As of November 12, 2026, three numbers frame the same problem. Roughly 5% of purchases on Douyin in October arrived through chatbot links, and more than 20% in categories that require research. JD's assistant drew about 80 million users during the festival. And in instant retail, the number of outlets open at midnight is roughly a third of the number open at noon. Put together, they describe a shelf an assistant can only see partially, and a customer who now asks questions at hours when that shelf is thinnest.
The assistant can only quote what the store reports
An assistant does not browse. It queries. When a shopper asks for cold medicine at eleven at night, the answer is assembled from whatever structured data the retailer exposes: item, price, store, opening hours, delivery window. If opening hours are missing from that data, the assistant either guesses or falls back to the nearest large store, which may be closed. The customer sees a promise, then a cancellation.
This is why on-shelf availability, a metric built for field sales, has become a consumer-facing number. A brand that knows its distribution coverage by region but not by store-hour cannot tell whether a late-night request is serviceable. The reporting that wins contracts in physical retail is not the reporting that answers a chatbot.
Coverage has to be counted by the hour
Most distribution reports treat a store as either covered or not. In an assistant-mediated market, coverage is a function of time. A store counts at ten in the morning and does not count at midnight. Two brands with identical outlet counts can have very different serviceable coverage once opening hours are overlaid, and the difference shows up as cancellations rather than as a coverage gap.
The stock file has to survive the shift change
Real-time inventory is a process, not a dashboard. It requires cycle counting on fast movers, a hard block on selling below a safety threshold, and a handover routine so the night shift knows what was picked. Operators that skip the handover produce a stock file that is accurate at 6pm and fictional at 11pm, which is exactly when the assistant is being asked.
Why the failure is quiet
A stockout in a store is visible: the shelf is empty and the shopper walks away. A stockout inside an assistant is invisible, because the assistant simply does not mention the product, or mentions a substitute the shopper did not ask for. The brand sees a lower share of voice and no complaint, and concludes that demand is soft. The cost is not a lost basket; it is a slow drift in who gets recommended.
There is a second-order effect that matters for category growth. Assistants tend to favour items with clean, complete attributes, because complete attributes make an answer easier to assemble. A brand with patchy product data loses recommendations even when its distribution is fine, and the loss is attributed to marketing rather than to data hygiene.
| Signal | What it hides | Where to look |
|---|---|---|
| Lower share of voice | Item not eligible for the answer | Attribute completeness by SKU |
| Rising cancellations | Store closed or unstocked at order time | Opening hours and stock by store |
| Substitution complaints | Assistant filling gaps without a rule | Substitution policy in the connector |
The three signals look unrelated in most dashboards because they sit in three teams: marketing, operations and customer service. Reading them together produces a single question that a category manager can act on: for this item, at this hour, in this store, can we actually deliver what the assistant just promised?
A short audit that answers it
Start with twenty stores and ten fast-moving items. Check stock and opening hours twice a day for a week, once at midday and once late at night, and compare what you find with what the system reports. The delta is the operating gap. Then repeat the exercise through an assistant: ask the same question a shopper would ask, and note whether the answer matches the reality you just measured.
Consumer demand is shifting toward interfaces that answer rather than display, and the stores that benefit are the ones whose data can be quoted without correction. That is a mundane advantage, built from fields and counts, and it compounds quietly for whoever does it first.
FAQ
Does distribution coverage still matter if an assistant is choosing the product?
More than before, and in a different shape. Coverage now has to be counted by outlet and by hour, because a late-night answer can only cite a store that is open and stocked at that moment.
How do brands audit something like this?
By sampling. Twenty stores, ten fast movers, twice a day for a week. The gap between the report and the reality is the size of the problem.
Is this only a China story?
The data is Chinese because assistant-led shopping is furthest along there. The mechanism is general: any market where an assistant reads retailer inventory rewards the same discipline.
What is the cheapest first fix?
Add an opening-hours field to the distribution report. It costs almost nothing and immediately separates stores that can serve a late order from those that only look able to.
