Written by: Nicole Ogonowska, IT Growth Manager, Digital Colliers
September 2026 marked the moment conversational commerce hit grocery at scale. Instacart launched Clementine, Shipt rolled out its own AI shopping assistant, and suddenly millions of users could shop by asking instead of searching. For brands selling through these platforms, the shift created an immediate measurement problem. You can't optimize what you can't see.
The attribution black box
Traditional eCommerce attribution is already messy, but at least you know when someone clicked your ad or searched your brand name. AI shopping assistants introduce a new interface layer where all of that visibility disappears. A user asks "I need ingredients for a healthy weeknight dinner for two" and the assistant picks products, builds a basket, and completes the order. From your perspective as a brand, you see the sale. You don't see why the assistant chose your pasta over the competitor's, or whether it even considered you.
The broader AI measurement crisis makes this worse. 95% of enterprise GenAI pilots deliver zero measurable P&L impact, and most of that failure traces back to the same root cause: teams can't connect the AI system to concrete business outcomes. When the AI assistant is owned by the platform and sits between you and the customer, the measurement gap becomes structural.
Three questions you cannot answer
The data you need breaks into three distinct asks, and right now platforms are not surfacing any of them:
Which orders came from the assistant? You can see total sales on Instacart or Shipt, but not which transactions originated from a conversational query versus traditional search or browse. Without that split, you're flying blind on whether this channel matters to your business at all.
How did the assistant choose you? When it recommends your product, what drove that decision? Was it price, margin the platform earns, past purchase behavior, inventory availability, or something else entirely? You need the selection logic to understand where you sit in the ranking and what levers you control.
What optimizations actually work? If you adjust your product title, enrich your description, or change your price, does the assistant start recommending you more or less? Customer acquisition cost across DTC brands has risen roughly 40% since 2023, so every new channel matters, but only if you can measure and improve performance inside it.
What you actually need from the platform
The winning move is not to wait for platforms to volunteer this data. They won't, at least not in a form you can use without asking. You need to open a direct conversation with your account manager and request three things:
First, a weekly report that tags conversational orders separately from search and browse. The format can be simple: order ID, timestamp, entry point. You just need the split.
Second, transparency on recommendation logic. Not the full algorithm, just the inputs: which attributes the assistant weighs when it considers your category. If it prioritizes organic certification over price, you need to know that so you can decide whether to highlight it in your listing.
Third, a feedback loop. When you make a change to your listing or pricing, you want to see how it moves your share of assistant-driven baskets over the next two weeks. The platform already has this data for its own optimization. You're asking for a subset of it, scoped to your products.
The operators shipping this are asking now
The pattern I keep seeing is that brands who escalate early get partial visibility, even if it's informal. Your account manager has access to queries that mentioned your product, even if they can't share the full dataset. A monthly call where they walk you through anonymized examples of how the assistant is positioning you gives you enough signal to start testing.
The risk is waiting until this becomes standard practice. By then, competitors who got visibility earlier will have spent six months optimizing their listings for conversational discovery while you were still guessing. More than 80% of AI projects fail, and a large share of those failures trace back to teams who assumed the data would arrive on its own. It doesn't. You have to ask for it, and you have to ask now.