or how many left without buying. Podcast attribution took off once hosting platforms put pixels on every download. AI shopping needs the same rails: assistants that pass along referral data and complete checkouts directly with retailers. As those arrive, we’ll fold them into the same measurement, and it’s the area we’re watching most closely.
Where to start
You don’t have to wait for the rails to be finished. Fabric shows how your products appear across AI assistants at each stage of a purchase, enriches the product data behind those answers, and measures the effect on traffic and sales. The place to start is the AI catalog assessment, which shows where your product data is complete and where an assistant would have to guess.


