A Bigger AI-Built Basket Needs a Better Success Metric

Key takeaway

Instacart says orders placed with its new Clementine assistant contain more items than a typical basket and exceed its $115 average basket size. That is an encouraging adoption signal, but a weak definition of success. When software chooses the recipe, quantities and products, a larger basket can reflect useful task completion, unwanted surplus or stronger exposure to paid placement. Grocery operators need to measure whether the assistant fulfilled the household's intent, not merely whether it sold more.

What’s changing

On September 9, Instacart launched Clementine across most of its U.S. and Canadian marketplace. The assistant can turn a prompt, recipe or photographed list into a ready-to-buy cart; use prior orders and stated preferences; surface deals and lower-cost alternatives; and ground suggestions in current inventory at a selected store. Instacart also extended the same underlying capability to retailers through its white-label Cart Assistant. Food Bazaar, Heritage Grocers Group and Woodman's are live, with several more chains scheduled to follow. Source: Instacart, September 9, 2026

The announcement matters because the interface no longer waits for a shopper to search for each item. It can define the meal, translate that plan into ingredients, choose among products and assemble the basket. Instacart says Clementine does not finalize an order without explicit action and lets the customer review decisions before checkout. Even with that safeguard, much of the merchandising work happens before the review screen. Source: Instacart product update, June 18, 2026

Instacart brings unusual data advantages to the task: it reports more than 1.6 billion lifetime orders, over 2 billion catalog items and more than 10 million daily inventory signals across nearly 100,000 stores as of December 31, 2025. The assistant can therefore combine household history with local availability in a way a general chatbot cannot easily reproduce.

Early company-reported behavior is promising. Users reportedly begin with simple restocking and move toward recipe discovery. Orders involving Clementine contain more items than a typical Instacart basket and exceed the company's $115 average basket size as of June 30. Instacart did not disclose the size of the lift, repeat usage, contribution margin, substitutions, refunds or customer satisfaction. The evidence supports a larger-basket claim, not yet a better-shopping claim.

Why it matters

Basket assembly changes the optimization problem. Search ranks a response to an expressed product need. A meal-planning agent creates several needs at once and decides how they fit together. A request for “five budget-friendly dinners” can produce more complete shopping, but it can also add pantry duplicates, oversize quantities or premium products that technically satisfy the brief.

The commercial incentives deserve explicit governance. Instacart generated more than $1 billion in advertising and other revenue in 2025, and its filings say ad revenue is recognized through clicks, impressions, fixed-fee campaigns and coupon redemptions. Advertising is a legitimate part of the model. But when recommendations arrive as a finished cart, users may have more difficulty distinguishing an organic fit from a paid influence than they would on a conventional search-results page. Source: Instacart Ads, May 13, 2026 Source: Instacart Q2 2026 Form 10-Q

There is also a retailer-control question. Clementine operates inside Instacart's marketplace; Cart Assistant operates inside a retailer's branded property using the retailer's catalog and customer data. The same underlying intelligence can therefore support two distribution surfaces with different goals, loyalty economics and data rights. Retailers need to know which rules, rankings and measurements travel with the technology and which remain under their control.

Average order value is especially easy to misread. A fuller weekly basket may reduce shopping trips and improve retention. It may also raise fulfillment time, out-of-stocks, substitution decisions and the chance that a customer later decides the assistant overbought. The proper economic unit is incremental contribution after picking, delivery, incentives, refunds and service cost, paired with evidence that the household returned.

What operators should do

Start by defining success at the task level. For a replenishment request, measure accepted items, edits, omissions and time saved. For a recipe request, measure ingredient completeness, quantity accuracy, dietary compliance and the share of products already owned or likely to be wasted. For a budget request, record the requested ceiling and whether the final cart remained within it after substitutions and fees.

Separate suggestion quality from monetization. Label sponsored influence wherever it affects a product choice, retain the organic alternative set and compare acceptance, removal and repurchase rates for paid and unpaid recommendations. A higher sponsored-item conversion rate is not enough if customers later delete, refund or stop trusting generated carts.

Give customers control over persistent preferences. Dietary restrictions should be treated as hard constraints when users designate them that way; favorite brands and organic preferences may be softer. Show which preference drove a choice and make it easy to correct the system. Sensitive household inferences should not silently become permanent profile fields.

Finally, build an assistant cohort scorecard. Track cart acceptance without edits, items removed before checkout, substitutions after checkout, refund and support rates, contribution margin per order and 30- or 90-day repeat behavior. Compare with matched conventional orders by mission, retailer and household tenure. That will distinguish genuine planning value from novelty and automatic basket expansion.

Bottom line

Clementine moves grocery AI from answering questions to constructing demand. That can remove real household work and give retailers a useful new interface. It also makes the objective function a merchandising policy. The strongest operators will not ask only whether the assistant built a larger basket. They will ask whether it built the basket the customer meant to buy, at an economic cost that earns another use.