Google Merchant Center Briefly Exposed Its Own AI Prompts Instead of Insights

Google Merchant Center briefly showed the actual AI prompt behind one of its insights reports, instead of the polished summary it's meant to generate — a rare, if minor, look behind the curtain at how Google's AI-generated retail insights actually work.

What was spotted

Vipul Kumar posted the find on LinkedIn: "Looks like Google Merchant Center accidentally gave me a little peek behind the curtain today... Instead of just showing the final AI-generated insight, Merchant Center was actually displaying what looked like the instructions being used to generate it."

According to Kumar's screenshot, the exposed prompt scaffolding included distinct sections for:

• What the AI was supposed to analyse — the scope of data it was told to look at
• The rules it needed to follow — constraints on tone, format or claims
• Click changes for 20 products — the specific dataset fed into that instance of the report
• How it decided which product to highlight — the selection logic behind picking one product as the headline insight
• Templates it could use for the response — pre-built sentence structures the model fills in rather than generating freely
• The sentence it was going to generate — a preview of the model's planned output before final rendering
• A final check before showing the result — a validation step run after generation, before the insight reaches the merchant

Why it matters

This isn't a security issue or a meaningful leak of proprietary logic — it's a rendering glitch that briefly showed prompt scaffolding instead of the finished output. But it's a useful, concrete glimpse into how Merchant Center's AI-generated insights are actually built: a structured, template-constrained pipeline with an explicit selection step and a post-generation validation check, rather than a single freeform generation handed straight to the merchant. That structure explains why Merchant Center's AI insights tend to read in a fairly consistent, formulaic style — it isn't writing freely each time, it's filling in one of a fixed set of templates after running through a defined selection and rules process.

Worth keeping in mind next time a Merchant Center AI insight reads as slightly generic or templated: that's evidence of how the system is built, not a bug in a specific report.

Sources

• Vipul Kumar on LinkedIn — original screenshot and spot