This is the fourth in a 5-part series, You Can’t Prompt Produce, on the AI grocers keep shipping, and most shoppers keep skipping.

Why are grocers pouring budget into customer-facing AI when shrink alone eats close to their entire profit margin?

That should stop every CFO in the room. It doesn’t, because a personalization demo looks better in a slide deck than a shrink report ever will. I’ve sat through both pitches. Only one gets applause.

The front office mirage

Grocery tech conferences all look the same now. Presentations promote smart carts, AR aisle navigation, and even some concept pitched as “the metaverse of grocery” that nobody in the audience asked for.

Please note: this is NOT an argument against customer-facing AI. It’s definitely needed to secure the next generation of grocery shoppers. A personalization tool that lifts basket size a couple of percent is a real win. I won’t pretend otherwise.

But that lift, for a business running a 1.6% to 2.1%1 net margin, is still small next to what's sitting untouched in operations. The gap comes down to ROI. It's real, just not as dramatic as people make it when they're chasing a headline instead of running the math.

Some of this is pure ego. A CEO announcing an AI shopping concierge gets a stage and a press hit. Nobody claps for “we cut spoilage in the dairy case by 8%,” even though that number shows up in EBITDA every quarter, on time, without a demo.

The grocers who thrive will balance investments in AI that customers can see with the fixes they’ll never notice.

Where the real money sits

The real money sits in parts of the store nobody photographs. The five AI projects below carry more weight than their price tag suggests, and none of them needs a customer to open an app.

Dynamic Markdown Pricing: Pricing algorithms are tied to SKU-level expiration data to catch a product hours before it becomes a total write-off. Forty cents recovered on a dollar you were about to lose outright is margin with no acquisition cost attached.

Automated Shrink Detection: Shrink prevention runs bigger than most operators say out loud. Computer vision at checkout catches sweethearting and scan avoidance in the moment. Point the same tech at the loading dock, and it flags inventory gaps before finance has to ask what happened. Point it at the produce aisle and less of it ends up in the dumpster.

Demand Forecasting: Predict SKU-level demand down to the store and the day, including weather and local events, then trigger reorders automatically. No marketing campaign will ever do that. Fewer stockouts walk straight to Walmart. Less overordering spoilage rots on the shelf.

Predictive Maintenance: Refrigeration and HVAC get ignored because they're the least glamorous line item in operations, and that gets expensive fast. A compressor failure in a frozen foods case doesn't just cost the repair; it costs everything in that case at once, and the sensor data that would've flagged it three weeks earlier usually just sits there unwatched.

Micro-Fulfillment Routing: Batch picking at the store level cuts labor cost per online order. Digital fulfillment bleeds margin badly right now. This is the one lever actually pulling it toward breakeven.

Running the numbers

Take a $1 billion chain running with a 1.8% net margin. That’s $18 million in profit. Shrink runs 2.7% of revenue on average2, so $27 million is already gone before anyone opens a strategy deck.

Cut shrink 100 basis points, from 2.7% to 1.7%, and you recover $10 million. That was money already spent on product that left the building and earned nothing. Getting it back drops almost straight to net income. Profit goes from $18 million to $28 million, a 56% jump from one fix.

Give the front office side its best case. A well-run personalization push drives a 2% conversion lift on $100 million in digital sales, a result most retailers would happily take. That’s $2 million in extra revenue. At a 1.8% margin, that’s about $36,000 in real profit, before licensing fees and the months it took to build.

Ten million against thirty-six thousand. I gave the front office example every benefit of the doubt, and the gap still isn’t close.

These usually come from different budgets and different teams, so it’s rarely a true either-or on the org chart. But someone still decides where the next dollar goes, and where the next round of exec attention goes, and that person is picking the wrong side of this math.

Where this leaves us

Grocery runs on thin margins and high volume. The operators still standing in five years won’t be the ones with the slickest app. They’ll be the ones who also fixed shrink, forecasting, and fulfillment while everyone else just demoed a chatbot on stage.

AI isn't a strategy here. It's a capability, and it only pays off when it's aimed at what's actually bleeding your margin.

Ten million dollars doesn't need a demo. It just needs someone to stop chasing the shiny things long enough to fix the boring ones.

Tell me where I'm wrong.

You Can’t Prompt Produce is a five-part series on how the AI grocers keep shipping, and most shoppers keep skipping.

1. The AI Mass Adoption Mismatch

2. The 95% Graveyard

3. Your Chatbot Sucks

4. Ten Million Against Thirty-Six Thousand

This is the third in a 5-part series, You Can’t Prompt Produce, on the AI grocers keep shipping, and most shoppers keep skipping.

Ask any grocery chatbot to rebuild last week’s order for your family, but now has a peanut allergy. Watch it recommend a snack bar that’s out of stock. Watch it swap in a “nut-free” substitute that isn’t, because nothing in that decision chain might have checked the label against the allergen flag.

That’s a real task. Complex tasks like meal planning, dietary restrictions, and rebuilding a cart from purchase history are exactly what a chatbot is supposed to do well. And it still gets it wrong.

Please note: this is NOT another rant against chatbots. They have an important role in securing future generations of loyal customers. It’s not “chat versus search for a gallon of milk.” Nobody’s trying to replace the produce grid with a text box, and no grocer is dumb enough to force that fight.

The focus is on ensuring your chatbot delivers the hard stuff it was built for.

Half the room hasn’t shown up

Even Walmart’s own research supports this. They found 47% of consumers would trust a digital agent to buy household essentials within a set budget. That number drops to 38% once you get specific about food and groceries, the lowest category they tested1. The company running the most visible AI shopping assistant in the country is telling you, in its own data, that groceries are the category people trust it with least.

And that’s despite real usage. Roughly half of Walmart’s app users have tried Sparky, and those users carry a noticeably higher average order value than the ones who don’t. So it’s not that nobody’s touching these tools. It’s that even heavy usage hasn’t closed the trust gap for the one category that matters most here.

Somebody else’s chatbot isn’t your chatbot

Here’s the part that gets skipped in the boardroom deck. National chains can build in-house, fund the engineering, and iterate on hallucinations weekly. Most regional and independent grocers can’t, and they know it.

That’s exactly why Associated Wholesale Grocers, a distributor, built a shared AI shopping tool, SmartMeals, in partnership with Breez AI, and handed it down to its independent grocery members2. Not because those grocers wanted a wholesaler’s chatbot. Because the alternative was falling further behind better-funded national chains with no realistic path to catch up on their own.

That’s the two-tier gap in its purest form. A middleman had to step in because small operators structurally can’t compete on this front alone, no matter how much they want to.

Where it actually breaks

Three specific failure modes keep showing up, and none of them are about typing being annoying.

Substitution logic is a known pain point across the industry. RetailWire’s coverage of online grocery substitutions found real, sustained shopper frustration, with cases where the replacement item sent wasn’t even a plausible stand-in for what was ordered3. That’s not a rare glitch. It’s a structural problem with how these systems guess at alternatives.

Inventory disconnection is the mechanical reason substitution keeps failing. Grocery inventory turns over multiple times a day, and pricing can shift within hours, but many retailers still only sync their product feeds overnight. A bot working off a stale feed will confidently recommend something that sold out that morning, and there’s no amount of clever prompting that fixes a data pipeline problem.

Hallucination is the third leg, and it’s the hardest to catch because it looks confident. The bot invents a product, a price, or a promotion that doesn’t exist, and the shopper only finds out at checkout.

The wrong answer costs more than a bad search result

There’s a reason getting this wrong stings more than a clunky menu. A shopper who hands a chatbot something complicated, such as a dietary restriction, a budget, or a week of dinners, is trusting it to think, not just to click. They’re not skimming a shelf on autopilot. They’re trusting the system to reason through something that actually matters.

When that trust gets a confident but wrong answer, the damage isn’t friction. It’s a shopper who bought the wrong bread for a kid with celiac disease because the tool she was told to rely on didn’t actually check.

The gap between the demo and the deployment

Vendors aren't lying when they show off a slick model in a sales meeting. They’re showing you the model’s ceiling, not the floor it hits once it’s stitched into your actual inventory system.

Baymard Institute ran a version of this test on the auditing side of the house. When they had GPT-4 evaluate live ecommerce pages against the same usability issues human experts had already found, the model caught only 14% of the real issues, and 80% of what it flagged was a false positive4. That’s the easier job: spotting a problem. The harder job, the one your chatbot is being asked to do, is not spotting friction but removing it in real time against a live catalog.

Retailers are moving faster than shoppers are willing to follow them. FMI’s early 2026 survey found more than two-thirds of food retailers now say they’re using AI, up from less than half the year before5. Meanwhile, groceries sit at the bottom of consumer trust in the very same Walmart data above. That gap will matter more than any individual product launch.

Build the chatbot worth trusting

Chat is a reasonable tool for genuinely ambiguous asks like meal planning around a budget, substitutions when shelves are empty, or shoppers working through dietary restrictions. These are genuinely hard problems, and language is a fine interface for solving them.

The mistake isn’t building a chatbot. It’s benchmarking yours against a company with ten times your data and a hundred times your engineering budget, then acting surprised when your version hallucinates a product or misses an allergy. Here’s what actually separates the two.

1. Sync your feed in real time. An overnight sync means the bot is recommending things that have already sold out. If yours updates in hours instead of minutes, find out why.

2. Write the substitution rules yourself. Leave allergen and diet swaps to a model’s guesswork and eventually something dangerous gets through, not just something annoying. Write the rules and own them.

3. Get a real number on how often it’s wrong. No vendor hallucination rate means no one’s actually measured it, and you’ll be the one who finds out in production. Ask what the bot does when it isn’t sure, too. A tool that always has an answer is worse than one that occasionally says it doesn’t know.

4. Figure out if it’s actually built for grocery. Plenty of what’s on the market is a general model wearing a grocery skin. This is usually where a national chain’s tool and everyone else’s start to look nothing alike.

5. Leave the door open, and ask if anyone wants in. Give shoppers a real way back to search and grids. And find out if they even want this before you bet the launch on it, because high usage with low satisfaction just means people feel stuck.

6. Decide what success looks like before you scale. Pick your baseline metric first: cart completion, time to checkout, number of complaints. Then put someone in charge of the failure log every week, with a number that actually triggers pulling the plug.

Get those six right, in-house or with a vendor, and you’ve got something people can rely on.

A downloadable version of these six rules, with supporting questions for each rule, is available on the Resources page.

Your turn, shoppers

What’s the dumbest thing a grocery AI has ever recommended, hallucinated, or gotten flat wrong for you? Maybe it swore an out-of-stock item was sitting in your cart, swapped in a substitute that ignored an allergy, or invented a price that never existed.

Drop it below.

You Can’t Prompt Produce is a five-part series on how the AI grocers keep shipping, and most shoppers keep skipping.

1. The AI Mass Adoption Mismatch

2. The 95% Graveyard

3. Your Chatbot Sucks

4. Ten Million Against Thirty-Six Thousand

This is the second in a 5-part series, You Can’t Prompt Produce, on the AI grocers keep shipping, and most shoppers keep skipping.

A grocery chain’s board approved a seven-figure budget for a GenAI pilot last year. Nobody on the exec team can tell you exactly which use case it was supposed to solve anymore. The tool is still live, but nobody’s opened it in months.

That’s not a hypothetical. It’s the default outcome across grocery right now, and the industry keeps issuing press releases like it isn’t happening.

Most of that capex never got signed off by “the board” in the vague way people like to blame. A CTO or CIO built the business case, a CFO signed the number, and the board wanted the comfort of knowing its company wasn’t the one that sat out.

Grocers still run on razor-thin net margins. That’s the environment these AI pilots got funded in, which makes the whole thing worse, not better.

Somewhere in a testing environment right now, a chatbot pilot from eighteen months ago is quietly dying. Nobody killed it. Nobody’s writing the postmortem. It just stopped coming up in the QBR.

That’s the graveyard. Almost nobody wants to admit how full it already is.

The Data Nobody Wants On The Board Slide

Here’s the number that should open every grocery AI strategy meeting instead of getting buried three slides deep in an appendix.

MIT’s NANDA initiative concluded that despite $30–$40 billion in enterprise AI investments, 95% of generative AI pilots produce zero measurable impact on the P&L.1 The researchers looked at the results of 300 publicly-disclosed AI deployments, surveyed 350 employees, and conducted 150 executive interviews. and the failure wasn’t about model quality. These failures largely came down to integration and priorities that never connected to an actual business problem.2

(Please note: this stat has taken its own share of hits since it went viral, with critics saying the methodology gets oversimplified. Fair warning before you drop it into a board deck cold. That said, it doesn’t undercut the pattern grocery operators are living through though, and the research below backs that up from a different angle.)

Gartner isn’t offering much cover either. Last year, the firm expected over 40% of agentic AI projects to be canceled by the end of 2027, blamed on ballooning costs, murky business value, and risk controls nobody bothered building before launch.3

Meanwhile, the money keeps moving. FMI reported that US grocery capex-to-revenue ratios have ballooned to 1.9%4 (with Walmart hitting 3.5% to 4%). Over 80% of food retailers are actively expanding technology spend to drive operational efficiency.

Capex is up, cancellation risk is also up, and most of what’s funded never touches the bottom line. This is not a strategy. It’s a lottery ticket with a seven-figure buy-in, and grocers keep buying more tickets.

Grocery also carries three handicaps most industries never have to deal with. Their ERPs are older than cloud computing itself, the supply chains have dozens of handoff points where data quality rots, and the industry margins are thin enough that a six-month ROI delay kills a pilot before it gets a fair shot.

None of that will shows up in the pitch deck.

The Boardroom Fight Nobody Talks About Publicly

Picture the room. A grocery CTO is staring down a renewal contract for a “predictive AI platform” that hasn’t moved a single operational metric yet.

Across the table, the vendor’s slide deck still says “transformational.” Their renewal ask is up 22% year over year. And nobody on that side has set foot in a distribution center floor since the deal closed.

The CTO asks the only question that matters: what did this move on shrink, forecast accuracy, labor hours saved? The answer pivots into softer language: foundational infrastructure, the next phase, features slated for “Q3” that were supposed to ship eighteen months ago.

The vendor can't answer because the pilot was never built to. It was built to answer a question asked a year earlier, several floors up: would this company show up in a trade article as the retailer that did nothing while a competitor announced an AI push?

The board bought insurance against that headline. That's reputational risk management, and it's why metric-free pilots are still getting funded a year after the failure data went public.

The CTO isn’t the villain. Neither is the vendor rep. Both are negotiating over a metric that was never the actual goal of the money behind it.

A pilot funded to protect a headline was never going to survive a real P&L review.

Three Rules For Killing Vanity AI Before It Kills Your Budget

Grocery leaders don’t need a new framework. They need permission to say no more often, and something better than a headline to hand the board.

1. No pilot gets funded without a named operational metric before day one. Shrink rate works. Forecast accuracy works. Labor hours per case picked works. If nobody can name the specific P&L line it’s supposed to move, it doesn’t get capex, it gets a whiteboard.

2. Every pilot gets a 90-day kill clause, no exceptions. That’s a hard date on the calendar. By day 90 the project shows a measurable delta against its named metric or it dies. Vendors will fight this one hard, and that fight is exactly how you know you’re negotiating seriously.

3. Back-office AI gets funded before anything customer-facing. Demand forecasting, inventory accuracy, shrink reduction, these have an actual ROI track record in grocery ops. Chatbots and generative shopping assistants photograph better in a board deck. They rarely pay for themselves.

None of this is complicated. It just gives a board something better than a headline to point to the next time a competitor announces its AI initiative. A capex rule nobody can second-guess in an earnings call beats a chatbot demo every time.

A one-page version of these three rules, with text to paste straight into a vendor contract review, is available on the Resources page.

Has your organization seen actual, measurable P&L yield from GenAI pilots, or are you just burning capex to satisfy the board?

You Can’t Prompt Produce is a five-part series on how the AI grocers keep shipping, and most shoppers keep skipping.

1. The AI Mass Adoption Mismatch

2. The 95% Graveyard

3. Your Chatbot Sucks

4. Ten Million Against Thirty-Six Thousand

This is the first in a 5-part series, You Can’t Prompt Produce, on the AI grocers keep shipping, and most shoppers keep skipping.

Scroll through any conference keynote or ‘thought leader’s’ post this year, and you’ll hear the same assumption stated as settled fact: everyone uses AI now, daily, instinctively. Many point to Pew’s latest number: 49% of US adults now use AI chatbots.1

That number is not nothing. It’s just smaller and stranger than the hype implies. Grocers have the most at stake here, because their actual customer base looks like the excluded half of the country, not those on the tech conference stage.

What “using AI” actually means

Let’s get back to that number. Pew found 49% of US adults now use AI chatbots, up from 33% in 2024 and 23% in 2023. ChatGPT alone reached 44 percent of adults, up from 18 percent in 2023 and 34 percent in 2025.

But what do they mean by “use”?

Dig deeper, and you’ll find that only 24% use chatbots daily. Half the country has never opened one. Another chunk opened one once, maybe out of curiosity, and never went back. Trying something and living inside it are different behaviors, and the topline adoption number conflates them into one figure.

Layer on age and that gap stops being a footnote and starts being the whole story. 66% of those 18 to 29 ‘use’ it, and that number drops to 42% for 50- to 64-year-olds, and 23% for those 65 and older. Reverse those numbers, and you get the plain sentence nobody puts in a keynote: most Americans over 50 have never used a chatbot.

Education plays a role. A Brookings analysis released alongside Pew’s numbers found a steep education gradient, with generative AI use hitting 67% among adults with a bachelor’s degree or higher vs. 24% for those without.2

And geography adds a third axis. Microsoft’s county-level diffusion data for early 2026 found metropolitan counties carry an average AI user share of 32.9%, more than twice the 16.2% average in rural counties.3 Rural America is using AI at roughly half the metro rate.

(None of that is the same conversation as passive AI, such as the recommendation feeds and spam filters running quietly in the background. That layer is functionally universal. This isn't the same as opening a chat window and typing a prompt.)

Why the industry can’t see it

Full disclosure: I live in San Francisco, a city that tech journalist Kara Swisher once called “assisted living for millennials.” It’s hard to escape AI in these parts, and that can easily warp perceptions of what happens outside of the Bay Area bubble.

The people making the loudest claims about universal adoption are the same demographic where it's already near-universal: under 40, college-educated, urban, knowledge workers. A chatbot slots into their job at a keyboard; it has no obvious slot in stocking shelves or driving a route. Ask SF Bay Area techies if "everyone" uses AI, and they’ll answer based on everyone they know.

That's not bias. It's sampling error. Silicon Valley is measuring a room and reporting it as the country.

The wrong playbook

Most grocers are trying to mimic the wrong retailers: Walmart, Amazon, and Instacart.

Walmart’s own online numbers make the point more precisely. Lower-income households are still Walmart’s largest online grocery segment by volume. But its fastest-growing slice, the one driving recent gains, is households earning $200,000 or more, growing almost five times faster than the overall base.

That’s exactly where Walmart pointed its AI efforts. In March, it embedded Sparky, its shopping assistant, directly into ChatGPT and Gemini, chasing what CEO Doug McMillon called a “native AI experience.” Amazon and Instacart built the same pattern into their own tools.

That’s three layers of mismatch, not one. Online grocery is still only about 19% of total US spend.4 Inside that minority channel, the affluent shoppers Amazon, Instacart, and Walmart are chasing are a minority too.

All three retailers have the talent, scale, and capital to make those bets. Most grocers don't, but they keep copying that playbook anyway. They build in-house chatbots that go unused, aimed at a shopper base that doesn't look like the one Walmart, Amazon, and Instacart are chasing, in age, income, education, or geography.

Copying Walmart’s answer to a different question doesn’t fix it.

The mismatch

The AI mass adoption story is real. It’s just landing almost entirely on one narrow slice of the population, the slice that also happens to write the roadmaps and fund the pilots. Call it a mismatch.

And Silicon Valley is still clueless about who’s actually in the grocery aisle.

Is your non-tech inner circle, or your actual store’s shopper base, prompting AI tools on a weekly basis? Or are we all just describing the same twelve people in San Francisco to each other?

You Can’t Prompt Produce is a five-part series on how the AI grocers keep shipping, and most shoppers keep skipping.

1. The AI Mass Adoption Mismatch

2. The 95% Graveyard

3. Your Chatbot Sucks

4. Ten Million Against Thirty-Six Thousand

Here’s the problem nobody puts in the board deck: every order routed through a marketplace app, white-labeled or not, is an order where the grocer never saw the search, never touched the loyalty data, never got a cut of the ad revenue. As digital penetration climbs, the ownership of the customer quietly disappears underneath it.

Grocers can still turn this around, but not by trying to out-market Instacart. The move is to make their own tools the default habit, so a shopper opens them before they even think about a competitor’s. Whoever gets that first click keeps the margin and the customer.

Everyone else is just fulfilling someone else’s sale.

This video is a summary of The Marketplace Trap, a 10-part series on why grocers are handing customer ownership and profit margins to platforms, and what they can do to take it back.

The full series:

1. Ownership > Growth

2. The $1,000 Problem

3. Digital Penetration Is a Vanity Metric

4. Customer Ownership Is the Real Moat

5. Mid-Tier Grocers Are Stronger Than They Think

6. The Instacart Search Question

7. The Retail Media Margin Shift

8. The Loyalty Illusion

9. Defensive Distribution Strategy

10. If I Were a Grocery CEO

    Bonus: The Marketplace Dependence Curve

I laughed hard when I saw this 80 yo brass mechanical grocery list on Reddit. It’s pre-loaded with 40 categories, and you slide the little tabs left to flag what you need. There’s no battery, no sync issues, and no sponsored results pushing a Mongolian butter brand you’ve never heard of before. You walk the store, you slide the tabs back as you go, and you’re done.

Then I stopped laughing, and PTSD set in.

I spent months working on a few grocery list pilots for a national grocer. We had beautiful designs, detailed stakeholder decks, and a product manager (me) who genuinely believed we were going to change how households shop. Every project cratered. The plan was solid. The technology wasn't there. Other teams weren't either.

That brass tab-slider from the 1940s still works better than anything I’ve seen demoed since. It had two states per item, no account to create, and no location permissions to dismiss.

The biggest competitor in grocery list tech isn't Instacart. It isn’t an 80-year-old piece of brass. It's a pen and the back of an envelope.

The Kroger app has a 4.8-star rating. So does Walmart Grocery. Instacart sits at 4.9. These apps do a lot of things: weekly ads, coupons, pharmacy, fuel points, order tracking. And yet the grocery list, the feature every household actually uses every week, is the worst part of all of them. A 4.8 with a broken core feature means the bar for grocery apps is pen and paper.

We are using 2010-era digital checklists in an era of predictive AI, and giving them five stars because they didn’t make us cry in the cereal aisle.

The Friction Is the Feature (And That’s the Problem)

Picture Tuesday evening. You’re standing in your kitchen, halfway through prepping chicken piccata. You’ve got your phone out because you remembered you need capers, and now you’re trying to add them to your grocery app while one hand is covered in lemon juice.

You open the app. It wants your location. You dismiss the prompt. You tap the search bar. The keyboard slides up and covers the “Add Item” field. You tap again. You type “capers.” The app serves you four sponsored results: a brand you’ve never heard of, a bulk 5-pound jar from a restaurant supplier, some kind of artichoke antipasto, and buried fourth, the Goya 4-ounce jar you actually buy.

By the time you’ve added the item, you’ve overcooked the chicken.

That’s not an edge case. That’s grocery list apps performing exactly as designed, which is to say, poorly. It compounds: laggy household sync, dead wifi mid-aisle, item entries that default to wrong sizes, and zero awareness that you just cooked a meal and already know what you need.

Grocery apps were never built around the actual moment of use.

The Ad-Tech Hijack

Here’s what happened. Grocers looked at their apps, saw a captive audience with purchase intent, and turned them into billboards.

Retail Media Networks now generate high-margin billions. Instacart’s retail media arm pulled in $1 billion in 2023. Every major grocer is right behind them. CPG brands pay serious money to sit at the top of your search results.

The operator math is easy. The shopper math is what you feel every time you open the app.

When you type “butter” into your grocery app and get three sponsored results for brands you don’t use, in sizes you don’t buy, sorted by who paid the most, the app has stopped working for you. It’s working against you.

This is where utility breaks down. The grocery list is the highest-intent moment in consumer retail. You’re not browsing. You’re not impulse-buying. You know what you want. And the app’s job, at that exact moment, has been flipped from “find what you want” to “show you what someone else paid to show you.”

Every sponsored result injected into a list search is a direct tax on the person who opened the app. It adds time, creates confusion, and often forces a second search. The grocer has monetized user frustration.

Brain-Dead “Household Intelligence”

There’s a concept in software called “cold start.” No data on the user, start from scratch. It’s a well-known problem in recommendation systems. Netflix solved it over a decade ago.

Every time you open a grocery app, it cold-starts.

It doesn’t know your household goes through a gallon of 2% milk every five days. It doesn’t know you buy coffee filters in 100-count packages and you’re probably running low. It doesn’t know you’ve ordered chicken thighs every third week for eight months.

It has your order history, timestamps, and frequency data. And it does nothing with any of it.

The average American household shops from a stable repertoire of roughly 150 to 300 SKUs, with modest seasonal variation.1 Grocery shopping is almost entirely cyclical, and yet every session starts the same way: blank list, search bar, waiting for you to remember everything yourself.

Recipe integration makes this worse. Most apps that offer recipe-to-list functionality scale the ingredients and then dump them into your list without checking what you already have. There’s no pantry state, and no “skip if in stock” toggle. You get told to buy paprika when you have three jars at home.

The Walled Garden Fallacy

Grocers build their apps on an assumption that is empirically false: that you shop with them exclusively.

Nobody does this. The actual pattern for most American households looks more like Costco for paper goods, Trader Joe’s for snacks and seasonal items, the local grocer for produce and fresh meat, and maybe an ethnic grocery or farmers market on top of that. Four or five stores for a single household’s weekly food supply.2 None of those stores share data. None of their apps talk to each other.

The grocer’s app assumes 100% wallet share. Industry estimates put the typical primary grocer’s actual share of household food spend closer to 16% to 25%,3 directionally, though that number varies significantly by market. A rural household with one dominant grocer nearby looks nothing like an urban household splitting trips between Whole Foods, Trader Joe’s, and a neighborhood butcher.

Third-party apps have tried to solve this. AnyList, OurGroceries, and others position themselves as store-agnostic. The problem is that without deep API integration, they’re just digital paper. They can’t access past orders, real-time inventory, or fulfillment routing. AnyList didn’t fail because the interface was bad. It stalled because a great list app without grocers’ data connections is a notepad with pretty UX.

Grocer APIs, where they exist, are proprietary, poorly documented, and locked behind partnerships that protect the grocer’s own app. It's not negligence. Locked APIs are a margin protection strategy.

An open standard connecting a neutral list app to Kroger, Publix, and H-E-B would make price comparison trivially easy, and that destroys margin. The wall stays up because the wall protects the business model.

Which also means no grocer fixes this voluntarily. Any grocer that opens its API to a neutral list platform is handing a price comparison tool to every household it serves. That’s not a product decision. It’s a business model decision, and the economics run the wrong way.

The Holy Grail

Here’s why any of this matters beyond UX pedantry.

The grocery list is the single most valuable piece of commercial real estate in retail e-commerce, and nobody has properly built on it.

Think about what the list actually represents. You built it at home, before you opened any app, and before any brand got near you. The person who puts capers on a list is going to buy capers. That’s not true of someone who sees a banner ad for capers and maybe clicks through. List conversion isn’t browse conversion. It’s not even close.

That said, this moment is already being contested. Grocer apps push weekly circular notifications. Recipe features inside store apps seed the list with sponsored ingredients. Alexa adds items to an Amazon list by default.

The window of opportunity is real, but it's closing. Whoever captures the list-building session early enough, before the shopper opens the grocer’s app, owns the uncontaminated version of that intent. That window is the prize.

Marketers have a term for the moment a consumer first becomes aware of a need: Zero Moment of Truth. Google coined it in 2011 and built much of its search advertising model around it.

The grocery list is earlier than that. It’s pre-ZMOT: the moment a household commits to a purchase based entirely on internal knowledge of their own consumption, before any brand or algorithm weighs in. The app that lives at that moment owns the most intimate commercial relationship in consumer retail.

Whoever controls the grocery list controls the basket. Whoever controls the basket controls grocery e-commerce.

The obvious question is: why hasn't anybody won this yet?

Amazon is the most credible incumbent answer. Subscribe and Save is a crude but functional version of replenishment cycle intelligence. Alexa captures voice-driven intent at the moment it forms. Whole Foods gives Amazon purchase history across a major grocery format. If anyone had the data and distribution to build what I’m describing, it’s Amazon.

But Amazon can’t be the neutral layer, because Amazon is a grocer competing on its own platform. Kroger won’t open their APIs to a company that sells groceries in direct competition with them. A neutral third party without that conflict is the only entity grocers have any rational incentive to cooperate with, which is exactly why the winner here doesn’t look like a retailer at all.

It looks like a company that already owns the upstream moment. Think of a meal planning platform with tens of millions of active households, or a recipe app that lives on your phone before you ever think about where to shop. Any service that captures what your household intends to eat this week, before any grocer gets near the session, is sitting on the most valuable real estate in grocery commerce.

That company, with the right API partnerships, becomes the demand layer that sits above the entire fragmented grocery ecosystem.

The platform that wins will be store-agnostic, routing list items to the lowest-cost or fastest fulfillment option across your actual shopping ecosystem.

It will have genuine household intelligence that learns consumption cycles, suggests replenishment before you run out, and connects meal plans to list generation without making you own three jars of paprika. It makes money at the list-building stage, through placement that fits the context instead of hijacking it. Suggest the off-brand olive oil when I’m browsing, not when I’ve already typed in the one I want.

These platforms exist. The deep grocer partnerships don't.

The person standing in aisle nine, consulting their phone at 6:47 PM on a Wednesday, is not browsing. They’re executing. Execution requires frictionless retrieval, not an ad for a product they didn’t ask for.

The prediction models exist. The API infrastructure can be built. The household data is sitting in purchase histories that grocers are actively failing to use.

The obstacle was never engineering. It’s that grocers spent a decade monetizing friction instead of eliminating it, and the ad revenue was good enough that nobody felt the urgency to build something better.

And the one company with the scale to force the issue is structurally prevented from being the neutral party grocers would actually work with.

That's the opening, and it's not a feature gap. It's a structural one. And it’s sitting there unclaimed, worth more than anyone inside the grocery industry has admitted out loud.

An 80-year-old piece of brass already figured out the interface. The rest is just the will to build it.

PS: Want to embarrass your kids at the store? Amazon carries a version of this thing: https://amzn.to/4fDPD28 (yes, Bezos cuts me a check). More on the tool itself here.

PPS: If you’d prefer a digital version, a reader vibe-coded just one for you (it’s wonderfully ridiculous): https://mysimplelist.base44.app.

This is the final post in a 5-part series, The Unit Problem, on how grocers can reverse unit volume contraction. Here’s where we started.

Most grocery apps greet you with a coupon library. Browse hundreds of offers, find the relevant ones, clip each manually, and remember them at checkout. Shoppers doing that in 2026 are already doing more work than they should.

One banner figured out a different greeting. She opens the app and sees: “We’ve saved you $4.75 this week. Already applied.” No clipping. No browsing. The grocer did the work before she even wrote her list.

44% of consumers are leaning harder on coupons and promotions right now.1 That number will not surprise anyone reading this. What should surprise you is how few grocers have figured out what that shopper actually wants from the experience.

Every extra click you make her take is unit volume walking out the door to someone who did the work for her.

What AI Is Actually Good For Here

68% of food retailers surveyed by FMI in early 2026 said they use AI in their business, up from 47% a year ago. Technology spending averaged nearly 2% of total sales in 2025, double what it was in 2024.2

Most of that investment is going into consumer-facing AI: chatbots, search improvements, and product recommendation carousels. These are real applications. None of them are where the unit-volume leverage lives. (Review the Grocery AI Pyramid.)

The AI that actually moves units works at the household prediction level. It is not visible to the shopper as “AI.” It shows up as an offer that feels like the grocer actually knows her.

McKinsey’s data shows grocers currently report 35% of promotions are fully personalized, and expect that share to reach 55% in two to three years.3 The grocer that reaches 55% by mid-2027 instead of late-2028 has a loyalty and unit advantage during the exact window when shoppers are most open to switching banners.

How the Engine Actually Works

The architecture is not complicated. It starts with purchase velocity: AI identifies which households are running low on which categories based on purchase cadence, pack size, and days since last purchase. A household that buys a 32-ounce Greek yogurt every 10 days is probably out by day 8. That is a known fact sitting in POS data that most grocers are not using.

From there, the offer writes itself. Not “buy 2 get 1 free” on a shelf tag, but a push notification: “You’re probably out of Greek yogurt. Buy 3 and save $2.50, good until Saturday.” That unit pull is real because the household was going to buy one anyway and is now buying three.

The last piece is basket completion. When a digital cart reflects a recognizable mission but is missing something the household always buys on that mission, surface one targeted offer to close the gap. A household that always pairs chicken thighs with salad dressing but only has chicken in this basket gets one prompt: “Complete the meal, 20% off your usual dressing today.”

These offers are not generic. The unit volume goes up because the prompt answered a question the shopper was going to answer somewhere else.

The 44% Coupon Cohort

The shopper clipping coupons in 2026 is not who you think she is. She is not stretching a tight budget and hoping for the best. She is a high-frequency, high-intent shopper telling you exactly what she values. The problem is the coupon experience most grocers hand her.

A paper coupon in a Sunday insert requires her to find it, clip it, remember it, and present it at checkout. A digital coupon library requires her to browse hundreds of offers, identify the relevant ones, and manually clip each before she shops. Both formats add friction at exactly the moment you want it eliminated.

The move is automatic application. Before each shop, the system identifies every offer in the weekly set that is relevant to the household’s purchase history and applies it to the loyalty account without any action required. She opens the app and sees: “We’ve saved you $4.75 this week. Already applied.” She did not have to do anything to earn that. The grocer did.

The behavioral consequence is straightforward. A shopper who gets that feels recognized. She connects that feeling to the banner. She is less likely to do her coupon sweep at a competitor because the competitor requires her to do the work herself.

This is not a complex feature to build. It is a policy decision about how coupons work, combined with a loyalty data integration that already exists in most modern grocery tech stacks. The tech is easy. The willingness to change is not.

Connecting the Series

The promotional engine is the piece that connects everything. It is the connective tissue that makes everything in this series work.

The mission-based app architecture from the second post generates the basket start. The friction fixes from the fourth post keep shoppers in the app long enough to build a full list. The AI engine fills it with personalized offers and bundle recommendations. The GLP-1 telehealth partnership from the third post feeds new household acquisition into the loyalty database, where the engine immediately begins learning the new shopper's behavior.

These are not four separate strategies. They are one strategy with four execution layers. The grocer running four disconnected initiatives loses to the one that built a system.

The 90-Day Starting Point

Pull 12 months of POS and loyalty transaction data for the top 20 SKUs by household penetration. Run a repurchase interval analysis: average days between purchases by SKU and household segment. Identify the five SKUs with the most consistent repurchase cadence. Those are the first five items for the bundle offer test.

Build a control group of 500 loyalty households and a test group of 500. Push a replenishment offer to the test group for those five SKUs in week 3 of the next promotional cycle. Measure unit pull at the SKU level and total basket impact at the household level over four weeks.

That test costs almost nothing. It runs on existing loyalty infrastructure. It produces a result the CFO can read. And it is the data that justifies everything else in this series.

Grocery unit volume will not recover because the macro improves. It will recover because individual grocers build systems that earn back the trust and trip frequency of shoppers who are currently hedging across four banners and pruning every cart.

The grocers that pull ahead will be the ones that figured out how to win during the contraction. The ones that waited for it to end will find the winners already dug in.

The Unit Problem is a five-part series on why US grocery unit volume is falling and what traditional grocers can actually do about it.

1. The Unit Problem

2. Your Tools Are Teaching Shoppers to Spend Less

3. The Partnerships You Haven’t Made Yet

4. You Fixed the Wrong Friction

5. $4.75, Already Applied

This is the fouth post in a 5-part series, The Unit Problem, on how grocers can reverse unit volume contraction. We began here.

The grocery industry spent five years making the physical checkout faster. Amazon walked into grocery, scared everyone with what frictionless looked like, and the industry responded by throwing billions in capital at the last thirty feet of the shopping journey. With exceptions, most of it failed12.

The physical checkout is not where unit volume is leaking.

The Three Friction Points Nobody Fixed

The shoppers who abandon their carts at digital checkout are visible. The transaction fails, and the data captures it. The shoppers who get three substitutions they did not want on their pickup orders, accept them, and never place another order are harder to see. And the shoppers who built a mental list on Sunday night, opened the app Monday morning, could not find two items, and quietly decided Instacart was easier never show up in the data at all.

There were no failed transactions, no abandonment events tracked. The orders never happened, baskets never got built, and unit volume crashes.

Grocery has not solved that problem. It has been optimizing around it.

There are three points in the digital grocery journey where unit volume actually disappears. Checkout is not one of them.

The first is list-building. The shoppers who plan their trip in the app are the highest-value shoppers in the building. They arrive with intent, buy closer to their full list, and are far more likely to be loyalty members you can actually learn from.

Most grocery apps treat list-building as a notepad bolted onto a storefront. Search returns poor results for anything but an exact match, and there is almost no personalization pulling shoppers toward items they have bought before. They build half a list, get frustrated, and finish it in their heads on the drive over. Half those mental items never make it into the basket.

The second is substitution. Shoppers who place a pickup or delivery order and receive three substitutions they did not want do not complain to customer service. They place their next order somewhere else.

Substitution is where the grocer’s inventory problem becomes the shopper’s problem, handed off with a notification and a hope. The industry has put real money into substitution logic on the fulfillment side, better picker tooling, smarter item matching, and almost nothing on the shopper side.

Shoppers should be able to pre-authorize preferences by category, rank what they will and will not accept, and trust that the system knows they will take Fage but not store brand on Greek yogurt. Most apps give them a binary: accept or reject after the fact. One of those outcomes is a satisfied customer. The other is a churned one.

Grocery has the lowest digital cart abandonment rate of any retail category, around 57%, a figure that has held across multiple industry analyses. Most grocers read that as a win. The likelier explanation is that volume loss is happening before the cart, in the list-building and discovery phase, where nothing gets tracked.

The third is pickup and delivery window selection. This one sounds operational, but it is a conversion problem. Shoppers who cannot find a window within two hours of when they need it either abandon the order or move to a competitor with more availability.

Kroger, Walmart, and Instacart have all invested in dynamic slot management. Most other grocers are still running static windows on a 24-hour refresh cycle. The shopper who wants a 5-6 PM delivery on Thursday and sees nothing until Saturday morning is a shopper Instacart just acquired.

What Checkout Optimization Actually Solved

Digital penetration varies widely by banner and market, but the direction of travel is the same everywhere: the share of orders touching a digital surface before checkout is growing, and the friction points upstream are growing with it.

Faster checkout reduced one real cost: associate labor at the front end. On margins that leave almost no room for error, that matters, and nobody should pretend otherwise. Scan-and-go also improved the experience for a specific type of shopper, the one buying eight items who does not want to stand in any line at all.

Those shoppers were already loyal. They knew what they wanted. They were not the conversion problem.

The conversion problem is the shopper still deciding whether any of this is worth the effort. They tried delivery twice and got bad substitutions. Their app keeps surfacing items that are out of stock. They cannot find a window that fits. They are one friction event away from deciding Instacart is just easier, margin be damned.

No amount of physical checkout optimization reaches that shopper.

Where the Next Infrastructure Dollar Goes

List-building, substitution preference, and slot availability are not glamorous investments. They do not photograph well in a press release. Nobody is cutting a ribbon on a new substitution preference UI.

But these are the three places where the digital grocery shopper decides whether to stay or go. Get list-building right, and unit volume grows among the shoppers worth keeping. Get substitution right, and you stop losing pickup and delivery customers after their second bad experience. Get slot availability right, and you stop sending Instacart new users every time your Thursday evening windows fill by noon.

The physical checkout investments made the stores faster for shoppers who were already there.

Grocers that do not fix these three things are not losing a price war. They are losing to a slow app, a bad substitution, and a full pickup window. That never shows up in the quarterly ops review. It shows up six months later, in a cohort chart nobody wanted to explain.

The Unit Problem is a five-part series on why US grocery unit volume is falling and what traditional grocers can actually do about it. Next, we'll explore the promotional engines and close with planning for the next 18 months.

1. The Unit Problem

2. Your Tools Are Teaching Shoppers to Spend Less

3. The Partnerships You Haven’t Made Yet

4. You Fixed the Wrong Friction

5. $4.75, Already Applied

This is the third post in a 5-part series, The Unit Problem, on how grocers can reverse unit volume contraction. Here’s the first.

Three forces are pulling grocery unit volume out of traditional grocers right now, and none of them are competing retailers.

The first is GLP-1 adoption. The second is gasoline prices. The third is meal planning. Each one is forming purchase intent before shoppers walk into any store or launch an app. Almost no grocer has built the partnerships to intercept all three.

The GLP-1 Basket Is Already Here

16% of consumers report current household GLP-1 use, with another 7% reporting past use, according to McKinsey’s State of Grocery North America 20261. That is not a niche population. That is a meaningful share of grocery-buying households that have reorganized what goes in their carts.

The category shift is specific. After GLP-1 adoption, spending falls in chips, sweet bakery, soft drinks, cookies, candy, and desserts. It rises in nutrition bars, meat snacks, fresh produce, and yogurt. A grocer’s center store is built for the basket these shoppers are actively dismantling. The perimeter is where their spend is going.

The question is not whether to respond. It is whether you own the moment when new GLP-1 users reorganize their grocery identity, or watch it happen from a distance.

That moment happens exactly once per adopter. The first four to six weeks after starting the medication, eating patterns are disrupted, and old purchasing habits break down. Whoever reaches them in that window with a relevant, easy-to-execute basket owns their grocery behavior going forward.

The telehealth platforms that prescribe and support GLP-1 users (Noom, Hims & Hers, Ro, and Weight Watchers) have direct relationships with these shoppers at exactly that moment. A co-branded nutrition program, where new subscribers get recipe ideas and curated digital baskets of high-protein, low-sugar, nutrient-dense items available at your stores, puts you in that window.

The telehealth partner promotes heavily during onboarding and ongoing care. The grocer handles the assortment, the loyalty integration, and weekly basket nudges calibrated to what a GLP-1 household actually buys.

ShopRite has already moved here, launching wellness kits for GLP-1 patients to bridge the gap between pharmacy and nutrition. A wellness kit is a start.

The co-branded basket model with a telehealth platform is the more ambitious version, and it is still emerging. No operator has built it at scale. But the conversation with a platform like Noom or Ro is worth starting now, before a competitor does.

Nearly 90% of grocers expect to increase shelf space for high-protein products by more than 10% over the next two to three years, according to McKinsey. Shelf space without a demand generation strategy is just inventory cost. The telehealth partnership is the acquisition channel that fills the expanded shelf with actual shoppers.

The Gas Price Problem Is Solvable

Kroger and other national grocers have run a fuel rewards program for twenty years. Most operators at every other tier have not. When gas prices climbed over 20% across the US in March 2026, that gap is no longer just a missed loyalty feature. It is a weekly pain point shoppers feel every time they fill up, and grocers that address it directly earn goodwill a competitor sitting on the sidelines cannot match.

It’s not a complicated idea. Every dollar spent in the store earns fuel cents per gallon that’s redeemable at their stations or within a partner fuel network. A loyalty card or app QR code handles redemption at the pump. Shoppers connect the grocery trip directly to relief at the gas station. That connection changes the calculus on trip frequency, driving distance, and banner loyalty.

Gas at 20% above last year makes fuel rewards less of a perk and more of a response. A grocer without a fuel reward is invisible on the issue shoppers care about most right now.

"We know gas is expensive. Shop here and we'll help." Simple, direct, and emotionally resonant for a household running the numbers on where to spend this week.

The Meal Planning Play

The technology that turns recipe content into grocery baskets exists. Many grocers have not plugged into it.

Chicory runs a shoppable recipe network across more than 5,200 food blogs and recipe sites. Its “Get Ingredients” button sits on content ranging from Martha Stewart to thousands of independent food bloggers. Shoppers find a recipe, click the button, select a preferred retailer, and the full ingredient list populates in that grocer’s digital cart. Chicory is already integrated with more than 70 retailers, including Albertsons, Giant Eagle, and Wakefern.

The integration requires no proprietary engineering and no national platform deal. There is no good excuse for not having it. The playbook already exists. The content network already exists. And the shopper behavior is already there.

Shoppers planning a meal add the complete ingredient list. Shoppers browsing without a meal anchor add items they remember they need. When the basket is built around a meal, every item belongs. There is nothing to prune.

This is not an emerging opportunity. It is an available one. The question is why more grocers have not taken it.

The shopper's next basket is being built right now, on a telehealth platform, at a gas pump, on a recipe site. Most won't be. Will you?

The Unit Problem is a five-part series on why US grocery unit volume is falling and what traditional grocers can actually do about it. Next we'll look at the bottom of the purchase funnel.

1. The Unit Problem

2. Your Tools Are Teaching Shoppers to Spend Less

3. The Partnerships You Haven’t Made Yet

4. You Fixed the Wrong Friction

5. $4.75, Already Applied

This is the second post in a 5-part series, The Unit Problem, on how grocers can reverse unit volume contraction. We started here.

The grocery apps and sites were supposed to drive basket size. They’re doing the exact opposite.

A shopper building a cart online has something no in-store shopper has ever had: a running total, visible at all times, and updated with every item added. The behavioral consequence is predictable. When the number gets uncomfortable, items come out. The planned items remain, but the high-margin items, the impulse additions and premium versions, would have gone in the cart without a second thought in the store.

This is basket pruning. It is a structural feature of the digital grocery channel. The shoppers who removed three items before checkout are not behaving irrationally. They’re doing exactly what the interface trained them to do.

Showing the cart total is not the problem. What the shoppers think about before they get there is.

McKinsey’s State of Grocery North America 2026 documents the channel shift driving this pattern. Purchase frequency increased 5% year over year from August 2024 to August 2025, while units per trip fell across every major grocery channel1. Shoppers are coming more often and buying less each time, splitting missions across banners, channels, and delivery platforms based on where the value lands this week.

I saw this same pattern eight years ago: omnichannel shoppers allocate 31% of their grocery spend to a given retailer, compared with 26% who only shop in-store. Cross-channel shoppers spend more. The digital channels, as most grocers have built them, are quietly eating that advantage.

The fix is not incremental.

From Convenience Tool to Mission Builder

Current grocery apps and sites are built around a search bar and a cart. Shoppers arrive with vague lists, search for items, and add them as they watch the total grow. Then they remove anything that pushes past a mental threshold. The channels are neutral containers, and neutral containers get pruned.

Mission builders change the entry point. Instead of a blank search experience, shoppers arrive at intent-driven entry points: “Tonight’s Dinner,” “Stock the Pantry,” “Reorder Essentials,” and “Weekly Value Grab.” Each path feeds a curated or predicted cart of products built for that specific trip type, with items pre-selected from past purchases and the current promotions.

Shoppers who enter through “Tonight’s Dinner” will not prune a basket. They’re choosing between two meal configurations that are already complete. The usual behavior has nowhere to operate because the carts were assembled before they touched them.

Kroger and some grocers are already seeing more than 20% of online orders come from recommended or preassembled carts, according to McKinsey. The behavior exists. Most grocers are not building toward it deliberately.

The Smart Basket Push

The highest-leverage version of this is a pre-built weekly basket delivered to the shopper before the shop begins.

The mechanic: every Thursday, the app pushes a household-specific replenishment basket built from purchase cadence, average pack size, time since last purchase, recommended recipes, and the week’s personalized offers. The shopper opens it, reviews, adds or removes items, and confirms. The cart starts full, not empty.

The behavioral difference matters. A shopper editing a complete cart removes less than one building from zero. Defaults are sticky. Items that go in before the shopper touches the experience tend to stay in.

Reframing the Checkout Moment

Before the cart total appears, reinforce the savings summary first: loyalty credits applied, bundle discounts captured, and private label savings versus national brand equivalents.

The sequence matters. Shoppers who see “You saved $8.40 this week” before they see the cart total are in a different mindset than those who see $94 and start pruning. Right now, most digital tools present the number that triggers anxiety before the number that provides relief.

Pair this with a basket completion prompt. When a digital cart reflects a recognizable build but is missing something, surface one targeted offer to fill the gap. Not three. One, specific to what is actually missing.

The grocer that knows a household always buys chicken thighs and salad dressing together. When there's only chicken in the cart, surface the salad dressing.

The Physical Store as Digital Activation Engine

The same shoppers who pruned their baskets online are standing in your store with full purchase intent and no running total in their heads. Digital is invisible at this moment when shoppers are most ready to buy.

End caps, fresh perimeter displays, and seasonal fixtures carry enormous foot traffic but don’t deliver digital conversions. This is free media inside a physical space the grocer already owns, and it is almost entirely unused. Adding a QR code to a display that links to a curated basket costs almost nothing to produce. It converts a physical browse into a digital cart start.

Shoppers who scan tend to complete the cart. The ones who do not scan buy what they remember and leave.

Consumers who are buying fewer items, trimming non-essentials, and cross-shopping three banners will not respond to a better search bar. They will respond to an experience that makes value visible before their anxiety sets in. Give them a complete cart to edit, not a blank slate to fill. And make sure it shows up in their Thursday planning window.

That is the difference between digital tools that prune and those that deliver.

The Unit Problem is a five-part series on why US grocery unit volume is falling and what traditional grocers can actually do about it. Next, we’ll cover how partnerships can move unit volume externally.

1. The Unit Problem

2. Your Tools Are Teaching Shoppers to Spend Less

3. The Partnerships You Haven’t Made Yet

4. You Fixed the Wrong Friction

5. $4.75, Already Applied

This is the first post in a 5-part series, The Unit Problem, on how grocers can reverse unit volume contraction.

There is a metric that will never appear on a grocer’s financial report: unit sales. That absence is not an oversight.

When unit volume contracts, grocers’ thin 1% to 3% profit margins get squeezed because fixed costs are spread across fewer items moving through checkout. So the reporting machinery moves to protect the narrative. Identical sales growth, propped up by inflation, leads the earnings call. Basket depth gets bundled into "average ticket size." Raw unit losses get buried in qualitative management discussion. The financials look stable while the volume story quietly deteriorates.

The Numbers Grocers Aren't Showing

Bain & Company went looking for that volume story in their analysis of NielsenIQ’s data1. Unit growth went negative in mid-2025. Then in February 2026 it stepped down sharply enough to pull actual dollar sales lower across the US. By June 2026, units were down 1.8% year over year. Twelve months earlier, they were up 0.1%. That is a nearly two-point deterioration in a single year, and it happened while prices were still climbing 2% to 3%.

Inflation did not cause the unit decline. Inflation was supposed to mask it. It stopped working.

How Shoppers Are Responding

Bain’s analysis points to overlapping pressures rather than a single shock. SNAP benefits were scaled back in late 2025, tighter eligibility rules followed in early 2026, and gas prices climbed over 20% in March. Consumers were already absorbing a 33% cumulative rise in grocery prices since 2019.

Among those trimming their bills, 56% are trading down to lower-priced brands, 49% are buying fewer items, and 44% are leaning harder on coupons. NielsenIQ panel data shows 22% of shoppers are now actively cross-shopping multiple banners to chase the best price on whatever matters most this week.

Value players, from discounters to mass and club, are taking trips from traditional grocers. Gaining shoppers does not resolve the unit problem. People are still buying less overall. The consumer who walks into your store instead of a competitor’s store is still buying fewer items than she was a year ago.

This is a contraction. Gaining share during a contraction is still winning, but it requires a completely different set of moves than the ones that worked during inflation.

The Trip Has Changed

McKinsey’s State of Grocery North America 2026 report surfaces the structural shift underneath the numbers2. Growth is now being driven more by purchase frequency, which increased 5% year over year from August 2024 to August 2025, than by basket expansion. Shoppers are not planning one large weekly mission and executing it fully. They are splitting trips across value stock-ups, fresh and prepared food occasions, convenience-led delivery runs, and targeted coupon sweeps.

The traditional large-format grocer was built for the opposite behavior: big baskets, infrequent trips, and planned purchases. That model is under direct pressure from club stores on staple volume, c-stores and prepared food operators on immediate consumption, discount grocers on price perception, and delivery platforms on convenience.

What this means operationally is that grocers who are optimizing for basket size are optimizing for a shrinking shopping behavior. The grocer that learns to win on trip frequency, capturing smaller baskets more often across more missions, is playing the game consumers are actually playing right now.

What Comes Next

The edge goes to grocers who price sharply on the products shoppers notice, and use promotions, loyalty, and private brands with enough precision to build a value story shoppers can actually trust.

Trust is the operative word. The 33% cumulative price increase since 2019 did not just stretch budgets. It eroded the default assumption that a familiar banner is a safe choice.

Shoppers who used to drop $180 in a cart without much thought are running the numbers now. They know what eggs cost. They know what ground beef costs. They are watching those items as a proxy for whether the store is being straight with them.

That is the real problem underneath the unit problem. Consumers are not just buying less. The implicit contract between grocer and shopper has been tested, and a significant portion of the customer base found it wanting.

Rebuilding that contract is the work.

The Unit Problem is a five-part series on why US grocery unit volume is falling and what traditional grocers can actually do about it. Next, we’ll cover why digital channels became pruning tools, and what to build instead.

1. The Unit Problem

2. Your Tools Are Teaching Shoppers to Spend Less

3. The Partnerships You Haven’t Made Yet

4. You Fixed the Wrong Friction

5. $4.75, Already Applied

A picker walks into a Kroger, pulls up an order, heads to the dairy cooler, and the shelf holding the chocolate milk is empty. The app reroutes. The substitution goes out. The customer gets frustrated. That friction, repeated across millions of orders each week, is where grocery economics quietly bleeds out by the billions each year.

Instacart just bought a fix for that. And it’s a much bigger deal than their press release made it out to be.

Why POS data doesn’t solve this

Here’s a question that sounds like it should have an obvious answer: don’t grocery stores already know what’s on their shelves? They scan everything at checkout. Sell three cans of soup, and the system knows three fewer cans exist.

The problem is that sell-through data tells you what left the store, not what’s on the shelf right now. Inventory counts constantly drift from reality due to shrink, misplaced products, vendor mispacks, employee errors, and theft. Columbus Consulting estimates that 56% of grocery inventory records are actively inaccurate at any given moment.1 The system says 12. The shelf says zero.

“System real time” and “shelf real time” are not the same thing for most grocers. The only way to know what’s actually there is to look, and that meant manual audits on a fixed schedule.

Instacart’s acquisition of Arpalus gives it a computer vision system that turns standard smartphones into shelf-scanning hardware with better-than-95% accuracy. Point a phone at a shelf, and Arpalus maps what’s there, item by item, in real time.

Instacart’s plan is to pipe that into its 600,000-plus shopper network, turning every order fulfillment into a passive shelf audit. A continuous, physical feedback loop on what’s actually in stock.

No grocer has built this at scale. Instacart is about to.

The found rate is the number that matters

There’s one operational metric at the center of all of this, and most grocery tech investments ignore it: the found rate.

The found rate is the percentage of ordered items a shopper actually locates and delivers. Every missed item costs a substitution, a support ticket, or a lost customer. A grocer can build a beautiful personalized recommendation engine and still bleed money on every order if the found rate is soft.

Arpalus directly protects that number by giving pickers real-time intelligence on what’s available, so they don't waste time searching for product that isn’t there.

Consumer-facing AI doesn’t touch this metric. That’s the whole problem.

The number grocers don’t like to say out loud

Out-of-stocks are costing North American food retailers 5.9% of total retail sales, according to IHL Group.2 On margins that already sit between 1% and 3%, that’s not a rounding error. That’s the difference between a good quarter and a bad year. And a loyalty program with points and tiers doesn't survive one too many substitutions.

The customer damage compounds fast in digital channels. A shopper who places a delivery order expects the items they ordered. When substitutions stack up, they don’t complain; they switch. According to FMI, the average American household shops at 5.4 different grocery banners per month.3 Switching costs nothing.

The pyramid problem

If you’ve read The Grocery AI Pyramid, you know the argument. Most grocery AI budgets flow toward the top of the stack: personalized search, generative shopping assistants, chatbots, and other consumer-facing features look good in demos. But the bottom of the stack gets treated like infrastructure nobody wants to put on a roadmap: demand forecasting, inventory accuracy, shelf availability.

Consumer-facing AI is visible. You can screenshot it, show it to a board, and put it in a press release. Shelf data infrastructure is invisible until it fails, and when it fails, it gets filed under ops, not technology. So the budgets keep going to the top.

Instacart buying Arpalus is a direct shot at the bottom. A personalized recommendation engine is worth nothing if the item isn’t on the shelf when the picker goes to grab it.

The part grocers should be worried about

Instacart is building the most accurate, continuously updated map of physical shelf inventory across thousands of stores. They’ll use their own shoppers, their own technology, and even their own carts. The data belongs to Instacart.

CPG brands already pay serious premiums for shelf positioning and promotion compliance data. The grocer who controls that data has pricing power. The grocer who outsources fulfillment to a platform running its own shelf-intelligence layer has handed over that leverage, along with the customer relationship and the delivery margin. It’s a slow transfer, and most operators don’t notice until the renewal conversation.

Caper Carts extend the reach. They’re now deployed in 100 cities across 15 states, with Instacart tripling their store count in 2025. In stores where the carts are running, every pass through an aisle silently updates the shelf map. The more stores adopt, the harder it is to unwind the dependency.

What the acquisition actually signals

With the Arpalus acquisition, Instacart is building at the bottom of the pyramid, while grocers keep spending at the top. The gap doesn't close while you're busy perfecting your chatbot.

Jennifer Shawgo published a post worth reading on the FMI blog.1 It’s measured, as trade association writing tends to be. Read past the careful framing, though, and AI’s impact on grocers’ loyalty programs is rough.

Her post opens with a simple observation most operators already know: shoppers tell them when they want an exact item, when a substitute is fine, and when they’d rather skip the purchase. Her argument is that as grocers’ AI agents take over list building and purchasing, those one-time choices can become standing instructions. Always buy this coffee. Do not replace this yogurt.

The loyalty programs grocers spent years building are their agents’ starting point. The question is whether that’s enough to keep the basket from moving elsewhere.

Please note: This isn’t an argument for building a better app or racing to deploy an AI agent. It’s an argument for understanding what these agents will reveal about existing loyalty investment before they act on it.

The Habit You’ve Been Calling Loyalty

I’ve made this argument before2, but Shawgo gives it new teeth. Only about 1% of grocery shoppers are genuinely loyal to a single banner.3 The rest are loyal to convenience. The rest are loyal to convenience. The store on the way home. The app already on the phone.

Grocers built loyalty programs on that friction and called it retention. It wasn’t. Shawgo is polite about it: “A familiar app or convenient location may drive repeat trips today.” What she doesn’t say is that the grocer’s own agent will reveal how much of that repeat behavior was friction, not preference.

Frequency tells you what happened. It doesn’t tell you what the shopper will protect once decisions get delegated to AI.

The Rise of the Protected Preference

Shawgo introduces a concept she calls “protected preferences.” When a shopper hands their basket to an AI agent, some items are excluded from the optimization. The shopper names them explicitly, removing them entirely from the agent's discretion.

Protected preference isn’t just a product decision. It’s a routing decision.

Shawgo lays out the structural logic. A national brand can anchor a product choice within the grocer’s system. Without a protected preference, the grocer’s agent may substitute or optimize around it. A private label product does the opposite. Protecting the item also protects the retailer, because you can only get it in one place. A small share of the basket becomes the reason the entire basket starts at a specific store.

When nothing is protected, the agent optimizes for price, delivery speed, nutrition, or whatever parameter the shopper sets. Grocers aren’t a destination in that world. They’re a default, one algorithm update away from being optimized out.

The Tension Between Suppliers and Retailers

Suppliers and retailers want different instructions embedded in the shopper's agent, and Shawgo is the first trade voice I've seen say it plainly.

A supplier wants the shopper to say: buy this exact product wherever you can fulfill it best. That protects the brand while making the retailer interchangeable.

A retailer wants the shopper to say: start with this store and choose the best products available there. That protects the banner while leaving product selection open.

The strongest outcome protects both. When neither is protected, the grocer's agent fills the gap with its own logic: substitution rules, private label defaults, and algorithmic ranking within the assortment. That's not a loyalty relationship. It's a transaction, and one algorithm update away from losing the trip.

What Your Promotions Are Actually Doing

Shawgo raises a question that should make every trade marketing team uncomfortable: is a promotion building a protected preference, or is it just setting the conditions for replacement?

Did the shopper buy again at full price? Did the product become a named preference? Did the promotion strengthen the retailer relationship, or did the purchase move the moment another deal appeared?

Immediate return on ad spend doesn’t answer any of that. A shopper can instruct the grocer’s agent to substitute if a preferred product exceeds a certain price. The promotion didn’t build preference. It built a condition for replacement.

Promotional mechanics can get encoded directly into the agent’s standing rules. Run enough promotions, and you haven’t built loyalty. You’ve trained the agent to wait for you.

The Product Metric No One Is Tracking

The KPI that actually matters in an agent-mediated market isn’t DAU. It isn’t session length, coupon redemption rate, or monthly active users. Those metrics describe a human interacting with a UI that the agent is increasingly bypassing.

The metric is: what share of your customer’s basket would they protect from agent optimization?

Call it Protected Basket Rate. It belongs on the product roadmap, not in a campaign brief.

Start with a proxy. Pull the top 10% of customers by spend and identify items that have had no substitution history over the past 18 months. No swaps accepted, no alternatives taken, no refund requested. Those are the protected anchors. That’s the baseline Protected Basket Rate. Everything else is at risk.

It’s an approximation. Zero substitution history reflects revealed behavior, not stated preference. But until grocers have agent-level instruction data, it’s the closest signal available.

From there, ask which items those customers would explicitly name in a standing instruction to an AI. Which store visits are habitual, and which are chosen? The gap between those two numbers is the actual vulnerability.

What to Do About It

Shawgo's call to action: understand which products, attributes, and retail experiences shoppers are most likely to protect, then strengthen the value behind those choices.

Here’s what that means in practice.

Stop optimizing the app for engagement. The grocer’s own agent is now making many of those decisions before the shopper ever opens it.

Find the fifteen things a customer would fight their agent to protect. Start with the prepared food section that makes Thursday night work. Look at the private label line where the product and the banner are inseparable. Find the produce department nobody in the market matches.

Those are the protected anchors. Everything else is a commodity. Commodity gets optimized.

Build the thing people hardcode. That’s the job now.

App Zero: Get Mobile Right the First Time is a series for PMs at established companies who've been handed a mobile mandate and don't want to blow it. This is the fourth and final post, and here’s the first.

The average mobile app loses more than 75% of new users within 24 hours1. Not because the core product is bad. Because the team that built it forgot one thing: few download an app in a calm, focused, ready-to-learn state of mind.

They download it while waiting for coffee or between calls. They get the app at 11 PM because something in their feed reminded them it existed. But after a week, more than 88% of them are gone, and after a month, it’s 94%2.

That window is tiny. And most apps spend it explaining themselves.

The Motivation Illusion

The install spike is not commitment. It’s curiosity. And curiosity runs out fast.

A new user has enough patience to get through maybe three friction points before they close the app and never come back. Most teams respond to this by building a six-screen walkthrough that asks for push notification access on screen two and introduces pitches features the user can’t access now.

This is the motivation illusion. You see high install numbers and assume those people are pulling for you. They’re not. They’re skeptical, distracted, and two back-taps away from something else.

The reframe is this: onboarding is not a tutorial. It’s a conversion funnel with no second chances. Every screen you add costs the user motivation they aren’t getting back.

Two Different People Just Downloaded Your App

Most growth teams think of “new users” as a single population. They’re not.

For any app with an existing web presence, two completely different people hit the same install screen at the same time.

The first one already has an account. They’ve been on your desktop product for months. They downloaded the app because they want their data in their pocket. They do not need a tour. They need to log in and see their stuff, and every screen you show them before that moment is something they resent.

The second one found you through the App Store or an ad. They have zero context, no account, and no reason yet to trust you. They need proof of value before they’ll hand over an email address, let alone payment info.

When you build one flow for both of these users, one of them churns at twice the rate. Usually the web user, because they’re the most irritated by being treated like a stranger. Fixing this requires actual intention at the top of the funnel:

• Use deferred deep links in web sessions. If an existing web user taps “Get the app,” that link should carry a signal. Land them directly on a login screen and skip everything else. This is a growth cheat code.

• For cold installs, let users experience something before you ask for an account. Duolingo starts a lesson before you’re registered. That’s not a growth hack. It’s a choice to earn the relationship before demanding it.

The Four Decisions That Actually Determine Whether Onboarding Works

Forget the six-screen walkthrough debate. Onboarding comes down to four specific decisions. Each one creates a leak if you get it wrong.

Permissions Timing

Asking for push notification access on screen one is an amateur mistake that’s all too common. It keeps happening because permissions feel like a logistics problem. They’re a trust problem.

When a user sees “Allow Notifications?” before they’ve done anything in the app, they have no idea whether your notifications will be useful or annoying. So they tap “Don’t Allow,” and you’ve lost that channel permanently. Getting it back requires a manual trip to Settings that almost nobody makes.

The rule: earn the ask. Stack-rank your permissions by how much value the user needs to have experienced before each one makes sense. Push notifications come after the first completed action, minimum. Camera or location access gets triggered at the exact moment the feature that needs it becomes relevant, with one sentence explaining why.

Getting to the Aha Moment First

Every app has a moment when a new user first understands why the product will matter to them. For a grocery app, it might be when a previous order reappears, saving them 20 minutes of list-building. For a password manager, it’s the first time autofill works without friction.

Most onboarding flows bury that moment under setup steps and feature previews.

Map the shortest possible path from install to that specific moment. Then cut every screen that isn’t on that path. Advanced settings, social features, tutorials for things the user hasn’t tried yet, all of it can wait. The job is to get them to the point where they think, "Okay, this is actually useful," before you ask for anything else.

Account Creation

Every field in your sign-up form costs you users who decided the trade wasn’t worth it. Email plus password as the only option in 2026 is a slow bleed.

Sign in with Apple or Google removes most of the friction and gives you verified contact info as a bonus. Passkeys are catching up fast for apps where security is part of the value.

Delayed sign-up (let users experience something before requiring an account) works well for discovery apps. It falls apart for apps where personalization is the core product, because you can’t personalize anything for a ghost. Know which one you’re building before you pick a strategy.

Define What “Activated” Actually Means

“User opened app twice in the first week” is not an activation event. It is noise.

A real activation event is specific, observable, and actually predicts whether someone will still be using your app in 30 days. For a grocery app, it might be “completed checkout with at least five items.” For a ride-share, it might be “booked a second time within seven days of the first.”

To find yours, pull retention data on your day-30 users and work backwards. What did they do in the first session that churned users did not? That action is your activation event. Build the entire onboarding flow around getting new users there as fast as possible. Everything else is decoration.

The Personalization Trap

Asking users what they want during onboarding can reduce early churn. It can also make them feel like they wandered into an HR intake form when they just wanted to try the app.

One rule covers most of this:

A fitness app that asks “What’s your main goal?” and routes users to different workout tracks based on the answer? That works. A budgeting app that asks five questions about financial habits before showing any interface, then presents the same generic dashboard regardless of what you said? That destroys trust faster than no personalization at all.

Questions should be visual and fast. A card selection with big tap targets takes ten seconds and feels interactive. A form with text fields and dropdowns takes ninety seconds and feels like homework.

Two questions maximum. And only if you can point to exactly what changes on screen for each possible answer. If you can’t, cut it.

Keep Working on Onboarding After Launch

Onboarding is the highest-leverage surface in a mobile app. A 10% improvement in activation rate compounds across every new user from that day forward. Most teams touch it once at launch and don’t go back.

This is where most of the real opportunity is sitting, unworked.

The sprint structure that actually moves the number:

Week 1: Find where users are actually leaving. If you don’t have event-level tracking on every onboarding screen, that’s step zero. You need to know which specific screen has the highest drop rate, not just that churn happens in the first 24 hours. Set up a funnel in Amplitude, Mixpanel, or PostHog that shows you screen-by-screen falloff. Find the worst one. That’s your first test.

Week 2: Write a real hypothesis. Not “we think this screen is bad.” Something you can actually test: “We believe users drop at the permissions screen because we’re asking for notifications before showing value. If we move that prompt to after the first completed action, we expect drop-off here to fall by 20%.” One variable. Resist the urge to change three things at once because you’re impatient to see results.

Week 3: Read the results and move. 95% confidence before calling it. If you don’t have enough traffic to hit significance in a week, either run it longer or accept that acquisition is the bigger constraint right now. That’s a real choice, not a failure. Apply the winner, find the next worst screen, repeat.

One other tool worth using: session recordings. Pull 20 sessions from users who churned on day one and watch them. The patterns are usually obvious. Someone rage-tapping the same button repeatedly, a permission prompt appearing before any value, a flow that assumes the user knows what a feature does before they’ve used it. You can see all of this in an afternoon, and it will tell you more than any A/B test result.

The onboarding flow you shipped at launch is probably the one your users are still running through today. Most teams never go back.

That’s the whole game.

App Zero is a 4-part series for product managers at established companies launching their first consumer mobile app.

1. Don’t Make Mobile a Smaller Web

2. Getting Everyone Aligned Before You Build

3. A Mobile MVP Nobody Will Fight You On

4. 94% of Your Users Will Never Come Back

This is the third post of App Zero: Get Mobile Right the First Time is a series for PMs at established companies who've been handed a mobile mandate and don't want to blow it. We began here.

Your MVP is already bloated. Every stakeholder had one feature. Completely reasonable. Completely necessary. Completely capable of doubling your timeline.

I’ve said yes to all of them. Never got a v1. I just planned disasters with a splash screen.

Apps took nine months to build. Launch dates slipped. I’ve lost credibility with my developers. Users handed me 2.8 stars and an App Store page of one-star reviews I couldn’t delete.

Nobody was proud of them. Nobody learned anything. "Mobile" became a dirty word in every company I worked in.

So I built a structured way to push back. This is it.

The Parity Trap

The most common mistake I've made in mobile product thinking was a framing decision. Someone in the room said, "The mobile app should do everything the website does," and nobody laughed.

Web parity is a lazy default dressed up as a requirements doc.

Stakeholders get nervous when familiar features are “missing.” Engineers default to what they’ve already built. And you, without a structured defense, capitulate because it feels safer than the argument. You end up porting a desktop experience onto a 6-inch screen, wondering why engagement is terrible.

So reframe the conversation entirely. The only question that matters:

The Three-Pillar Defense

Pillar 1: Find the 2-3 Native Micro-Jobs

Mobile does a small number of things dramatically better than web: high-frequency tasks, location-dependent actions, and anything that benefits from a push notification at the right moment. That’s it.

Your job is to identify the 2 or 3 tasks your users will want to do on their phone, on the go, without sitting down to think about it. Things like order tracking, loyalty check-ins, quick reorders, and appointment reminders. These are the things that genuinely benefit from native.

Everything else waits. Use a web view, add it to the backlog, or cut entirely.

Audit every proposed feature against one question: does mobile handle this exponentially better than a browser? If the honest answer is “not really,” cut it.

Pillar 2: The No-Hedge Scope Document

Here's the pattern that kills most mobile PRDs: everything that doesn’t make the cut gets labeled “Phase 2.”

I've written more Phase 2 parking lots than I can count. None of them shipped.

Write a scope document where non-inclusion is an explicit decision with explicit conditions attached. Write it with a trigger: “User profile editing is deferred until we hit 50K MAU or onboarding completion exceeds 60%, whichever comes first.”

Now you have a contract. Stakeholders know the feature isn’t dead. They know what unlocks it. And you’ve bought yourself a launch without the feature.

This also forces you to be honest about what you’re actually prioritizing. If you can’t name the metric that would trigger a deferred feature, you don’t believe in it yourself.

Pillar 3: The Mobile Effort/Impact Matrix

Week-one engineering estimates for mobile are always wrong. Not a little wrong. Often 2x wrong.

Things that are “cheap” on web are deceptively expensive on mobile. Complex forms, legacy auth flows, and anything that touches the OS directly (camera, location, notifications) all carry hidden costs that don’t show up in your web estimates.

Before you finalize scope, sit down with your engineers and get a mobile-specific estimate for every feature. How long on iOS? How long on Android? What changes when you factor in OS permissions, offline states, and two rounds of App Store review?

The number will go up. Plan for it.

Then build a simple matrix: estimated mobile effort on one axis, expected day-one user impact on the other. Anything in the high-effort/low-impact quadrant is not a v1 feature.

Launch Sequencing over Big Bangs

The companies that do mobile well don’t launch with everything. They launch with one or two things done so well that users tell their friends about it.

Instagram launched with filters and a feed. That's it. They built the thing people actually wanted on their phone, nailed it, and used the momentum to justify every feature that came after.

You have the same option. Launch narrow. Ship quality. Then show the data to earn more scope. A 4.4-star App Store rating from a constrained, polished v1 is a much stronger argument for headcount and budget than a 2.3-star rating.

Narrowness pays off another way. Launch two features and your metrics are legible. Launch five and you're guessing.

The 7-Question Scope Review Checklist

Copy this into your PRD. Make your team answer them. Features that can’t clear the bar get deferred with explicit unlock conditions, not vague “Phase 2” labels.

1. Does this feature serve one of our 2-3 core mobile micro-jobs?

2. If omitted, does it break the critical path, or just cause minor user friction? (Minor friction is a web view. Or a support email. Not a build.)

3. Are we estimating this based on actual mobile architecture, or copy-pasting web engineering assumptions?

4. Does this feature affect more than 5% of our target day-one user base? (If not, it’s an edge case. Edge cases don’t make v1.)

5. Can this be handled by an elegant error state, a web view, or a support flow instead of native code? (Honest answer: usually yes.)

6. If this feature delays our App Store submission by two weeks, are we still glad we built it?

7. Which specific mobile metric does this feature move, and is logging ready for it? (If you can’t name the metric, you can’t justify the feature.)

Your PRD is open. Start at the top of your feature list and run every item through these seven questions. Anything that fails gets a deferred trigger, not a Phase 2 label.

You’ll cut more than you expect. That’s the point.

App Zero is a 4-part series for product managers at established companies launching their first consumer mobile app.

1. Don’t Make Mobile a Smaller Web

2. Getting Everyone Aligned Before You Build

3. A Mobile MVP Nobody Will Fight You On

4. 94% of Your Users Will Never Come Back

This is the second post for App Zero: Get Mobile Right the First Time is a series for PMs at established companies who've been handed a mobile mandate and don't want to blow it. Here’s the first post.

The C-suite announced the initiative. Everyone nodded. Someone said: “Customer-centric.” There was a slide with a rocket emoji. People agreed it’s needed.

Then the meeting ended.

A mandate is not alignment. It’s just permission to start the hard part. And the hard part is the people. Engineering doesn’t want a second codebase. Design has never thought in touch targets. Leadership has a vague picture in their head that looks like Instagram but costs nothing. Getting all three moving the same direction before you’ve written a single line of code is the actual job.

Most companies don’t fail at mobile because they picked the wrong features. They fail because they never actually got aligned in the first place. They just assumed the clapping meant something.

The Scope Fight Is Never About Scope

I’ve sat in a lot of mobile scoping meetings. They all go the same way. Engineering wants to cut. Stakeholders want to add. Design wants three more weeks. Nobody is actually listening to each other. An hour passes. Everyone leaves annoyed, and nothing is resolved.

What’s actually going on in that room has nothing to do with features. It’s fear.

Engineering’s fear is the hidden one. They’re scared of the codebase they don’t know how to staff or maintain. The senior engineer in the corner isn’t worried about your MVP. He’s worried about being on-call at 2 AM when the iOS build breaks and nobody on the team knows what they’re looking at. He says “we need more time” because he can’t say “I don’t trust this plan.”

Design’s fear is about losing craft control. They’re afraid of losing the ground they’ve spent years earning. They have a design system. It works. Mobile breaks their assumptions about layout, interaction, handoff. What they’ve built their professional reputation on suddenly doesn’t apply, and nobody is acknowledging that.

Leadership’s fear is what keeps them up. It’s the 1.8-star app store review that lives in Google results forever. They’ve watched competitors embarrass themselves publicly on mobile. A bad launch doesn’t disappear. It sits there.

You cannot argue people out of fears they haven’t admitted to. What you can do is address the actual exposure each group is carrying.

For engineering: put it in writing. Agree in writing to a staffed mobile pod, a defined on-call model, and an explicit owner. Remove the ambiguity before it becomes a standoff. For design: pull them in before you’ve wireframed anything. Let them define quality on mobile rather than defending against decisions already made. For leadership: get go/no-go criteria agreed on now, before launch pressure makes everyone stupid.

Persuading harder doesn’t work. Changing what people are actually worried about does.

How to Actually Get Aligned

Run a Mobile Discovery Sprint: Don’t call it a brainstorm. Run a time-boxed sprint, two weeks max, with eng, design, and product in the room, and a specific deliverable at the end. Answer three questions: What does the user need to do on mobile that they can’t do well on web? What are the hardest technical constraints? What’s the minimum version we could ship and learn from?

Keep pulling the conversation back to constraints, not possibilities. End with a written, signed-off list of what’s in scope, what’s explicitly out, and the rationale for each call. That document will save you later.

Write an MVP Definition Leadership Will Actually Sign: The mistake you make is writing an MVP definition engineering can execute, but leadership can't picture.

The version that holds up has three parts.

Write a user story in actual English, not PM-speak. “A customer can pay their bill from their phone without calling support, in under three minutes.” That’s it. That’s your north star.

After that, list the features required to make that story true. Just those. That's the whole list. Don’t add anything else.

And include a list of what’s explicitly not in this version, along with a rough placeholder for when it’s reconsidered. This is the part everyone skips. Don’t skip it. The exclusion list is what keeps scope from rotting.

Handle the Scope Creep Stakeholder: There is always one. The senior leader who decided their feature must be in version one, or they will escalate.

Don’t fight it. Don’t fold either.

Make the cost visible: “I want to take this seriously. Can I walk you through what we’d have to cut or delay to fit this in V1? Once you see the full picture, if you still want to prioritize it, I’ll back that call and take it to [exec sponsor] to reset the timeline.”

That's the whole move. You're not saying no. You're handing them the bill and asking them to sign it. Most of the time, they don't want to sign it.

How to Not Die in the Last Month

Web-first companies always blow up mobile launches the same way. They build for six months, pick an arbitrary date, and then spend the last two weeks discovering that nobody ever defined what “done” actually meant.

What’s the acceptable crash rate? Nobody knows. How long does App Store review take? Someone guesses three days. It’s actually a minimum of ten, and that’s if nothing gets rejected. Are analytics instrumented? Mostly. Does the rollback plan exist? There is no rollback plan.

Fix this now, before you’re in it.

Before the project starts, get written agreement on:

• Performance: Spell out P95 load time, the crash-free rate threshold, and that there’s no ANR on primary flows.

• App Store: Set submission ten-plus business days before launch. Get metadata signed off. And talk about what happens if it gets rejected.

• Analytics: Core flows need to be firing. Baseline verified in staging. And you should be able to answer one basic question: did the user finish what they came to do?

• Feature flags: Every major feature can be turned off without a re-release.

• Go/no-go: Define what's a hard blocker and what's negotiable, in writing, before launch pressure turns every opinion into a crisis.

The reason to do this up front isn't process religion. It's that everyone's judgment degrades under launch pressure. The CEO wants the date they announced. Marketing already sent the teaser. Engineering is running on nothing. In that environment, a pre-agreed checklist is the only thing left with any authority.

The Escalation Nobody Wants to Do

You did everything right. The sprint was clean. The MVP held. The milestones were real. And there is still one person blocking the whole thing for reasons that, if you’re honest with yourself, are mostly political.

Rookie PMs wait. They absorb it, absorb it, absorb it, and then eventually say something they regret. Now there’s a relationship problem sitting on top of the product problem.

The move is to go earlier, not angrier.

Write it down: the blocker, the impact, and your recommendation. Bring it to your exec sponsor not as a complaint, but as a decision that needs to be made at a higher level. “Here’s the situation, here’s what’s at risk, here are the options I see, here’s what I’d do. I need your call.”

That’s it. Clean, documented, no drama. You’re not throwing anyone under the bus. You’re doing the actual job, which is surfacing decisions to the people with authority to make them.

One more thing. The job people hired you for is building an app. The job you’re actually doing is building your company’s ability to build apps. Every time you handle one of these moments without making it a crisis, you’re teaching the org something. You’re making the next mobile project easier.

You’re the first one through the door. Make it look like you’ve done it before.

App Zero is a 4-part series for product managers at established companies launching their first consumer mobile app.

1. Don’t Make Mobile a Smaller Web

2. Getting Everyone Aligned Before You Build

3. A Mobile MVP Nobody Will Fight You On

4. 94% of Your Users Will Never Come Back

App Zero: Get Mobile Right the First Time is a series for PMs at established companies who've been handed a mobile mandate and don't want to blow it.

A former coworker hit me up for drinks a few weeks ago. Halfway through the first round, it became more like product therapy.

His team had been tasked with launching a mobile app for their division. It wasn't going well. The project was taking three times longer than expected. Their MVP that nobody could agree on felt bloated. And the early users? Opened it once, maybe twice, and disappeared.

I know what you’re thinking:

Yes, but I've seen this before. A lot.

And yet in 2026, the setup is almost always the same. Leadership decides mobile is the move. Someone pulls up the web product in a conference room, points at the screen, and says "we basically just need this, but for mobile." The room nods. A timeline gets sketched on a whiteboard. Engineering gets pulled in.

That's where it goes wrong. Every time.

, watching established companies with real products and real users launch apps that nobody opens more than once. The bar being low doesn't mean most teams are clearing it.

There’s a meeting that happens at almost every established company the moment leadership decides to launch a mobile app. Someone pulls up their web product on a laptop, points at the screen, and says something like “we basically just need this, but for mobile.” The room nods. A timeline gets sketched. Engineering gets looped in.

That’s the point where most mobile ideas go wrong.

Porting your web experience to mobile is not a mobile strategy. It's a way of avoiding the harder question: what does mobile actually let you do that web doesn't?

Companies that skip that question ship apps that work fine and get used once, because the app doesn't do anything your website couldn't already handle. Users figure this out fast, even if they can't articulate it. The downloads trail off. The ratings drift down. The investment quietly underperforms, and nobody wants to be the one to say why.

Why the port-the-web instinct is so hard to kill

The reason this keeps happening is organizational gravity, not stupidity.

Your web product already exists. It has features that stakeholders fought for. Users who depend on them. An engineering team that knows exactly how they work. When mobile comes up, every stakeholder’s first move is to find their piece of the web product and make sure it’s in the app.

The PM who pushes back looks like they’re playing politics. The PM who doesn’t push back ships a mobile app that’s basically a slower, worse version of the website.

And the context problem is real. Web assumes someone with time, a keyboard, and actual attention to give. Mobile assumes the opposite. Short sessions. One thumb. Constant interruptions. Someone who opened your app while waiting in line and will bail the second you ask too much of them.

Those two contexts produce different products. Features that make your web experience powerful tend to make your mobile experience heavy. Flows that feel smooth on a 1440px screen feel punishing on a 390px one.

Designing mobile like it's just a smaller browser window is how you end up with something nobody opens twice.

The question that changes everything

Before you talk about features, before you pick a platform, before anyone writes a line of code, there's one question worth actually sitting with:

That framing cuts through a lot of noise. Mobile is good at fast, specific, in-the-moment actions. Check a balance. Track an order. Book something. Capture something before the window closes.

It’s not good at complex multi-step flows, dense information, or anything that expects the user to remember where they left off.

Go look at your web analytics segmented by device. Find where users are already hitting your product on a mobile browser. Those sessions are probably a mess. That’s actually useful information. Those are the use cases people care enough to attempt even when the experience is bad.

That’s your starting point. Nail those before you try to do anything else.

What your web product tells you about your mobile opportunity

Your existing product is full of mobile signals if you know how to read them. Short mobile sessions with high completion rates? Build those flows. Long sessions with high drop-off? Leave them on web.

Look at your retention numbers by device. A gap between mobile and desktop retention is the clearest possible signal that there’s a real job to do. Parity means wait. Mixed signals usually mean someone in leadership wants this more than the users do.

The output of that audit isn't a feature list. It's a mobile thesis. One paragraph. What specific thing will your app do better than any other option, including your own website? If you can't write that paragraph, you're not ready to build.

The stakeholder conversation you need to have early

The conversation you need to have with leadership early is commercial, not philosophical. Every feature you add to v1 is something that ships later, gets tested less, and burns runway you could spend proving the core works.

A narrow app that does one thing well and keeps people coming back is worth more than a comprehensive app that nobody opens twice. And it’s a much better foundation for everything you’ll build next.

My coworker’s team had plenty of alignment. Everyone agreed on the scope, the timeline, and the launch date. What they didn’t have was a mobile thesis. That’s the difference between alignment and just having a lot of meetings.

Write the paragraph first.

App Zero is a 4-part series for product managers at established companies launching their first consumer mobile app.

1. Don’t Make Mobile a Smaller Web

2. Getting Everyone Aligned Before You Build

3. A Mobile MVP Nobody Will Fight You On

4. 94% of Your Users Will Never Come Back

Grocers have heard this pitch before. In 2020, Instacart called it a lifeline. It was a leash. Every few years, Silicon Valley finds a new way to tell grocers that giving away their customer relationships is a gift.

This time it’s called the Universal Commerce Protocol.

Google launched UCP at NRF in January1 with a guest list built to make you feel safe: Walmart, Target, Shopify, plus Visa, Mastercard, and Amex. The pitch is seductive. Shoppers talk to Gemini, build a basket in conversation, and check out without leaving the chat. You keep your data. You keep your payments. You stay the “Merchant of Record.”

Read that last part again, because it’s the whole trick.

Google lets grocers keep the title while it takes their job. Merchant of Record means you handle the transaction. It says nothing about whether the customer ever sees your name, opens your app, or forms a single thought about your brand. You become the warehouse. Gemini becomes the storefront.

Google is selling five promises. Let's see who actually pays for each one:

1. The Universal Cart Kills the Part of the Basket That Pays Your Rent

A grocery store isn’t a vending machine. It’s a psychological environment engineered to push the customer into buying more than they came for. It’s the endcap, the impulse rack, and the private label set at eye level next to the national brand at a 30% better margin. None of that is an accident. It’s the profit model.

Hand basket construction to Gemini. The customer says, "Add milk," and the AI adds milk. Not your store brand, where you make money. Whatever the model picks, optimized for the customer and the platform, never for you. Private label penetration is one of the few margin levers a mid-tier grocer fully controls, and an AI basket-builder doesn’t know your store brand exists.

You don’t get a smarter cart. You get a smaller, lower-margin one, and you get to fulfill it.

2. “Smart Substitutes” Is You Mailing Google Your Weak Spots

This is the one that should keep you up at night.

To make substitutions work, you feed real-time inventory into Google Merchant Center. Sounds operational. It’s the most dangerous line item in the whole protocol.

Every substitution is a confession. It says a primary item was out of stock. Captured once, it’s noise. Captured across thousands of orders, it becomes a map of exactly where your inventory accuracy fails, which categories chronically stock out, and which supply chain relationships are quietly broken. That’s not marketing data. That’s operational intelligence, the kind that drives replenishment and forecasting and hits your margin directly.

We have already watched this movie. Instacart built the Consumer Insights Portal on data from nearly 100,000 stores2, then launched Data Hub at CES3, a clean room reselling first-party purchase signals to the CPG brands you negotiate against every year. Your weak spots are refined by their AI and sold to the other side of your own negotiating table.

UCP asks for the same confession, except now it’s a live, structured feed piped straight into the largest AI company on earth. A company whose stated job is finding “the best value for the consumer,” which is corporate for routing your shopper to whoever has the item cheaper.

You’re not enabling smart substitutes. You’re handing Google the schematic to your vulnerabilities and trusting it never to use them. Read the protocol docs. The word “consumer” shows up a lot more than the word “you.”

3. Local Delivery Means You Do the Work, Google Gets the Loyalty

Plug your fulfillment directly into Google’s ecosystem and compete with the big chains. Sounds great. Look at who pays for what.

You run the stores, staff the pickers, and carry the costs of the cold chain and failed-delivery refunds. Google contributes an API. When it works, the customer thinks “Gemini got me my groceries,” not “my grocer nailed that delivery.” You did the work. Google banked the loyalty.

I call this the ALDI special: pay a platform to run your storefront, then pay again when it sells your secrets back to your suppliers.

4. “Merchant of Record” Is a Legal Fiction, Not a Brand

Google’s cleverest line is that you keep full ownership of customer data, transaction history, and payments.

Fine. Who cares?

Loyalty in grocery isn’t built in a payment ledger. It’s built in the thousand small moments of the trip, the trust that the produce is fresh, the habit of the weekly run, and the feeling that this is my store. Replace all of that with a Gemini conversation, and what do you own? A row in a database confirming a faceless basket got paid for.

That’s not a customer relationship. It’s a receipt. Google keeps the part that compounds. You keep the part that gets audited.

5. Micro-Promotions Are a Race to the Bottom With a Friendly UI

The brisket-bundle demo is slick. But the chat window is a frictionless comparison engine. The instant your promotion appears, the AI can place it next to every competitor’s offer on the same items, in the same breath, with zero switching cost. Google already pilots exactly this, called Direct Offers4.

You’re not running a clever promotion. You’re in a live auction against every grocer in the market, refereed by the company that profits no matter who wins.

In a 1-3% margin business, “real-time discount optimization” is a polite name for the thing that kills you.

The Pattern Is Always the Same

Strip the jargon, and UCP is the oldest platform move there is. A company that owns distribution offers to “help” a margin-thin industry reach customers, in exchange for sitting between that industry and its customers forever.

The help is real. So is the dependency.

Walmart and Target can afford this game. A 200-store regional grocer isn’t negotiating with Google. It’s accepting an SDK.

AI is a capability, not a strategy, and UCP is a capability too. The question is whose strategy it serves. Google’s is clear: become the front door to all commerce and rent that doorway back to everyone who used to own it.

Yours should be the opposite. Own the interface. Own the relationship. Own the data that trains the AI that decides who wins this industry over the next decade.

That means the unglamorous work. Build an app your customers actually open. Keep your first-party data. Point your AI at the back room before you point it at the storefront. None of it demos as well as the basket building itself in real time. But all of it is yours.

Google built you a beautiful cage and labeled it a doorway. The grocers who walk in keep their title and lose their business. The ones who stay out keep both.

You can be the Merchant of Record. Or you can be the merchant.

Maryland’s surveillance pricing ban is a gift to grocery operators. Most of them just don’t know it yet.

The law itself isn’t particularly strong. It only covers retailers over 15,000 square feet. It bans price increases based on personal data but says nothing about decreases, which means a retailer could raise baseline prices, selectively discount for targeted shoppers, and stay technically compliant. Loyalty programs are fully exempt. Fines top out at $25,000.

Consumer advocates are already calling it gutted. But it forces a conversation the industry has been avoiding.

Grocery operators have been pouring money into AI that figures out what you’ll pay. Personalization engines. Willingness-to-pay models. Dynamic pricing algorithms that watch your behavior and adjust accordingly. The pitch was compelling: capture more margin from every transaction.

That’s not what AI is for in grocery. That’s not even close.

And now you’re paying a second price for it. Not just legally. Reputationally. Shoppers don’t trust AI in their grocery cart. They’re right not to.

The industry trained them to assume the algorithm is working against them. After watching airlines do it for years, they’re not starting from trust. You built the suspicion. You own it.

📉 Grocery runs on 1-3% net margins1. Shrink alone eats another 2-3% of sales2. Out-of-stocks cost the average grocer 4-8% of potential revenue3. Labor misallocation is harder to measure, which is exactly why nobody fixes it.

The battle isn’t about squeezing extra cents from a customer who might pay more for orange juice. It’s about not throwing that orange juice away. Having it on the shelf. Scheduling the right people at the right time.

Those are operational problems quietly killing margins every day, while the industry chases AI that looks impressive on a vendor deck.

AI built for demand forecasting, waste reduction, and labor scheduling doesn’t make headlines. But it compounds. A better forecast means less waste, fewer stockouts, and tighter labor deployment. That’s the model that actually moves the margin needle.

🎁 The gift is this: Maryland gives operators political cover to stop defending pricing models they shouldn’t have built, and redirect that budget toward AI that works on the 86% of the business that happens before anyone opens an app.

The retailers who figure this out will show up in 2027 with tighter margins, less waste, and a cost structure their competitors can’t explain.

That’s the real AI advantage in grocery. Not what you charge. What you stop losing.

AI Reality Check for Grocers is a 10-part series. This is the final post and here’s the first.

Every grocery executive right now is getting asked the same question by their board. “What’s our AI strategy?” And most of them are walking into that meeting with a deck full of pilots. A shopping assistant here. A personalization engine there. A vendor relationship with someone promising transformation. They’re calling that a strategy.

It isn’t. It’s a shopping list.

The numbers are already telling the story. MIT found that 95% of enterprise generative AI initiatives deliver no measurable P&L impact1. The share of companies walking away from AI initiatives more than doubled last year2.

Grocery is not going to be the exception. It's going to be the cautionary tale because operators have spent the last two years confusing capability with strategy.

🧭The Distinction That Matters

A strategy is a decision. It’s a choice about who you serve, how you’re different, and what you refuse to do. A capability is a tool that helps you execute that decision.

Capabilities serve strategies. Strategies do not emerge from capabilities.

The industry has this backwards. Most grocers today can describe their AI roadmap in exhausting detail. Ask them to describe their strategy, the actual business strategy that the AI is supposed to serve, and the answer gets vague fast. “Omnichannel excellence.” “Customer-centricity.” “Digital transformation.” Those aren’t strategies. They’re slogans that a strategy might produce.

Grocers have been treating AI as the answer. The real question nobody is asking is, an answer to what?

🔍How the Confusion Shows Up

Once you see the pattern, it’s everywhere.

Vendors pitch platforms as a strategy. “Our AI platform is your grocery strategy.” That sentence should make every operator in the room uncomfortable. The vendor is not your strategist. They’re selling a capability, and they want you to believe that buying their capability is equivalent to making a strategic decision.

It isn't. It's the opposite. It's outsourcing the decision you were hired to make.

Boards count pilots instead of outcomes. “How many AI initiatives do we have in flight?” is the wrong question. The right question is, “Which of our strategic priorities is this pilot serving, and how will we know if it worked?” Most grocers cannot answer the second question for most of their pilots. That’s the tell.

ALDI hands its digital storefront to Instacart and calls it innovation. It isn't, it's a surrender. The AI that will define grocery competition over the next decade runs on customer data, and ALDI just gave theirs away. They're renting their digital future from a competitor who hasn't been introduced as one yet. ALDI isn't adopting a capability. They're giving away a business.

Kroger announces AI partnerships and calls them a strategy. They aren't. They're procurement. A real strategy would tell you what Kroger said no to. Their press release never does.

What's missing from this pattern? Ask any grocery analyst to name the grocer doing AI right, and the answer is a long pause, then a hedge. "Walmart is doing interesting things." "H-E-B is probably ahead of where the press covers them." That isn't an endorsement. That's the industry admitting it can't find a winner yet.

The winners won’t be the grocers with the most pilots. They’ll be the grocers with the clearest strategy, and those grocers are still quiet.

The Reframe Operators Need

The right answer to "what's our AI strategy?" is: we don't have one. We have a business strategy. AI serves it.

That answer does three things. It puts strategy back above capability in the hierarchy. It forces the next conversation to be about the actual business, not the tools. And it signals that the executive answering won’t be stampeded by noise.

The same frame works in vendor meetings. When a vendor tells you they're selling an AI strategy, ask whose. If they can't name your priorities better than you can, you're buying a capability. Price it accordingly.

The grocer operating from a strategy first looks different. They invest where their strategy points, not where the vendor is selling. They say no to pilots that don’t tie to a strategic outcome. They buy capabilities, not transformations.

And they measure AI by whether it made the business better at being the business they chose to be. Not by whether it shipped on time.

What This Series Was Actually About

Nine posts of this series dismantled pieces of the confusion.

• The pyramid argued that most grocers invest in AI at the wrong layer.

• The forecasting post argued that inventory accuracy is the real scoreboard.

• The data-robbery post argued that grocers are handing their strategic asset to platforms that don’t share their interests.

• The labor flip argued that the real opportunity lies in the back of the house, not the customer-facing side.

Every one of those arguments was really the same argument. Grocers are making capability decisions without a strategy to evaluate them against. And when you do that, you buy whatever the loudest vendor is selling, and you end up in the 95% that delivers nothing.

AI is a capability. Strategy is still your job.

🛠️How to Stay Out of the Trap

Here are five things a grocery executive can do this quarter to avoid buying capability at strategy prices.

Write your strategy down on one page. Not a slogan. A decision about who you serve, how you’re different, and what you refuse to do. If your team cannot produce this page in a week, you don’t have a strategy problem with AI. You have a strategy problem with everything, and AI will be the most expensive place it shows up.

Audit every active AI pilot against that page. Ask two questions per pilot. Which strategic priority does it serve? How will you know if it worked? Fuzzy answer on either one, kill the pilot. Most grocers will find that a third to half of their current AI work doesn't survive this test. That's not failure. That's money you get back.

Change the vendor conversation. Stop asking what the platform does. Start asking which of your priorities it actually serves. If they can't name your priorities better than you can, you're buying a capability. Price it accordingly.

Change the board conversation. Replace the AI pilot count with the outcomes those pilots are supposed to produce. Activity metrics tell you nothing. Pick two strategic outcomes your AI should move. Report on those.

Invest where the pyramid says to invest, not where the vendor says to invest. If execution is your edge, your AI budget belongs at the bottom. Forecasting, inventory, labor, shrink. A vendor pushing a customer-facing pilot that doesn't tie to one of those is serving their roadmap, not yours. Say no.

The grocers who win the next ten years will be the ones who kept strategy as a verb. Everyone else is buying tools for a job they haven’t defined.

AI Reality Check for Grocers is a 10-part series that cuts through the hype and focuses on where AI actually creates value, from the shelf to the supply chain.

1. AI Won’t Save Grocery

2. AI Is Solving the Wrong Half of the Grocery Business

3. The Grocery AI Pyramid

4. Your Inventory System Is Lying to You

5. AI Has One Job in Grocery. This Is It.

6. Grocers Are Getting Robbed. Just Not in their Aisles.

7. Part 1: The Pyramid Flip.

8. Part 2: The Labor Flip.

9. The Grocery AI Mistake That’s About to Happen

10. AI Is a Capability, Not a Strategy