When an agent does the grocery run, where does the impulse buy go?

Every grocery store in America has a few feet of shelf that exists because people get bored. The checkout lane sells candy, gum, mints and magazines to a person who is standing still with a cart and a minute to spare. It is one of the most reliable bits of retail design there is, and it rests on a human being physically present at the end of the trip.

So it is worth paying attention when the company behind M&M's and Snickers says that trip is changing. In a report published in October, Mars said routine grocery shopping is heading into the background, run by AI agents. In an interview Digiday published on November 11, its global chief customer officer, Neil Reynolds, described how the company is preparing.

We don't think the impulse buy dies. We think it moves, and that most snack, treat and beverage brands are still planning for where it used to be.

What Mars actually predicted

The report, developed with Kantar and Synthesis, analyzed more than 180,000 data signals: 11 stakeholder and expert perspectives, more than 30 prior research reports, about 170,000 channel and shopper publications and 4,999 podcasts, collected from January 2024 through December 2025. It is a trend synthesis, not a survey of shoppers, and it should be read that way.

Its central claim is a split. Routine purchases become "a frictionless background service managed by agentic technology," while stores and digital platforms become places people visit for discovery and fun. For snacking specifically, Mars says brands and retailers will need new answers "as legacy browsing and impulse shopping behaviors give way to predictive fulfillment that happens in the background." The release puts the job in one line: retailers "must both earn a spot on automated lists and find new ways to spark spontaneity."

In the Digiday interview, Reynolds was measured. Digiday reports he does not expect AI to kill the in-store checkout lane, but does expect agents to take over more of the weekly grocery run. Mars's first step, he said, is getting a starting point for how its categories and brands show up in AI models: "how do we make sure our brands are discoverable by the LLMs and how do we make sure that once those brands are discovered, that they convert through to purchase?"

Shoppers are not handing over the wallet yet

Before anyone redesigns a snack portfolio for robots, the consumer data says the background-agent future is further off than the forecasts suggest. Circana surveyed more than 4,600 US adults; FoodNavigator reported the headline numbers on November 10.

Show the data

What shoppers said	Share of shoppers
Comfortable using AI to compare prices	53%
Influenced by AI to make decisions	34%
Trust AI agents to buy without their approval	16%

Source: Circana, via FoodNavigator, Nov 10, 2026. Survey of more than 4,600 US adults. Figures as reported by FoodNavigator.

Sally Lyons Wyatt, Circana's chief advisor for consumer goods, summed it up: "People are happy to let AI handle the hassle, just not the moment the money leaves their wallet." Her colleague Kiara Barrett predicted that autonomous replenishment "will be earned task by task, not adopted as one giant leap."

That fits what we have seen in other data. In our post on the handoff, Visa's survey found a similar gap between using assistants and trusting them to pay. The near-term pattern is not an agent buying groceries unseen. It is an agent assembling a basket and a person approving it.

The retailer agents are already doing grocery

The other half of the picture is that the agents doing the grocery run are mostly the retailers' own. Amazon said on November 6 that more than 400 million customers used Alexa for Shopping in the past 12 months, and that the number using its agentic features has grown more than eight times since the start of the year. Its new text option is pitched squarely at groceries. The examples Amazon gives are worth reading closely, because each one shows where a brand gets chosen.

Table: Grocery requests from Amazon's own examples, and where the brand choice happens (our reading)

Request, as written in Amazon's announcement	Who names the product	Where an impulse or treat brand can enter
"deliver eggs, bacon, pancake mix, and maple syrup in 30 minutes"	The agent picks brands for generic items	The occasion: a brunch basket can include extras
"I'm out of dog food, need a birthday gift for my sister, and we're low on paper towels"	Mixed; repeat items may follow history	Suggestions sent back as it finds them
"reorder the laundry detergent I bought last month"	The shopper's history	Almost none; repeat beats novelty
"add oat milk to my grocery list"	The agent, unless a brand is named	The list itself, if the shopper writes a brand
Text a photo of a recipe to add all the ingredients	The agent	Generic ingredients; brand is the agent's call

Requests are quoted from Amazon's November 6, 2026 announcement; the last two columns are our reading, not Amazon's.

The store keeps a job

It would be easy to read the Mars report as a eulogy for physical retail. It is closer to the opposite. The second half of its forecast is that stores and digital platforms become "immersive playgrounds designed for discovery and entertainment," and Reynolds told Digiday that Mars is working with retailers on store-of-the-future ideas for snacking. He also described a practical change in how Mars sells: "Why is our data team not talking to the data team at Walmart?"

For a smaller brand, the takeaway is that the two jobs pull in different directions. Discovery and trial still happen where people browse, in stores, on social feeds and in video. Repeat and routine purchases drift to agents. A treat brand needs both: a reason to be tried, and a name clear enough to be reordered without anyone looking at the shelf.

Where the impulse buy goes

Reading those examples next to the Mars and Circana data, we see three places the unplanned purchase moves to.

1. The occasion request. "Plan a last-minute brunch" is not a list; it is a brief. The agent decides what a brunch needs. Snacks and treats have always sold on occasions, from movie night to game day to the office party, and that is exactly the kind of request assistants now get. A brand that is consistently associated with an occasion, in the reviews, recipes, articles and retailer content assistants read, has a chance to be added to the basket. That is the same principle as tip 20: get recommended for specific needs, and an occasion is a need.

2. The add-on at the approval step. If most shoppers approve the basket rather than letting the agent buy blind, then the approval screen or message is the new checkout lane. It is a moment of attention at the end of the trip, with the basket in front of the person. Whatever the agent suggests there ("anything else for the brunch?") is the closest thing to the candy rack. Brands cannot buy that slot directly today, but retailers will almost certainly sell it, and the brands with clean data and clear pack sizes will be the easy ones to suggest.

3. The standing list. Earlier this month we argued that the shopping list is the new search box. For treats, the list is where an impulse becomes a habit. A shopper who adds your brand by name to a saved list once has handed you every future run. A shopper who adds "candy" has handed the choice to the agent. The brands that win here are the ones people name, which is an old-fashioned brand-building problem.

What we would do as a snack or treat brand

None of this needs a new agency or a new budget line. It needs a few existing jobs pointed at new places.

• Write for occasions, not just products. Make sure the pages assistants read connect your products to the occasions you sell into, in plain words: party size, lunchbox, movie night, road trip. That includes your own site, retailer content and the recipes and roundups you appear in.
• Make every pack size unambiguous. An agent adding a treat to a basket needs to know which of your 14 SKUs fits. Name, count, weight and per-unit price should read the same on every retailer, as tip 48 suggests for Amazon titles.
• Fix price and promotion data first. Circana's Barrett said many AI trust problems "are actually product data, pricing and inventory problems." An agent that suggests your product at a price that ended yesterday loses the add-on and the shopper's trust in the suggestion.
• Get on the list by name. Repeat-order and subscription features favor whatever the shopper picked last time. Promotions that get a first purchase through a retailer's app, where it then enters order history, may be worth more than they look.
• Find out where you stand now. Mars is starting with a baseline of how its brands show up in AI models. That is the right first step at any size. Ask the assistants your shoppers use for occasion baskets in your category and write down who gets added. Teams do this by hand, or in tools like SearchPanel, from XENA Intelligence.

The honest uncertainty

Mars's report is a forecast built from what people have been writing and saying, and Circana's numbers describe attitudes, not purchases. Neither tells us how fast agent grocery ordering will grow. Amazon's numbers are about its assistant broadly, not grocery orders, and Amazon has not said how many orders the text option has placed.

What the evidence does support is the direction. Routine purchases are drifting toward agents, slowly and with a person still approving. The checkout lane will be around for a long time. But the next impulse buy for a lot of households will happen at the end of a conversation, not at the end of an aisle, and the brands that think about that moment now will be the easy ones for the agent to suggest.