AI Shopping Visibility: Tracking What AI Recommends at Product Level

"Best LED strip for a workshop", "top robot vacuum under 300 euros", "what should I buy for X" — purchase-intent prompts behave differently from informational ones. The answers are list-shaped, product-named and price-anchored, and the engines lean harder on structured sources: product pages, shopping feeds, review aggregators. A generic prompt set won't surface how you perform here; the shopping set has to be tracked as its own population, with its own metric. That metric is recommendation rate: of the shopping-prompt runs in which any tracked product was recommended, how often was it yours? It's measured per brand and per product, with the same sampling discipline as everything else — repeated runs, sample sizes shown, and a dash rather than a zero when a product hasn't accumulated enough decisive runs to say anything honest.

Brand-level shopping visibility answers "do the engines like us"; product-level answers "which of our products do they actually sell". The gap between the two is routinely the finding. A catalog's hero product can carry the entire recommendation rate while the margin-makers never appear; or engines recommend a product you're phasing out; or they recommend yours — described with a competitor's spec. Product-level tracking means the catalog lives in the tool: each product with its aliases and SKU variants (engines rarely use your exact product name — matching "V-TAC LED Strip 24V Warm White" requires knowing its aliases), competitor products marked watch-only, and per-product recommendation and win rates over time. The win rate — of runs where any tracked product was recommended, yours won — is the shopping equivalent of share of voice, and the number a merchandising team can actually act on.

Here is the supply-side reality check from the account in our hero image. Of 200 product pages, the audit sampled 40; six were unreachable to the checker (its own finding). Of the 34 auditable pages: 3 carried Product schema, 3 carried Offer/price schema, and zero — none — carried review-rating schema. This is a real, trading shop with real revenue. Why it matters mechanically: when an engine composes "top X under Y euros", price-anchored answers need machine-readable prices. A product page without Offer schema forces the engine to parse prose, and engines under time pressure don't — they quote the competitor whose price was structured. Same for ratings: "best rated" answers are assembled from review markup, and a catalog with zero review-rating schema has opted out of every such answer. The audit turns this from theory into a fix list: which pages, which schema types, in priority order.

Nobody should type 200 products into a tracking tool. The catalog imports — Google Shopping XML, TSV or CSV feed, or synced directly from Merchant Center — and the prompt side generates from it: shopping prompts built around your actual categories and price points rather than guessed ones. Worth stating what this doesn't do: importing a feed doesn't make engines recommend you, and we make no such claim. It makes the measurement complete — every product that could be recommended is being watched, so when merchandising asks "does AI ever recommend our mid-range line", the answer is a number with an n, not a shrug.

Shopping is the AI-visibility surface with a transaction at the end of it, and the engines are building toward that transaction — ChatGPT has shopping results, price comparisons and, increasingly, checkout ambitions; Google's AI surfaces pull straight from Shopping data. When agents buy on behalf of users, recommendation rate stops being a marketing metric and becomes distribution itself. Which is the argument for measuring now, while your competitors' product pages are as schema-poor as everyone else's: the fixes are unglamorous (markup, feeds, aliases, liftable specs) and the measurement tells you which fix pays. Start with whether the engines can even quote your prices — the free GEO audit includes the schema check — and if you sell more than a handful of SKUs, track the shopping prompts as their own set before the category gets crowded.