The current reality is that less than one percent of online purchases involve an AI agent.² And, AI-referred traffic converts worse than most established channels.³ OpenAI hasn’t abandoned agent commerce, but it has shifted from frictionless buying to verified merchant protocols. Even the company building the most ambitious version of this new commerce ecosystem acknowledged it wasn’t ready for seamless checkout.⁴ Most AI-assisted purchases still end up on the brand’s own website anyway.

What I find fascinating (read: sad) is that the conversation about what to do about the biggest brand disintermediation of all time is being shaped almost entirely by people who don’t run brands. Instead, the discourse is shaped by tech companies, consultancies, platform companies, agencies, each with something to sell. The result is a set of solutions designed for the people selling them, not necessarily for the people buying them (or their customers). Ya’ll we’re choosing SAAS tools and fear over ownership and building.

Subscribe now

Immune Response

When a system gets threatened, sometimes it produces responses shaped by its own constraints, not by the actual problem it’s facing. A solution that takes multiple quarters to produce returns is, for the most part, structurally unsellable in this market. So the market doesn’t produce it. It produces what it can sell: tactical responses wrapped in enough urgency to justify the spend.

The average CMO at a major US advertiser lasts 3.1 years.⁵ CEOs have maybe five before the board wants a different direction. The board reports quarterly. Nobody in the chain is rewarded for patience. The rational move at every level is to invest in whatever is visible, reportable, and defensible in a presentation. “No one ever got fired for hiring IBM” and all that. This isn’t a failure of any single role, it’s what the system selects for. And the system right now is selecting GEO agencies, a new wave of data transformation projects, and AI shopping assistants. Those are the things, like acquisition growth numbers before them, that fit inside a budget cycle and produce enough early signal to look like progress. These are what we refer to as “vanity metrics.”

Some of this work is genuinely useful. Structured data, clean product feeds, content that answers real questions? Yes, please. The problem isn’t the tools. It’s the frame. The moment when business as usual housekeeping gets packaged as strategy — with an acronym, a price tag, and a conference slot — it starts competing for the budget and attention that should be going to the actual structural strategic work underneath. Funding for delivery, owned channels, product investment, and the things that give someone a reason to seek you out by name when an agent would just as happily substitute you are not the focus. That work loses the budget competition every time, because it can’t be sold in a quarter.

What Could Go Wrong?

Ask Ballantine’s. It’s a mass-market Scotch. On purpose. Decades of positioning as accessible, affordable, everyday. Recently, the head of digital at owner Pernod Ricard ran an audit after learning that two-thirds of Gen Z were using LLMs to research products. What he found dismayed him: the data on his brands was incomplete or incorrect across the board.⁶ Pernod Ricard also owns Chivas Regal and The Macallan, and when they audited how major language models represented their portfolio, they found that Ballantine’s was miscategorized as a prestige product.⁷ The system couldn’t tell the affordable whisky from its expensive corporate siblings. That sounds like identity theft by algorithm. Decades of careful market positioning undone because a model conflated brands in the same portfolio.

Making yourself legible to these systems is necessary. Obviously. But it isn’t strategy. Being findable by a machine is not the same as being chosen by a human. When you do the tactical layer first and the structural, value layer second, you get a brand that’s perfectly visible to machines and doesn’t have what it needs to build awareness with people post-purchase, precisely when brand leverage matters most. When you do it the other way around, the tactical work becomes a distribution question, not a survival question. The order matters.

The Amex Delta card became one of the most successful cobrands in history because both brands decided the relationship was worth more than either side’s margin on any given transaction. That arithmetic requires believing the relationship compounds. The credit card industry believed it because they could measure it. Most consumer brands have never been forced to develop that discipline — but the ones building it now, quietly, while everyone else chases readiness scores, are the ones who’ll own the ground by the time the rest arrive.

The Real Cost

None of this is easy to say with a straight face unless I’m honest about what the structural work actually requires. Specifically, money allocated against returns you can’t measure for quarters. It requires data infrastructure that most mid-market companies are still building. Most of all, it requires conviction — someone in the room willing to defend investment when the buzz is on something else entirely. And, it requires doing two things at once: doing the tactical work, because you have to stay visible in systems you don’t control, while simultaneously building the thing that makes visibility possible and worth having.

Most companies already have underutilized structural assets like customer data, direct channels, product reputation, community knowledge. These opportunities are chronically under-resourced relative to the acquisition side, under-measured, and under-leveraged. Before the next GEO contract gets signed, the more useful question is whether the structural investments that already exist are working as hard as they need to for the coming reality and what it would take to make them compound.

The new leverage is not in whatever is getting shouted at you on LinkedIn, it is in asking a different question first: what do we already own that can help us in this leverage shift to post-purchase? What systems are going to enable? If you run a company, creating the conditions for that question to get asked — and acted on — is probably more valuable than anything on offer at the next conference.

Footnotes

1. McKinsey, agentic commerce market projection, 2026.

2. BrightEdge, “AI Shopping Skyrockets,” November 2025. eMarketer, December 2025.

3. Kaiser & Schulze, 973 ecommerce sites (Aug 2024–July 2025). Contentsquare, 2026 Digital Experience Benchmark.

4. Stripe, “Developing an Open Standard for Agentic Commerce,” September 2025.

5. Spencer Stuart CMO Tenure Study, 2026. S&P 500 average: 4.1 years; Ad Age top 100 advertisers: 3.1 years.

6. Gokcen Karaca, head of digital and design at Pernod Ricard, partnered with Jellyfish to audit LLM brand representation. Created “Share of Model” metric.

7. HBR, March 2026. Acar & Schweidel, “Preparing Your Brand for Agentic AI.” Pernod Ricard / Ballantine’s miscategorisation by major LLMs.

Celebrity endorsements still work, but not the way they did when celebrity was scarce. Infomercials still exist, but not as a novel format commanding premium attention. Influencer marketing still reaches people, but “real person, honest opinion” is no longer a genuinely novel proposition. As each form democratized — more people could do it, more brands could buy it — the signal diluted because of its own success. The dynamic hasn’t changed from Suzanne Somers selling ThighMasters on late-night television to a twenty-three-year-old selling protein powder on TikTok. What changed is how many people can do it and the transparency around it, which is exactly what compresses the premium.

One of Us Is the Bougie Sister

My sister lives in suburban Virginia and thinks TripAdvisor is a reputable source. It was, once. Reviews from real customers, emergent ranking. It was a good guide. Then came review fraud, pay-to-play incentives, algorithm manipulation. I noticed and left. She hasn’t noticed and hasn’t left. We are both right about TripAdvisor in our own way (or so I tell her).

Trust erodes in layers, not all at once. The people who notice leave first. The people who haven’t noticed are still there. The same signal working and failing simultaneously, depending on who’s reading it. Sociologists call this stratification. Brands should call it a planning problem.

Only thirteen percent of consumers now turn to influencers for purchasing decisions.¹ Nearly half say most influencers are fake — while eighty-eight percent say authenticity matters to them.² This is Goodhart’s Law in action: when a measure becomes a target, it ceases to be a good measure.³ “Authentic recommendation” became the target. Every brand chased it. The authenticity evaporated.

Mega-influencers became indistinguishable from ads, so some trust migrated to micro-influencers, who felt more like peers. That layer is now compressing. The trust didn’t die. It just moved down a floor.

The influencer model isn’t going away completely. There are still demographics you reach effectively through it, the same way there are still demographics you reach on Facebook. The question is whether you know where your customers actually are versus what platforms tell you. These are frequently not the same.

Borrowed Proof

Every brand knows that you need some social proof in the path to purchase. So all DTC checkout pages boast the same visual syntax: “As seen in Forbes,” “Featured in Vogue,” trust badges stacked like medals. This is borrowed trust that is manufactured. Nobody’s reputation is at stake so it’s pretty meaningless. The cost of verification sits with the customer, not the brand. And when every brand is saying the same thing, “trust” dissolves into decoration. Mayer, Davis, and Schoorman studied what people are actually evaluating when they decide to trust. Competence, can you do what you claim? Benevolence, do you actually want good things for me? Integrity, are you consistent in your principles?⁴ The badges don’t give the buyer anything real to evaluate on any of those dimensions.

Borrowed trust that is earned looks different. In 2010, Mike Kurtz was selling chili-infused honey at Paulie Gee’s in Greenpoint, Brooklyn. Paulie Gee put it on his pizza. The product spread through the chef community — cooks pulling it into their own kitchens, staking their own plates on it. That’s not a badge. That’s the whole game. Four years in restaurants and on menus before the first grocery placement.⁵ Today: 30,000 retail locations. Forty million dollars a year. One product. Competence, benevolence, integrity not just claimed on a label but tested nightly, by someone else’s customers, on someone else’s reputation.

The Edit

Trader Joe’s carries about four thousand products. A conventional supermarket carries thirty to fifty thousand.⁶ Eighty percent of what Trader Joe’s sells is their own label. No sales. No coupons. The price on the shelf is the price. Every one of those decisions is trust built into the transaction, not into the marketing. The company stakes its reputation on every single item; nothing gets to hide. The result: $2,200 in sales per square foot, nearly four times Walmart.⁷ No traditional advertising. The brand doesn’t have trust. Trust is what the brand is.

Zucker drew a distinction that’s important between interpersonal trust, that is confidence in a specific person, and institutional trust which is confidence in systems, processes, and policies.⁸ Mike’s Hot Honey started as interpersonal trust. One chef, one pizza, one jar. Trader Joe’s is what happens when you build that into a company’s offering. The curation is the product. The edit is the trust.

Institutional trust is what lets the unknown survive. When you pull a novel off the staff picks shelf at your favorite bookstore because the person who picked works at the store you love and put their name on the card so you can chat about it at checkout. The institution does the editing. The unknown survives because someone else’s credibility said: this is worth your attention.

It is important to be clear that all this is different from a recommendation engine. A recommendation engine says: based on your past behavior, here are things you might like. It’s backward-looking, data-driven, and optimized for conversion. It doesn’t have a point of view, it has a model. And when it’s wrong, nobody’s reputation is at stake. What humans have done for each other, traditionally, is different. Someone with conviction edits the world down for you, and you trust them enough to let them propose your consideration set. The distinction matters because a recommendation engine can be gamed. Guided discovery can only be earned.

Own the Mechanism

Amazon does $717 billion a year⁹ offering the opposite of curation: unlimited choice, algorithmic sorting, same-day delivery. The infinite shelf wins by every conventional metric. But when you don’t know what you want yet, the algorithm doesn’t have a point of view. It has a grid.

Brands that built something durable often move the burden of proof onto themselves. They made decisions that cost them something: fewer SKUs, four years without a grocery placement, a return policy that absorbs the risk. Those aren’t marketing choices. They’re structural ones. And they compound over time rather than diminish, which is exactly why they work when the borrowed version stops.

You can keep renting trust from platforms and influencers and badges. It still works, for now, for some customers. But the premium is moving. And you don’t own any of what you’ve rented.

¹ Oracle and CRM Essentials, “37% of Consumers Trust Social Media Influencers Over Brands,” May 2022.

² Harvard Business Review, “How to Do Influencer Marketing That Customers Actually Trust,” December 2025.

³ Goodhart, C., “Monetary Relationships: A View from Threadneedle Street,” 1975. Popularly paraphrased as: “When a measure becomes a target, it ceases to be a good measure.”

⁴ Mayer, R.C., Davis, J.H. & Schoorman, F.D., “An Integrative Model of Organizational Trust,” Academy of Management Review, 20(3), 1995. Original terms are “ability,” “benevolence,” and “integrity”; paraphrased here for readability.

⁵ Kurtz began selling Mike’s Hot Honey at Paulie Gee’s in Greenpoint, Brooklyn in 2010, inspired by chili-infused honey he’d tasted on pizza in São Paulo. CNBC, “How Mike’s Hot Honey Built a $40 Million a Year Business with a Single Product,” November 2023. First Whole Foods placement 2014; now in 30,000+ retail locations.

⁶ Trader Joe’s carries approximately 4,000 SKUs versus 30,000–50,000 at a conventional supermarket. Private label accounts for approximately 80% of its product mix versus an industry average of 15–20%.

⁷ Trader Joe’s sales per square foot estimated at $1,750–$2,130, versus approximately $400–$500 for Walmart. Trader Joe’s does no traditional advertising.

⁸ Zucker, L.G., “Production of Trust: Institutional Sources of Economic Structure,” Research in Organizational Behavior, 8, 1986. ⁹ Amazon 2025 annual revenue: $717 billion. U.S. e-commerce market share approximately 37–40%.

One defense against the coming disintermediation is coalition. Not a partnership, not a collab, not co-marketing. Coalition in the older, harder sense. Entities pooling enough shared infrastructure that the coalition becomes the thing the market has to route through rather than around.

This is not a new idea. It is the oldest answer on the books for fragmented producers facing a consolidated buyer, and it has been working in other categories for a century. It is new for most of us because we’ve spent that century thinking of each other as competitors rather than as co-producers of a category, segment, lifestyle, or industry. The reality is, when we get squeezed, we all get squeezed together.

What a Coalition Actually Is

The cleanest example is American agriculture in the early twentieth century. Orange growers in California were going broke because buyers and middlemen — railroads, wholesalers, brokers — could extract whatever margin they wanted from a producer base of family farms too fragmented to negotiate back. The growers formed a marketing cooperative called the California Fruit Growers Exchange; it is now called Sunkist. Dairy farmers in Minnesota called theirs Land O’Lakes. Cranberry growers in Massachusetts created Ocean Spray. Nearly a century later, all three are still operating. We remember them as brands, but the logo on the box was a consequence of the coalition underneath, not the reason for it.

What made these coalitions hard to route around was the infrastructure they pooled. Shared quality standards that insurers and regulators could rely on. Collective bargaining power on inputs that made the unit economics work for members and made them stop working for non-members. Political representation with enough aggregated member count to actually get heard in the rooms where rules get written. Shared certification and lab capacity that absorbed the cost of being credible. Liability pooled across members, which made the coalition the natural place for a regulator, a retailer, or now an algorithm to verify a claim.

A coalition can be a member-owned cooperative, a joint venture consortium, a 501(c)(6) industry association, a nonprofit standards body, a certification scheme, or a member-owned LLC. The legal form is a downstream decision about tax, governance, and liability tradeoffs. The functional test is upstream: can the entity pool infrastructure that members cannot build alone, and is the governance neutral enough that no single member steers it. This is the base case Mancur Olson described in The Logic of Collective Action — coalitions hold together when non-membership costs more than membership, usually through selective benefits members cannot replicate alone.¹

The Trust Layer Job

In agentic commerce, this infrastructure takes on a second job. AI systems will eventually need external verification layers to consult because models cannot generate ground truth they were never trained on. The version that matters for a brand operator is narrower and more concrete: when an agent is asked a question where being wrong is expensive, it will consult a source that can absorb the liability for the answer. That source cannot be the brand itself. It has to be a neutral, funded, audited structure the agent operator can point at when something goes wrong.

This job is not hypothetical — it is exactly what Underwriters Laboratories has done for electrical products since 1894. Retailers, insurers, and regulators route through UL certification because self-attestation is problematic and UL absorbs the liability for getting it wrong. The Forest Stewardship Council plays the same role for timber, Fair Trade for commodity supply chains, and the Good Housekeeping Seal has done a consumer-facing version since 1909. Each exists because an industry needed something no single member could credibly produce, and the market built the coalition into its plumbing. What changes in an agentic world is the consulting party — not a regulator or a retailer but an algorithm executing purchases on someone’s behalf. The structural need is identical.

The liability argument needs a qualifier given how platforms have mostly skirted accountability in the last two decades. Agent-mediated commerce is a different legal animal than, say, content moderation. When an agent executes a purchase with someone’s money, it is structurally closer to a retailer than a publisher, and retailer liability has real case law underneath it.² Add the EU AI Act and the Digital Services Act, and the pressure is real. The risk is bigger for physical goods in categories where someone can get hurt, and the first movers will define a lot of the rules latecomers follow.

The Juice Is Worth the Squeeze

Most brands hate the idea of coalition because it feels like giving up sovereignty. For ninety-nine percent of brands, that sovereignty is already an illusion. The brand is already paying close to half its gross revenue to platforms for the privilege of being seen. Marketplace Pulse estimates Amazon’s true take rate, including the required advertising spend, referral, and fulfillment fees, at around fifty percent for a typical private-label seller, up from forty percent five years earlier.³ I call it a Discovery Tax; it is levied by platforms on every transaction the brand cannot route around. Oliver Williamson’s transaction cost economics anticipated this decades ago. When the cost of making each exchange through the open market gets high enough, producers either get absorbed into a hierarchy or pool into a coalition that internalizes those costs together.⁴ Coalition is not a loss of independence. It is the redirection of a tax the brand is already paying, into infrastructure the brand actually owns.

The part that should matter most to a brand operator is that the coalition is worth building even if the agentic thesis turns out to be overstated. If the only reason to create one is that an AI agent might one day consult you, it is a speculation and a bad one. But if the coalition produces collective bargaining on packaging glass, shared cold chain, pooled regulatory and compliance filings, joint legal defence, shared lab certification, pooled media buys, and political representation that actually moves legislation — then the coalition pays for itself in line items a CFO can find tomorrow. The agentic upside is the ceiling, not the floor.

A fair objection: even if agents do consult a coalition’s verification layer in 2027, the largest agent operators are building their own proprietary trust layers and will cut coalitions out. Probably. It does not change the calculation. Temporary moats that get used well compound into durable advantages — member-generated data, regulatory relationships, political capital, trust that was earned rather than bought.

The alternative is not keeping sovereignty. The actual alternative is being a permanent tenant on someone else’s trust layer.

What You Would Need to Believe

Coalition works when incentive alignment between members is tight enough that defection is materially costly. In 2026 terms, that cost has to be concrete: loss of access to the shared testing registry the major agents are trained to consult, removal from the safe-to-purchase list that absorbs liability for the agent operator, being locked out of the collectively negotiated supplier contracts that made the unit economics work in the first place. Members have to share real infrastructure and not just a badge.

You would need to believe that governance can be neutral enough that no single member steers it. The history of producer coalitions is largely a history of one member trying to own the group and the group dissolving. And you would need to believe that what the coalition produces is something members genuinely cannot produce alone at the same cost. These are the classic conditions Elinor Ostrom identified across a lifetime of fieldwork on commons governance.⁵ They have been studied for fifty years in other categories.

The work of making a brand legible to agents — schema markup, structured product feeds, agent-readable content pipelines — is necessary and insufficient. Algorithmic Legibility puts you in the table with your features and removes you from consumer view. Coalition infrastructure changes what the algorithm has to consult in the first place.

Orange growers in California figured this out with no internet, no AI, and no Digital Services Act to lean on. They had a consolidated buyer extracting margin, and the mathematics of surviving it alone did not work.

Footnotes

1. Mancur Olson, The Logic of Collective Action: Public Goods and the Theory of Groups (Harvard University Press, 1965).

2. Bolger v. Amazon.com, LLC, 53 Cal. App. 5th 431 (2020). Oberdorf v. Amazon.com Inc., 930 F.3d 136 (3d Cir. 2019). EU AI Act (Regulation (EU) 2024/1689) and the Digital Services Act (Regulation (EU) 2022/2065).

3. Marketplace Pulse, “Amazon Takes a 50% Cut of Sellers’ Revenue” (February 2023). More recent reporting puts the figure at 50–60% for typical private-label sellers.

4. Oliver Williamson, Markets and Hierarchies: Analysis and Antitrust Implications (Free Press, 1975) and The Economic Institutions of Capitalism (Free Press, 1985).

5. Elinor Ostrom, Governing the Commons: The Evolution of Institutions for Collective Action (Cambridge University Press, 1990).

At Visa, we divided spending into everyday and discretionary. Everyday is necessary stuff that barely moves when the economy shifts: groceries, utilities, gas, the commute. Discretionary is the stuff that contracts when things get tight: travel, certain kinds of dining out and entertainment, clothing. Different data patterns. Different risk profiles. Different businesses.

In research we often heard how consumers managed all this in their heads. People are running remarkably sophisticated internal systems. From a payments perspective, it meant one card for groceries because it maximizes cash back in that category. A different card for the big purchase that needs to float for a couple of months on the lowest APR. A third for travel because of the points program.

Richard Thaler won a Nobel Prize for describing the underlying mechanism: mental accounting. How we sort money into psychological buckets with different rules, different emotional registers, and a different willingness to optimize is going to be really important in this new agentic commerce age.¹

Even though money is fungible, consumers haven’t treated it as such. Historically.

The “Easy” Part

For everyday purchases — paper towels, dishwasher detergent, the stuff that runs on habit — the AI agent can be relatively straightforward. It does what consumers were already trying to do with their card-switching systems, it optimizes against a set of rules.

Recently, Google reintroduced “the invisible shelf” concept: when an AI agent shops on your behalf, it scans a surface and products no human ever sees. That “shelf” contains structured data, tagged attributes, and other machine-readable specs. If your brand isn’t built for that surface, the agent skips you entirely.²

For everyday brands, one move is obvious. Be legible. Compete on structured data. NielsenIQ data from March shows niche CPG brands gaining share because AI discovery rewards specificity. The brand that tagged “verified sustainable packaging” gets surfaced ahead of the mass brand with generic metadata.³ The invisible shelf favors that over scale. For now.

For discretionary purchases like travel, fashion, experience, or furniture; the agent is a different animal. It can find “hotels in Midtown under $300 with a gym and late checkout.” It can’t as easily understand that you don’t want to walk into a lobby and feel like the oldest person in the joint, wondering if the music has to be SO LOUD waiting to check in. Major chains like Hilton and Marriott have aggressively rolled out their own AI planning tools recently for exactly this reason: whoever owns the discovery layer owns the customer. They’d rather spend to keep that in-house than become a row in someone else’s database.⁴ That’s smart, but not a moat that’s guaranteed to last.

As usual, a binary is getting all the attention: commodity or luxury. But the terrain between everyday and discretionary is what’s super interesting. It is also where brand value actually lives.

What’s Your Pickle?

I buy all my pasta from a producer in Lari, a tiny hill town in Tuscany where I rented a house for a milestone birthday. Sure, I am absolutely fancy like that. AND, when I open my cupboard and see the yellow packaging I am transported to a hot July afternoon drinking Negronis with friends on a square. It’s dusty with semolina flour, you can see the pasta drying on racks, and view company’s bright yellow bags — the same color as the sunflowers surrounding the town — in the cool, dark store room you pass by on the way to your car.

An AI agent looking at my grocery list would see “dried pasta, 500g” and find me something excellent for way less money, probably delivered tomorrow. It would be solving a problem I don’t have. The value of this purchase is in a collection of memories I access every time I boil water.

In my family, we call this a pickle — the product you'd mourn if it disappeared, that you know more about than anyone in the room, that you ask for by name every time because settling feels like a small betrayal. Most of us have at least a few. It could be your toothpaste, your sneaker, your lipstick formulation, your choice of salt.

We all have our pickles. One of mine is bougie pasta. Others have different products or a category where you care disproportionately vs. the price point. My sister’s is pain relievers. Rachel was a very competitive athlete for many years then had a physical job as a firefighter and paramedic. She’s always had her share of aches and pains. Her coaches always gave her a particular pain reliever. Decades (and no small amount of R&D later), there is no veering from it, ever. Part drug, part talisman at this stage.

The pickle zone as I’m now calling it is everyday on the surface and discretionary underneath. The agent reads the surface. It misses everything below it.

New frameworks have to account for the fact that the pickle zone isn’t a property of the product. It’s a property of the person buying it. Paper towels are a commodity to me, Puffs Plus is a pickle. Toothpaste, socks, coffee, pens. For most people, these are autopilot. For someone else they are part of their identity. What this means is that almost no brand is categorically in any zone. Where you sit depends on who’s buying and why. Your customer base contains people in all three zones, and the ratio is your actual exposure map.

Subscribe now

The Waitlist for Beans

Rancho Gordo is one of my favorite examples. They sell hype-worthy dried beans. Generally speaking, dried beans are a commodity item. Spanish gigante beans in olive oil and preserved lemon at a specialty store is one thing but a bag of dried beans? An agent would likely optimize right past them.

Rancho Gordo is going to do just fine. It has 30,000 Bean Club members, another 30,000 on a waitlist, and it moves 2.5 million pounds a year.⁵ Near-zero churn. Founder Steve Sando deliberately limits growth. When asked why he doesn’t go mainstream, his answer: I created my own club and made it worth belonging to. I suspect Steve is not rushing to ensure his beans are legible by agents.

These beans are a lot of people’s pickle, it seems.

Pickle Zones Aren’t Equal

My niece Abby is in her twenties. For her, most commerce is about cost and ease. That’s not a failure of taste, it’s a rational response to her economics. Her everyday zone is large. Her pickle zone is small and her discretionary bucket is fairly small as well. The agent will serve her well for most purchases.

Pickle zones change. What you care about at twenty-five isn’t what you care about after additional decades of work, travel, relationships, and general adulting. And, it’s also true that these purchases, often the more expensive choice, aren’t equally available to everyone. The K-shaped economy is making it worse. There is a widening split between high-income and low-income consumer behavior. Income determines how much fancy pasta you can import.⁶ When income compresses, the purchases that once functioned as identity get rationalized into the everyday bucket. The agent accelerates this by offering the efficient substitute. Things get lost forever in those changes.

Who Are You and To Whom?

In addition to “how do I prepare for AI agents” brands should be figuring out where it actually sits in its customers’ mental architecture? This is going to be the new crucible in agentic commerce.

If you’re solidly everyday — for most of your customers, most of the time — the agent is your distribution channel. Embrace the invisible shelf. Optimize away.

If you’re discretionary, invest in the relationship. Build the direct infrastructure. Make the encounter worth having.

For almost everyone else, find out who has you in their pickle zone. That challenge is the most interesting and the least obvious. Your value will be invisible to every system being built to mediate commerce. You can’t win by optimizing for the machine, because what makes you matter isn’t machine-readable. You win by deepening the thing that makes your people care — the community, the expertise, the story, the specificity.

Martelli Pasta has been in Lari since the 1920s. Four shapes. Typically not for the tourists.

Share

Footnotes

1. Thaler, R. H., “Mental Accounting Matters,” Journal of Behavioral Decision Making, 12, 1999.

2. Google Cloud, “The Invisible Shelf: How CPGs Can Win Agentic Commerce in 2026,” January 2026.

3. NielsenIQ and Kearney, “The New Growth Frontier: Agentic Commerce and AI in CPG,” March 2026.

4. Accor (Q3 2025); Hyatt and IHG (Q4 2025); Marriott (early 2026); Hilton AI Planner (March 2026). Hospitality.today, March 2026.

5. Napa Valley Register, March 2026; Taste magazine, Steve Sando interview, 2023.

6. NIQ, “Decoding America’s Great Consumer Split: Inside the New K-Shaped Economy,” January 2026.

And a few days ago, there was a little piece of paper at the bottom of a Pika chicken potpie box. It just has the look of being typed in a farmhouse so I zoomed in. The pie was genuinely very good, and now I know they offer lobster, beef, and shepard’s as well.

The bar for brands standing out in the moment after the sale is so low it barely qualifies as a bar, honestly.

The Magic Window

If you don’t activate a new credit card customer in the first thirty days your P&L will suffer the consequences. Give them something meaningful, a real reason to change their behavior, or you’re in the sock drawer for a rainy day. The loyalty programs, the sign-up bonuses, the tiered rewards: all of it existed because of one hard insight: the relationship doesn’t happen by itself even, often, with a good product. You have to design for it. And, once the window closed, it’s rarely reopened.

There’s a behavioral reason the window matters precisely when it does. Kahneman’s peak-end rule: we don’t remember experiences as averages. We remember peaks and endings.¹ The moment you open something you ordered is not incidental. Done right, it is the ending of the purchase experience: the last strong impression the brand gets to make before the memory sets. The Sézane catalog wasn’t in the box by accident. The Pika insert wasn’t accidental either, even if it looked like it was made by someone’s low-tech grandma. Both were designed to be the things to engage you in a way more meaningful than the thumbs up of a doomscroll.

People have been designing for the post-purchase moment since S&H Green Stamps turned every repeat purchase into a future aspiration. The insight that a customer relationship needs tending is not new. What was new was the data. For the first time, banks could see precisely what a customer was worth at year five and build the activation infrastructure to match at month one. I had no idea how good I had it.

Sorry, Who?

Performance marketing made acquisition feel controllable. You could see the click, attribute the cart, measure the conversion. What happened after was harder to count and easier to defund.² What depleted wasn’t the touchpoints. It was the mental availability and equity — the things that makes someone choose you without a discount code, come back without being re-acquired, feel something when they see the name.

We tried to solve it with a mechanism. Subscription was the answer for a decade: recurring revenue, predictable LTV, a relationship by default. By 2023, three quarters of DTC brands had some form of subscription offering.³ Most found out the hard way that a subscription is not a relationship. It is typically inertia. Churn ran at roughly ten percent a month — the average customer was gone inside a year, perhaps because the product failed, but more likely because a relationship was never built.⁴

True loyalty — the deep, trust-based kind — fell to 29% in 2025. Sixty-five percent of consumers feel genuine affection for fewer than three brands.⁵ Out of however many they've ever bought from. Most brands aren't one of them and defunding the moments that might have changed that is how they got there.

Subscribe now

Build, Renovate, Teardown

The real problem isn’t acquisition. It’s re-acquisition. For most brands that sell direct, owned channels — email, SMS, direct — account for a fraction of total revenue. The rest flows through paid. That includes your repeat customers. The ones you already acquired. Testing across thousands of DTC brands finds roughly 40% of retargeting conversions aren’t incremental, meaning the customer was coming back anyway.⁶ You paid for the attribution, not the influence. That’s a poor retention strategy it’s paying twice for the same customer. Most brands are doing it on a loop.

A rough illustration: if your blended customer acquisition cost is $75 and an unactivated customer comes back through a paid channel, you’ve spent $150 to reach someone who already bought from you. A vellum insert like Sézane’s costs cents at scale. A black and white piece of paper like Pika Pies’ costs even less. The upfront investment in an email program amortizes to under $2 per customer at any meaningful volume and under $1 by year two.⁷ Performance marketing costs rise with every bid. Owned infrastructure costs drop with every order. That’s the crossover, and for most brands with any meaningful repeat purchase frequency, the math has already moved.

What the calculation requires is a decision about what kind of problem you have. For some this is a build — the infrastructure was never there. For others it’s a renovation, there are some basics gathering cobwebs without a real strategy. For a few it’s closer to a teardown: the dependencies and gaps are structural and change will take time and will hurt before it helps. None of this is cheap. All of it is less expensive than the alternative.

What’s In Your Box?

There’s one more reason to care about what’s in the box. The purchase moments before it arrives are continuing to collapse in agentic commerce.

In March, Amazon went to federal court to block Perplexity’s Comet AI agent from shopping on its platform. The injunction was granted, then stayed on appeal — the fight is still live. But Amazon’s argument didn’t change: the agent doesn’t see the ads, doesn’t get upsold, doesn’t generate the browsing behavior that feeds a $68.6 billion advertising business.⁸ Amazon’s ad network is a Discovery Tax at platform scale. Perplexity is trying to make it optional. Amazon is right to fight it and about what is at stake.

The agent will keep getting better at buying. Faster, cheaper, more integrated into the places where people decide things. The moments before the purchase — the discovery, the comparison, the slow build of wanting something — will keep compressing. What the agent cannot do, at least not yet, is open the box. It cannot read the insert. It cannot feel the texture of the vellum. It cannot notice that someone wrote the note by hand, or that the design of the card looks authentic. It cannot form the impression that makes a customer come back without being re-acquired.

That moment still belongs to you. To us. To the builders of brands and customer love.

You already know what it costs to buy your own customers back. I’m wondering what you’re going to put in the box. Most haven’t decided it’s worth designing for. The bar is still on the floor. That’s the problem, and it’s also the window.

Subscribe now

Footnotes

¹ Kahneman et al., “When More Pain Is Preferred to Less,” Psychological Science, 1993.

² Binet & Field, “The Long and the Short of It,” IPA, 2013.

³ PipeCandy and Rodeo, “State of DTC Subscriptions,” Retail Dive, 2022.

⁴ Recurly, “Churn Rate Benchmarks,” cited in Marketing Charts. Box-of-the-month subscriptions: 12.71% monthly churn.

⁵ SAP Emarsys Customer Loyalty Index, 2025; Snipp consumer loyalty survey, 2025.

⁶ Measured, “Incrementality Benchmarks,” 2024. Analysis of 1,000+ DTC brands.

⁷ Illustrative figures. Actual CAC, insert production costs, and program costs vary by category and scale.

⁸ Amazon v. Perplexity AI Inc., US District Court N.D. California, 2026; Amazon Q4 2025 earnings.

Two Firms. One Room.

An intermediary arrives with something genuinely useful: reach, efficiency, a customer base it has already built. The price is the customer relationship. It is not a reasonable trade and someone will be standing outside with their face against the glass.

In 2003, the question on the table for a large auto lending business was whether to join DealerTrack — a platform that let car dealers submit financing applications to multiple lenders simultaneously, letting lenders compete in real time. Two consulting firms were working on the same question without knowing it (awkward!) A boutique had done ride-alongs with dealer sales teams, spent time on the floor, crunched the numbers. McKinsey had a contract several times larger. The recommendations were opposite.

The lender had built a reputation in the subprime auto market — loans to borrowers with poor or no credit, deals others wouldn’t touch because they had a different risk thesis. Dealers knew this. They routed specific profiles to our lender. That selectivity, their brand, was the edge. DealerTrack’s grid had no field for it. Joining was a race to the bottom. That’s what we said.

McKinsey’s case won the day: the efficiency gains were real, the market was moving.

Subscribe now

Data Compounds

DealerTrack grew from five lenders to more than fifteen hundred.¹ The lenders who built the platform — JPMorgan Chase, Wells Fargo, AmeriCredit — made a rational calculation: efficiency gains and an equity stake in the infrastructure. When Cox Automotive acquired DealerTrack in 2015 for four billion dollars, the founding lenders collected their premium. The lenders who joined without equity got the volume, and the competition volume required — faster decisions, higher approval rates, terms the platform set.

The price for audience and volume, at the time, was not discussed as relationship. It was described as data-sharing. Twenty years ago, the architect of Tesco’s Clubcard loyalty program put it plainly: data is the new oil.² Customer behavior, repeat patterns, the profile of who comes back and why — that is the refined product. What looks like a byproduct at the point of entry becomes the relationship in its most durable form. In a data-driven economy, when you hand a platform your customer data in exchange for volume, you are giving away the thing that was going to compound.³

McKinsey was right about the near and medium term. We were right about the final destination.

The Sad Empty Dining Room

The restaurant on the other side of the glass was running a similar calculation. DoorDash and Uber Eats charge restaurants between 15 and 30 percent commission per order.⁴ In exchange: reach, and during the years when dining rooms were closed, a way to survive. Those are real things.

What is also real: restaurant net profit margins average three to five percent. A 30 percent commission could invert it. The platform could earn more per delivery order than the restaurant does. By 2025, 53 percent of restaurant operators said they were actively trying to reduce their reliance on third-party delivery.⁵ The math had caught up.

The platform knows which customers come back, how often, what they order, where they are. The restaurant knows it filled an order and maybe a few insights from the platform. The customer who loved the burger enough to want to explore the experience, found a window. You cannot build a robust business that way. You can only build a kitchen that uses a very expensive distribution channel that it doesn’t own and can’t control.

The brands that understood the importance of data built direct infrastructure. Starbucks invested in its own app and loyalty program while competitors leaned into third-party platforms; it now counts 34.6 million active US members, with nearly 60 percent of its American sales running through that direct relationship.⁶ The data those customers generate including what they order, when they come, how often, what brings them back; stays inside the brand rather than accruing to an intermediary. What Starbucks was building was the thing the platform was trying to own. The Discovery Tax — the compounding cost of ceding how your customers find you, return to you, and choose you again — was already accruing for every restaurant that didn’t.

The Frictionless Trap

When you compete through someone else’s interface, you compete on their terms. What you could express before — the judgment that made certain dealers route specific profiles to you, or something that makes someone want to walk through your door — stops being legible. The interface is not a limi

tation. It is the point.

The rational economic case for joining is always real. That is The Frictionless Trap. It’s not that the argument is wrong, but that it is right for a shorter time horizon than perhaps the one that matters. The efficiency gains arrive as described. The race to the bottom arrives later.

The Next Wave

In January 2025, OpenAI launched Operator — its AI agent — with a Shopify integration that never got beyond thirty merchants before the company pulled back on that approach. Brands read this as a retreat. They were misreading it.

The actual move was the Operator layer, announced November 2024 and live with Target, Instacart, Expedia, and Booking.com by 2025: ChatGPT completing transactions inside the conversation without surfacing a product page, a brand, or a choice.⁷ The purchases still route through the retailer’s own app — OpenAI is not the checkout. It is the layer that decides where you go. That is the more powerful position. A checkout button can be replicated. The intent layer cannot.

When the agent handles the purchase, your brand doesn’t get rejected — it simply never appears. The customer never browses, never compares, never decides. There is no moment in which your brand could make an impression, earn a preference, or begin to mean something. This is Agentic Invisibility: the brands in the consideration set are never seen by the buyer. The comparison site era and the delivery platform era both at least required a decision on the part of businesses to join or not. This time, the agent layer isn’t asking.

What is actually worth protecting, and how do you hold it when the infrastructure is already a speeding train? These are the biggest questions marketers should be asking themselves — because nearly 75% of all restaurant traffic now happens off premises, according to the National Restaurant Association. Sounds more like a bunch of kitchens to me.

Footnotes

¹ DealerTrack: 5 founding lenders (2001); 1,400+ by December 2013 (BusinessWire); 1,500+ current. Cox Automotive acquisition: $4 billion ($63.25/share), October 2015.

² Clive Humby, ANA Senior Marketer’s Summit, 2006.

³ Hagiu & Wright, “Data-enabled learning, network effects, and competitive advantage,” RAND Journal of Economics, 54(4), 2023.

⁴ DoorDash/Uber Eats commission rates 15–30%, confirmed via merchant portals, Q1 2026.

⁵ Toast/National Restaurant Association operator survey data, 2024–2025. Net profit margin figures: Toast Restaurant Industry Report, 2025.

⁶ Starbucks Q1 FY2025 earnings: 34.6M active US members, ~59% of US company-operated store sales.

⁷ OpenAI Operator: announced November 2024, research preview January 2025, full ChatGPT integration July 2025. Confirmed partners include Instacart, Target (joined November 2025), Expedia, and Booking.com.

I was in Baltimore recently, visiting my niece. We ended up on a Sunday stroll around her neighborhood of Hampden — a fancy cocktail at The Dutchess, then checking out a nice cluster of independent shops.

At one point, Abby walked into a store not even adjacent to my current demographic. I went in because I was with a twenty-five-year-old, and that gave me a kind of permission to be in a space that wasn’t curated for me. I was the wrong person in the right place.

The second I walked in the door, I saw a sweater. I fell instantly in love. Something not beige. Sometimes that’s the only point in a purchase.

I picked it up, confirmed it wasn’t cringe, and bought it. I wouldn’t ordinarily admit to this but I’ve worn it every day for the last five days. I’ve gotten at least one compliment every single day. There was even a day I didn’t leave the house, but my sister saw it on FaceTime and said she wanted to steal it. Typical.

The point isn’t the sweater. The point is that I never would have found it deliberately, and not just because I didn’t know it existed. I also didn’t know I wanted it. Kelly green. Tiger motif. V-neck cardigan. Big buttons. Definitely not my usual cashmere. The preference was constructed in the act of finding it — the permission of the context, the age-appropriateness check with the niece, seeing it from across the room as one day someone would see it on me and also fall in love. It adds up to something I haven’t taken off in five days.

If I had typed a description of what I wanted into any search engine or AI assistant, I could not have described that sweater. The discovery required a specific configuration: a niece, a Sunday, a neighborhood I wasn’t navigating, a store I wouldn’t have entered alone, a window display that looked like it was having fun.

You cannot instruct an agent to replicate that. The instruction requires knowing what you’re looking for. The preference construction research says the preference didn’t fully exist until I walked in the store.

Subscribe now

An Internal Compass

This is not just a commerce problem, and it’s not new. In the early 2000s, researchers began studying what happened to Inuit hunters in the Canadian Arctic when they started using GPS.

These were communities with a wayfinding tradition that was centuries old. Not cartographic, not written down, but embodied. Hunters navigated by reading the environment: wind direction, snow texture, the behavior of animals, patterns in ice formation, star positions. Knowledge was passed from experienced travelers to younger generations in the process of travelling, on the land itself. The experience was the knowledge.

When GPS arrived, it brought meaningful advantages. In whiteout conditions or heavy fog, a device that told you where you were was a safety requirement, not a luxury. Hunters could mark food caches precisely and return to them. This is life and death stuff.

Within a few years, researchers Claudio Aporta and Eric Higgs documented what was happening beneath the obvious benefits.

Many younger hunters who had grown up with GPS did not develop the depth of knowledge needed to move about safely without it. The practice through which that knowledge was built had been replaced by a practice that produced no knowledge at all. The choice wasn’t conscious. You followed the arrow. And the arrow, it’s worth noting, is not always reliable: Aporta and Higgs record explicit concerns about GPS giving “ultimately undependable advice” in Arctic conditions. Hunters now going out dependent on things that were far from foolproof — signal, power, extreme temperatures — limitations traditional wayfinding did not share.¹

What Aporta and Higgs described was not a technology problem. It was a cognitive outsourcing problem. The cognitive work the tool replaces builds something we need, and there’s little discussion — in the rush to agentic visibility, consumer invisibility, and commoditization — about what we’re losing.

What the Science Says

There is an assumption baked into every recommendation engine ever built: your preferences exist. Meaning, they are formed, stable, and waiting to be matched.

The problem with this model is that it’s the thing we’ve been optimizing for almost exclusively, built a mind-boggling amount of market cap and several industries around — and it gets a lot of things really wrong.

A substantial body of research in behavioral economics shows that many preferences are not retrieved from memory. They are constructed in the moment of decision. Paul Slovic’s foundational 1995 work describes preferences as “remarkably labile,” shaped by framing, context, and the process of evaluation itself. In important or unfamiliar situations, preferences are “constructed on the spot.”² The preference isn’t just an input to the discovery process — it’s also an output.

You’re not helping someone find what they want. You’re matching a preference profile built from past behavior and presenting it to a person whose preferences would have been partly — or maybe even vastly — different if they’d encountered different things. The agent optimizes for the you that already exists. Discovery is a process that can shape who you are. It helps you develop taste, identity, ultimately a sense of belonging.

If preference is partly constructed in the encounter, then brands are not just competing to satisfy demand. They are competing to shape the conditions under which demand becomes legible to the customer in the first place. This is the ultimate brand-builder’s goal — not representing culture, creating it.

The neuroscience adds a second layer. Kent Berridge spent decades at the University of Michigan separating two systems we tend to conflate: wanting and liking. The wanting system is motivational: the pull toward something, the anticipatory engagement of pursuit, the feeling of being drawn. The liking system is hedonic: the satisfaction of having received something good. Berridge’s research established that these systems are genuinely separable, and that dopamine contributes primarily to wanting rather than to the pleasure of liking.³

As an environment becomes more predictable, as an algorithm gets better at delivering what you’d already like, the wanting system has progressively less to work with. Anticipation, cue-driven motivation, the response to something unexpected — all of this diminishes. The liking stays approximately constant. The wanting atrophies. An agent that reliably delivers exactly what you’d prefer is an environment from which a certain kind of desire has been engineered out. The products are fine. Commerce becomes less alive. If this was a marriage, we’d start talking about taking each other for granted.

Wolfram Schultz’s work on reward prediction error established why unexpected rewards land differently neurologically than expected ones. Dopamine response is substantially higher when the outcome is a surprise.⁴ Eliminate the unpredictability and you don’t just lose the delight of surprise. You reduce the neurological intensity of the reward itself.

The cognitive science closes the loop. In a 2014 study, Robert Wilson and colleagues established that humans use two distinct modes of exploration: directed, where we seek information about specific options, and random, the productive noise that leads somewhere unintended.⁵ Most commercial agents are built to minimise unproductive variance. The kinds of detours that often generate discovery are treated as inefficiency. The noise gets engineered out, but the noise was doing the work of expanding the territory of what we might want.

Three disciplines, pointing at the same structure. The process of discovery is not packaging around the outcome. It is constitutive of the outcome’s value.

What We’re Missing

Christian Busch defines the concept of serendipity with useful precision. Serendipity is not luck. Luck is passive — something happens to you. Serendipity requires your participation: an unexpected connection, noticing what you weren’t looking for but that made an impact.⁶

His framework identifies three necessary conditions: agency, surprise, and value. Remove any one of them and you have something else. Agentic commerce removes agency by design. The value you receive is real and you might be surprised, but the serendipity is gone.

The walker who navigates a city without a map and discovers a shortcut has built something.

We are building commerce infrastructure that translates on our behalf, accurately and efficiently, and we are not having an honest conversation about what that kind of future creates.

Future Tastemakers

Brands that cede discovery accumulate Discovery Debt — the compounding cost of underinvesting in the experiences through which people develop relationships rather than transactions. Brands that optimize entirely for algorithmic legibility win agent visibility while losing the human attention in which identity, desire, and loyalty actually form.

The discovery experience isn’t just a mechanism through which brands get built. It’s a mechanism through which people know what they want. If preferences are partly constructed through discovery, and wanting things runs on anticipation and surprise — both of which algorithms are designed to eliminate — then a commerce infrastructure optimised for frictionless delivery is one that progressively narrows the scope of what people can want. They’re not being served efficiently. They’re being contained.

Call it Preference Compression: the narrowing of taste that happens when the discovery process is removed and only the preference profile that already exists gets served back. The person whose commerce is fully mediated by a well-trained agent is someone whose tastes are slowly calcifying around what the algorithm already knows. The you-that-doesn’t-yet-exist doesn’t get created, because the conditions for its creation have been optimized away.

Brands that create the conditions for discovery are not just protecting a marketing mechanism. They are, in a real and uncommercial sense, serving a human need that is older than commerce and more durable than any channel.

What we know about the brands we love was built in discovery moments we’ve mostly forgotten. It didn’t arrive in a recommendation or in a brown box. It arrived in Hampden, or a lipstick counter in Paris to go perfectly with the dress, or a corner of the internet that no algorithm had suggested and that we’d never been able to find again.

The brands that remember this, and build for it, are making the investment that compounds — and will be the only brands shaping taste alongside revenue and loyalty.

Footnotes

1. Aporta, C. & Higgs, E. (2005). Satellite culture: Global positioning systems, Inuit wayfinding, and the need for a new account of technology. Current Anthropology, 46(5), 729–754.

2. Slovic, P. (1995). The construction of preference. American Psychologist, 50(5), 364–371. Payne, J.W., Bettman, J.R. & Johnson, E.J. (1992). Behavioral decision research: A constructive processing perspective. Annual Review of Psychology, 43, 87–131.

3. Berridge, K.C. & Robinson, T.E. (1998). What is the role of dopamine in reward: Hedonic impact, reward learning, or incentive salience? Brain Research Reviews, 28(3), 309–369. Berridge, K.C. (2007). The debate over dopamine’s role in reward: The case for incentive salience. Psychopharmacology, 191(3), 391–431.

4. Schultz, W. (1998). Predictive reward signal of dopamine neurons. Journal of Neurophysiology, 80(1), 1–27. Schultz was awarded the 2017 Brain Prize for this work.

5. Wilson, R.C. et al. (2014). Humans use directed and random exploration to solve the explore–exploit dilemma. Journal of Experimental Psychology: General, 143(6), 2074–2081. The application to commerce is the author’s inference.

6. Busch, C. (2024). Towards a theory of serendipity: A systematic review and conceptualization. Journal of Management Studies, 61(3), 1110–1151. 2024 JMS Best Paper Award.

Performance marketing has run a similar illusion for twenty years. The platforms aren’t generating new demand. They are harvesting the brand equity you already paid to build, capturing customers at the very end of their journey, and selling them back to you as “new acquisitions.” You think your ad spend is generating returns, but you are just being fed your own money, until the reservoir runs dry.

First, A Fish Tale

To understand how this happens without anyone noticing until it’s too late, let’s look at the collapse of the Northern cod fishery in the 1990s.

In 1968, the fishery hauled in a record 810,000 tons of cod. By 1992, the biomass had collapsed to less than 7% of its historical average, leading to a moratorium and the loss of 39,000 jobs.1 Right up until the collapse, the fishermen thought they were doing great.

They were using advanced sonar and massive trawlers. Because their technology was so good at finding the remaining schools of fish, their “Catch Per Unit Effort” stayed incredibly high, masking the fact that the underlying population was entirely decimated.

In performance marketing, the attribution algorithm is the sonar, and ROAS (Return on Ad Spend) is your Catch Per Unit Effort. Just as those boats became hyper-efficient at scooping up the last remaining fish, today’s algorithms have become hyper-efficient at finding the last remaining high-intent buyers. And, the dashboard metrics look incredibly healthy until your brand equity runs out.

So Much Hype

The proposition behind performance marketing was always built on a single claim: unlike brand investment, we can measure it. Brand was soft. Performance was accountable. The CFO could see it in the dashboard, and what the CFO could see was fundable.

I spent the better part of my career (very reluctantly) living in this story — first on the brand side, reorganising marketing strategies, departments, budgets around what turned out to be highly inaccurate dashboards, then inside one of the largest platforms perpetuating the mythology. The attribution system that credited our campaigns looked convincing. It was designed to. What we were actually measuring was channel activity, not business outcomes.

There are two mechanisms worth separating out, because they compound. The first is equity depletion: cut brand investment and you draw down the reservoir of memory structures, associations, and instincts that make someone reach for your product without thinking. That reservoir empties slowly. Typically brands have twelve to eighteen months before efficiency collapses. The second is attribution error: performance platforms were claiming credit for purchases that were going to happen anyway — buyers with existing brand intent, who happened to see a paid ad while completing a purchase they’d already decided on. These two mechanisms reinforce each other. As equity depletes, there is less demand to misattribute. As the misattribution runs, the depletion goes undetected.

Les Binet and Peter Field spent twenty years analysing the IPA Databank — the largest database of marketing effectiveness cases ever assembled — and produced the most cited and least acted-upon finding in the industry: the optimal marketing mix is roughly 60% brand building, 40% sales activation.2 Brand building fills the reservoir. Activation converts it. The CFO could see the activation, the reservoir wasn’t measured.

Orlando Wood’s research for the IPA added a further layer: as performance channels came to dominate budgets, the creative running through them was optimised for click-through rather than memory — shorter, more rational, more feature-led. Better at triggering action in the moment; progressively worse at building the emotional associations that make a brand retrievable when the buying moment arrives. The ads were optimised for conversion at the cost of the very thing that made conversion possible.3

Byron Sharp’s work at the Ehrenberg-Bass Institute established the underlying logic: brands grow primarily through mental availability — being thought of in more buying situations — not through conversion efficiency or loyalty metrics. Performance channels can trigger that mental availability. They cannot build it.4

Cut brand building and you can still run activation for a while. The reservoir exists, it drains, and for a period your performance metrics can actually look better — because fewer uncredited touchpoints means the paid channel claims more of the attribution.

Airbnb’s Famous Findings

In 2020, Airbnb cut all performance marketing. Not as a deliberate experiment — the world had stopped travelling and the spend had nowhere useful to go. But the results were striking enough that they chose not to reverse them.

Airbnb retained 95% of its online traffic with marketing spend reduced to near zero. CEO Brian Chesky: “What the pandemic showed is we can take marketing down to zero and still have 95% of the same traffic as the year before.”

They used it as cover to permanently restructure. They cut $541 million in performance spend, reoriented toward brand investment and editorial coverage, and described the shift as moving from “buying customers” to building something people sought out. By 2022, 90% of Airbnb’s traffic was arriving direct and not from paid channels. CFO Dave Stephenson called the shift “incredibly effective.” Q4 2022: EBITDA up 52%.5

What the pandemic accidentally revealed was this: the performance campaigns hadn’t been generating demand. They had been converting demand that already existed — brand-built intent that attribution models were crediting to the last ad served. When the campaigns stopped, the intent didn’t go away, because the campaigns hadn’t created it. The measurement had been wrong for years.

The Airbnb has a caveat: they could make this pivot because a global crisis gave them cover to absorb the short-term pain of withdrawal. Most brands don’t get that option. Performance spend suppresses competitors as much as it generates returns — the moment you stop, someone else takes your position. It functions less like investment and more like a subscription you can’t cancel (ironic). This is the liquidity trap at the centre of the scheme: it perpetuates not because it works, but because stopping is too costly to contemplate. What the Airbnb case actually shows is not that the exit is easy, but it is possible.

Subscribe now

The End of Social

Apple’s iOS 14 update removed the device identifier Facebook had been using to match ad clicks to purchases. Here is how that had been working: you see a Facebook ad on your phone, click it, and later buy on a website. Facebook’s tracking pixel on that website sends a signal back — “this device converted” — and matches it to the earlier click. The conversion gets attributed to the campaign.

When Apple removed the identifier and 96% of US users opted out of tracking, Facebook could no longer make that match. Reported conversions collapsed. But in documented cases, actual sales — measured independently through order management systems — hadn’t moved.6

The campaigns weren’t suddenly failing. Facebook had been taking credit for purchases that were going to happen anyway. Buyers with existing brand intent, who happened to see a Facebook ad in the process of completing a purchase they’d already decided on, had been logged as performance marketing successes.

The measurement that was supposed to make performance marketing superior to brand investment was running the same margin of error marketing always had, it just provided the appearce of health and the illusion of speed.

There is a separate question — one that compounds everything else — about what the channel became while we were busy measuring it (poorly).

In November 2025, Reuters published internal Meta documents showing the platform serving approximately 15 billion higher-risk scam ads per day. Internal analysis projected that around 10% of Meta’s 2024 revenue — roughly $16 billion — came from ads for scams, banned goods, illegal gambling, and fraudulent investment schemes. The company’s internal policy required 95% confidence before banning a fraudulent advertiser. High-spending accounts could accumulate 500 violations without losing access to the platform.7

Your performance budget has been competing for space in a system deliberately structured to protect its most fraudulent customers. The context your ads run in is not a minor footnote.

Daily time spent on social media is down nearly 10% since 2022. Facebook’s share of organic content from people users actually know has dropped from 22% to 17% in two years; Instagram’s from 11% to 7%. The feeds have filled with AI-generated content, synthetic engagement, and automated slop. People are present but emotionally absent. Performance marketing running in that context is not the same as performance marketing running in the social media context that people loved. The ROAS figures have not reflected the difference, but they will.8

New Interface, Similar Issues

In February 2026, OpenAI launched advertising in ChatGPT at $60 per thousand impressions, with a $200,000 minimum spend. The ads appear below AI responses — outside the answer, OpenAI noted. “Ads do not influence the answers ChatGPT gives you.”99

That is exactly what Google said about AdWords.

Actually, that comparison undersells the problem. When Google introduced advertising alongside search results, there was still an impression. A user saw the ad. They might click or not — but there was a moment of human attention in which a brand could exist.

In an agent-mediated environment, there is no impression. The customer was never in the room. The agent evaluated, decided, and executed without any moment in which a brand could be encountered. Performance advertising on social could occasionally behave like brand advertising: interrupt, introduce, create a flicker of curiosity. The agent environment removes that possibility entirely. There is no discovery moment to sponsor. There is only the output and the brown box on the doorstep.

The Discovery Debt — the compounding cost of ceding how customers find you and decide you’re worth choosing — accumulates faster here than before. The reservoir isn’t just being drawn down by performance spend. The channels through which it was being partially replenished is being removed from the customer’s experience entirely.10

Channels Were Never Meant to Be Strategy

Performance marketing is not the problem. It is a set of channels — paid search, social advertising, display, programmatic — and channels do what channels do. They distribute. They report on distribution. They optimize for their own metrics. None of that is wrong.

The error was in ever treating a set of channels as a strategy. When channels become the plan, organizations manage for channel metrics, because those are the metrics available. CFOs who could see the dashboard defunded the efforts that filled the reservoir they couldn’t see. The 60/40 ratio inverted (or worse). For years the performance metrics continued to look acceptable, because they were drawing on equity built before the inversion happened.

The brands that came through this period in better shape are the ones that refused to make that trade. On Running and Hoka weren’t built on performance spend. They were built on product conviction and community — run clubs, niche trail athlete sponsorships, silhouettes distinctive enough to start real-world conversations before a digital ad ever served. The performance channels, when used, amplified something that already existed. That is what they are for.

What the platforms extracted, over twenty years, was the relationship between brands and the people who might choose them. The belonging, the identity, the sense that a brand was something you found and chose rather than something that found you. They intermediated that relationship so much that many brands now have no direct path to the people who buy them. And now those same platforms are selling access back — first as performance advertising, then as AI placement, then as whatever the next channel is. The platforms became the managers of relationships they didn’t build. The brands are paying rent on something they used to own.

The cod fishery wasn’t destroyed by fishing. It was destroyed by fishing without good measurement and without a plan for renewal. The trawlers were technically excellent. The sonar was accurate. The catch was real. What was missing was any investment in the stock the catch depended on because the measurement metrics were just plain bad.

The moratorium was declared in the year I graduated from high school (Go Saxons!) and I can tell you, that was a very long time ago. The Northern cod stocks still haven’t recovered. The fishers who kept hauling right up until the last day still talk about how healthy the water looked.

Ponzi schemes are deceptive, dangerous, and, ultimately, doomed.

Every brand steward has a version of this story. The moment the briefs shifted. Perhaps it arrived via a new funding round and bonkers expectations or perhaps just a new set of KPIs that did not include a column for things like wonder or wander. We started picturing the funnel not as a journey but as a series of ways to, essentially and sometimes literally, keep the customer “in line” and moving towards checkout.

Marketers who’d spent careers understanding humans and their relationships to brands weren’t absent from these conversations. We were out-hyped, out-shouted. And, this is the part that doesn’t get said out loud: marketing shifted from a job that attracted people with empathy to folks who think everything can be explained in 0s and 1s. Performance marketing came with a dashboard and created a veneer of health and speed. Everyone had the same enemy: Friction. Arguing for “brand” to a leadership that leaned into extraction wasn’t a battle that came with much armor.

Of course, our dashboards were measuring only the part of the story that was easier to quantify.

Once Upon a Time, We Leaned Into Wonder

In July 1909, Harry Gordon Selfridge’s famous department store had been open for four months. He’d staked everything on a single idea: that shopping should be an experience worth having whether you bought anything or not. Wanderers were as welcome as buyers. You could come in, look at things and leave empty-handed, but you probably didn’t.

When Louis Blériot crossed the English Channel on July 25th, Selfridge moved immediately. Within days, Blériot’s monoplane — wire, ash wood, a 25-horsepower engine, still smelling of oil and canvas — was in the store window. 150,000 people came in four days to stand at the rope. Browsing through a store with a brain lit up with wonder? Cha-ching.

Simply being awed was the initial offer. The purchase was a byproduct. Selfridge understood emotion and resonance, he knew humans want to be inspired. Feeling something is a revenue driver, just not one that fits neatly in a dashboard.

Recently, I wanted to explore an author I’d been meaning to read. Ordinarily that would have meant surveying friends who knew my taste and going down a few search rabbit holes. Doing the kind of mapping that takes a bit of time and leaves you with a sense of the territory before you enter it. Instead, I asked an AI where to start. It gave me the most acclaimed work. It was excellent. But I came to it without the journey: without the early work that explained the obsessions, without the middle period where the style crystallized. Going back to the earlier work afterward felt like watching the highlights before the match. The AI answered the question I asked but removed the opportunity for me to arrive at a better question for how I like to explore like: what’s a good way in?

That’s what optimizing does. And it usually looks good…on paper.

Handing Over the Keys

My first inkling of what this would become started with comparison sites in the mid-2000s. Credit Karma, CreditCards.com, and the other aggregators turned years of card brand-building into rows of APR numbers and annual fees in a database. Brands that had invested in actual relationships found themselves losing ground to institutions that had learned to compete on the headline number, recouping it elsewhere in fees and terms the algorithm wasn’t designed to surface.

The real casualties were the consumers, of course. But the smaller lenders and credit unions didn’t fare well either. The credit unions lost the fight for billions in new customer spend not because their product was worse — quite the contrary in many respects — but because the aggregators were measuring the easy things, the wrong things, and therefore so were credit card shoppers.

The comparison table converted (huzzah), but it was reductive.

SaaS Marketing, Please Take Your Toys and Go Home

Enter DTC. The pitch turned idealistic: cut out the intermediary, build something people actually love, own the customer. We got more excited about “disruption,” “efficiency,” and “the model” than we did about the products. And the consumers? Just numbers going up or down on a spreadsheet.

Do you remember Fab.com? It launched in 2011 around an interesting idea: a daily design marketplace where the chief creative personally curated what went live. There were limited quantities, 24-48 hour windows, and objects you wouldn’t find anywhere else. The model was built around a question: what cool thing might I find today that I didn’t know I wanted? Within six months, a million people had signed up.¹

Then Andreessen Horowitz arrived with $40 million. You take on that kind of money and growth targets are going to be the only thing you talk about. And, whatever model you were using is probably going to break as you try and achieve “scale.” In most cases, the pressure of that money means the very things that made a brand experience worth returning to will be eliminated. In Fab’s case this meant merchandising teams optimizing for volume and losing the cachet. The curated flash sale became a catalogue. The catalogue became a warehouse. The warehouse needed to move product. By 2013, Fab was trying to be the Amazon of design. At some point, they started selling steaks.

Traffic dropped 75% in under a year. Valued at $1 billion in 2013, Fab.com sold in 2015 for somewhere between $15 and $50 million. The founder wrote a memo to his staff: “I guided us to go too fast. I enabled us to lose our core focus.” The discovery was the product. Once it was gone, there was no reason to come back so people didn’t.

A decade after Fab’s fire sale, I launched a DTC menopause brand leaning beauty and lifestyle. Subscriptions were how we planned to justify what it actually costs to build a brand: the slower work, the development of narrative, audience, and advocacy. In practice, every tactic was optimized for the cart. Upsell, cross-sell, add-to-cart, you might also like. The whole architecture was oriented toward more sales without earning more attention.

And, the customer who receives something month after month without thinking isn’t loyal, they’re inert. Loyalty survives a better offer. Inertia doesn’t. Building the behavior without building the relationship is a very short-term strategy.

Marketing stopped being interesting (let alone fun) for years before I really understood why. None of this was sustainable. Extraction doesn’t compound. Relationships, however, do.

Stitch Fix followed the same old board-driven logic. Launched in 2011 on a genuine idea: human stylists who learned what you liked and sent you things you wouldn’t have found yourself. It IPO’d at $1.6 billion.² Under pressure to grow, it built toward an “AI-first” model and in doing so, systematically dismantled the human styling layer and its point of difference. The stock is down more than 90% from its IPO price and it’s currently in a turnaround.

What Stitch Fix became was just another algorithm competing against Amazon’s recommendation engine and a proliferation of similar services. In a room where everything has to be quantified and is expected to become more efficient, what’s human gets optimized out.

Along Came Shopify

Shopify. The great enabler. In January 2026, Tobi Lütke — CEO and co-architect of the Universal Commerce Protocol — called the agentic era’s central promise serendipity. “I would have never searched for this product,” he said, “but somehow it found me.”³

What Lütke is describing is targeting. Serving things up. A passive customer experience.

Christian Busch, who’s studied serendipity for years, has a clear definition based on research and actual human behavior, not wishful thinking: luck happens to you, serendipity requires your participation, its an unexpected connection made by a person that went looking.⁴ Lütke didn’t go looking; the algorithm did, and presented him something accordingly. No discovery, just purchase. One builds attachment, anticipation, ownership. The other is just a brown box showing up on the doorstep.

The comparison site. The subscription. Now the agent. This is the Frictionless Trap: every step may feel like progress but this path strips away things we — brands, customers, investors — need.

BCG and Moloco found 33% of consumers now discover brands through AI agents.⁵ The Discovery Tax: the rapidly compounding cost of ceding how customers find you, was already accruing before agents arrived. This is when the bill comes due, and it’s going to be a doozy.

Go Put a Plane in Your Window

Gentle Monster, a South Korean eyewear brand, employs six people to design its glasses and sixty to design its stores.⁶ The stores are immersive art installations — kinetic sculptures, robots, themed environments that shift with each collection. You go for the experience and then you buy glasses. They’ve built something the algorithm cannot evaluate: a reason to walk in.

Glossier went back to what Emily Weiss had built it on — community, conversations, products developed from actually listening to what people wanted. Weiss had been pushed aside in favor of a growth model the brand couldn’t sustain. The recovery came from returning to the founder’s original instinct: the community is the product, the goods are the proof. By 2024, they were profitable.⁷

Both remembered that wonder (and wander) should be part of the offer.

The brands that will survive what’s coming are going to be the ones giving people a reason to form a relationship. That’s the investment the dashboard can’t measure and the agent can’t shortcut. It’s also, as it turns out, the only kind that compounds.

Selfridge didn’t remove friction. He put a plane in the window and wanted people to come see it.

Most brands stopped looking for the plane the moment the spreadsheet told them wonder didn’t have a conversion rate.

Footnotes

1. Fab.com founded 2011 by Jason Goldberg and Bradford Shellhammer. Raised $333.7M total. Reached 1 million members within six months. Peak valuation approximately $1B (2013). Sold to PCH International in 2015 for an estimated $15–50M. Goldberg’s internal memo to staff, October 2013.

2. Stitch Fix IPO, November 2017. Approximately $1.6 billion valuation. Founder Katrina Lake. Restructured toward an algorithmic model under investor pressure, reducing reliance on human styling. The company has faced sustained share price decline from its IPO valuation.

3. Tobi Lütke, quoted at the Shopify/Google Universal Commerce Protocol launch event, January 2026. UCP is a proposed open standard enabling AI agents to handle product discovery, cart management, and checkout across platforms.

4. Busch, C. (2020). The Serendipity Mindset: The Art and Science of Creating Good Luck. Riverhead Books.

5. BCG and Moloco, “AI is Collapsing the Marketing Funnel,” 2026.

6. Gentle Monster brand profile. Founded 2011, Seoul. 81 stores across 14 countries as of 2025.

7. Glossier profitability reported 2024. Emily Weiss stepped back from CEO role 2022; brand recovery credited to return to community-first positioning.

The Rise of the Machines

In September 2025, OpenAI launched Instant Checkout. A way for consumers to buy directly inside ChatGPT, powered by Stripe and will be integrated with over a million Shopify stores.¹ In January 2026, Google announced Universal Commerce Protocol at the National Retail Federation conference. It created an open standard for agent-based purchasing, backed by Shopify, Etsy, Wayfair, Target, Walmart, and my former employer Visa.²

And then there’s Amazon, the godfather of platform disintermediation. Its “Buy for Me” has been live since April 2025. It lets an AI agent purchase from third-party sites without the customer ever leaving Amazon. Tens of thousands of items were available by the end of the year, including some from brands that never agreed to participate.³ The big walled garden just expanded its acreage and absolutely no one should be surprised. And, it will continue.

eBay’s updated user agreement goes live tomorrow, explicitly banning “buy-for-me agents, LLM-driven bots, or any end-to-end flow that places orders without human review.”⁴ It’s enacting a policy against something that barely existed twelve months ago and taking a pretty unique stance.

Disintermediation is not a 2030 conversation

Salesforce and Adobe independently converged on the same finding: around 39% consumers are already using AI to shop.^6,7 Not “curious about,” not “open to trying;” already doing it.

Brand teams are paying attention. Most are quickly trying to make content more parseable by AI. Structuring product data for agents, not humans. Trying to win what I’d call Top of Algorithm: being the brand the AI recommends.

And it’s not like they’re wrong, precisely. Brands cited in Google’s AI Overviews see 35% more organic clicks and 91% more paid clicks than those that aren’t.⁸ AI-referred traffic spends 45% more time on site.⁹ The traffic is smaller but richer. This is a great result for marketers.

Algorithmic Legibility works. Ish. And, while your board asks the same question at every meeting: exactly how fast you can win this new game? The question rolling around in my head is: what are brands actually winning and how much of this “optimization” can they survive?

You ask an agent to buy moisturizer. It evaluates a thousand options against your criteria — price, reviews, ingredients, delivery speed — and purchases one. You never saw the packaging. Didn’t see the brand story. There was no comparison against the one you used to love or the brand your friend raves about (Rhode, of course). The agent matched criteria and completed the transaction. You didn’t choose that brand, you received it.

Yes, brands that customers specify by name will still win. But how many of your customers, let alone prospects, will name you?

An agent that purchases you is not a customer who chose you. Algorithmic Legibility and Agentic Invisibility are two sides of the same coin. You can be perfectly readable to the machine and become completely invisible to the human. The agent matched criteria. The customer received a product. Nobody chose anything.

Google Gemini is out-Googling itself. The industry’s only response is to optimize harder for the thing that’s replacing a bunch of important customer touch points the brand used to own.

Right Now, It’s No Better Than A Coin Toss

So, folks who are jumping on the Legibility bandwagon, aren’t wrong; but they’re also not in charge.

SparkToro found less than a 1% chance the same brands appear across two AI responses to the same query.¹⁰ Your brand isn’t being rejected. It’s flickering, present in one answer, absent the next with zero transparency either way for the brand or consumer.

If you ask me, big problem actually isn’t inconsistency (though that is a bit of a whopper). Or even our own lack of understanding about how the tech works. We’ve been at the mercy of the black boxes for a while now. The difference is that with traditional performance marketing, you appeared in the results and the customer still made a choice. With agents, the cognitive labour of evaluation — the comparing, the weighing, the deciding — shifts off-screen entirely. The customer never sees the analysis. The agent evaluates, the agent buys, and your brand was either there or not. This is evaluation compression, and it’s a fundamentally different kind of invisibility than anything we’ve dealt with before.

On this side of the equation, the news is not so nice. Organic click-through rates have dropped 61% on queries where AI overviews appear. Paid clicks are down 68%.⁸

Replicating the Performance Marketing Ponzi Scheme with Higher Stakes

The industry is solving a commercial problem: how do I stay visible to AI agents? And few are talking about the extent of our burgeoning brand problem: what happens to meaning, attachment, and discovery when the customer is removed from the process entirely?

The commercial problem centers on Algorithmic Legibility which is about data structure, readability, being in the agent’s consideration set. Necessary, but insufficient. An AI agent doesn’t think of anything. It evaluates. It compares. It optimizes. Mental availability, that prize of being the brand you think of first, is irrelevant to a machine that has no mind.

Take for example, CPG brands. They’ve never owned the customer relationship; retailers have always held the data. They survived on mental availability, of being the brand you reach for without thinking. An agent has no mind to be “top of.” Fashion is arguably worse: the entire value proposition is built on discovery, desire, identity. The BoF-McKinsey State of Fashion 2026 indicates whoever controls the AI agent and its training data controls how the brand is perceived — when a bot, not a person, is doing the buying.¹² A luxury handbag reduced to a line item in a chatbot response has literally lost almost everything that made it a luxury handbag.

And even if you win the visibility game and your product data is perfectly structured, your reviews are plentiful, your content is beautifully optimized — you’ve won the right to be purchased. Not chosen. Purchased. The customer never browsed, never compared, never felt anything. That’s not brand loyalty. That’s procurement. And nobody has ever fallen in love with sourcing.

CFOs Should Be Losing Sleep

This is not just a brand problem. It’s a margin problem. A brand that is not chosen has no pricing power. No differentiation that survives a filter. In the agent’s eyes, a commodity. And commodities compete on price until someone loses.

Algorithmic Legibility or Agentic Invisibility. Two problems, not one. The industry is only working on the easier one.

The harder one, how you preserve what makes your brand mean something when discovery itself is being automated, is the one that will actually determine who survives this.

Every quarter you don’t address it, you’re paying a Discovery Tax: the compounding cost of ceding how customers find you, experience you, and decide you’re worth choosing again.

That tax has components. Reduced premium tolerance as customers lose attachment to brands they never chose. Substitution volatility as agents swap you out on marginal differences in rating or price. Rising retention costs to replace the relationship building that used to come from discovery. Growing dependence on agent-facing optimization you don’t control means your also at the mercy of a whole new set of platforms.

The math on this is not abstract. It’s in your P&L.

What’s your brand’s Discovery Tax? That’s not a rhetorical question. And it’s the one your board should be asking this quarter.

Footnotes

1. OpenAI, “Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol”, 29 September 2025.

2. Google, “Agentic Commerce: New AI Tools and Protocol for Retailers”, 11 January 2026. Announced at NRF 2026.

3. Modern Retail, “Brands are upset that ‘Buy For Me’ is featuring their products on Amazon without permission”, January 2026. See also CNBC.

4. EcommerceBytes, “eBay Bans AI Shopping Agents, Updates Arbitration Provision”, 21 January 2026. Effective 20 February 2026.

5. BCG, “Consumers Are Rewriting the Rules of Year-End Sales Events”, November 2025. 48% of consumers used or planned to use GenAI during year-end sales.

6. Salesforce, “With AI Adoption Surging, Shopping Behavior Is at an Inflection Point”, 2025. 39% of consumers using AI for product discovery; over 50% of Gen Z.

7. Adobe, “Generative AI-Powered Shopping Rises with Traffic to U.S. Retail Sites”, 2025.

8. Seer Interactive, “AIO Impact on Google CTR: September 2025 Update”, September 2025. Analysis of 3,119 queries across 42 organisations, June 2024–September 2025.

9. Adobe, “AI-Driven Traffic Surges Across Industries”, 2025.

10. SparkToro, “NEW Research: AIs are Highly Inconsistent When Recommending Brands or Products”, December 2025.

11. PwC, “CPG Industry Self-Disruption”, 2025. Survey conducted April–May 2025.

12. Business of Fashion & McKinsey, “The State of Fashion 2026”, 2026.

I spent a decade at Visa watching money go from a plastic to invisible — friction disappearing from payments one tap at a time. Then two years at Meta watching algorithms get extraordinarily good at predicting what people wanted. We were removing friction everywhere, and calling it progress. Turns out, we were also removing something we didn’t know we needed.

Finding and discovering are not the same thing. The industry uses the words interchangeably but they are not synonymous. That confusion is about to cost brands everything.

Finding is what you do when you know what you’re looking for. You need jeans. You search for jeans. You find jeans. Transaction complete.

Discovery is what happens when you don’t know what you’re looking for — and then you do. You went in for jeans. You left with cords. Hadn’t considered them in about twenty years, but seeing the texture or color sparked a memory of autumn and being twelve and feeling a certain way. You didn’t find cords. You (re)discovered them. And that discovery changed what you wanted.

The difference isn’t semantic. It’s neurological. The brain processes expected and unexpected information through fundamentally different pathways. When you search for something and find it, you get a small dopamine hit from the reward of task completion. When you encounter something unexpected that turns out to be valuable, research suggests a 30-40% stronger response.¹ The surprise matters. The unexpectedness is the point.

This is why the product you stumbled upon stays with you longer than the product you were handed.

The cognitive science is unambiguous about what discovery requires.

Christian Busch, who’s spent years studying serendipity, identifies three necessary conditions.² All three must be present:

Agency. The discoverer must be actively engaged in the process.

Surprise. The outcome must be unexpected.

Value. The discovery must be recognized as worthwhile.

Remove any one and you don’t have discovery, you have something else. Luck is surprise without agency. Targeted search is agency without surprise. Noise is surprise without value. Delegate your agency and optimize against surprise — that’s the Agentic Commerce Bargain.

Agentic commerce disintermediates brand by design. You tell an AI agent what you need. The agent acts on your behalf. Products appear. The efficiency is extraordinary. But where is the agency? You delegated it. Where is the surprise? The agent optimised it away. What you’re left with is retrieval — sophisticated, preference-matched, increasingly accurate retrieval. Useful. But not discovery.

I watched a version of this at Visa. We talked about “top of wallet” — the card you reached for first. Physical wallets had a bit of friction built in: reaching past other cards, seeing alternatives, maybe a moment of reconsideration. Digital wallets collapsed it. Your default became invisible. Issuers rushed to claim the top spot because they understood once claimed, displacement was highly unlikely.

AI agents will be “top of wallet” for everything – “Top of Algorithm,” really. You won’t even see the wallet. You’ll delegate; the agent will execute. The brands it doesn’t surface don’t just lose the transaction, they aren’t ever visible.

I keep coming back to a phrase tech loves: “serve up.” We serve up recommendations. We serve up offers. We serve up content. The language frames discovery as something done to consumers. We serve. They receive.

But the science says discovery only works when it’s done by consumers. The value of discovery – neurologically, psychologically, and it turns out, commercially — depends on the discoverer being an active participant, not a passive recipient.

This isn’t a new tension. But agentic commerce makes it explicit in a way previous technologies didn’t. An algorithmic feed still required you to scroll, to stop, to engage. An AI agent doesn’t even require that. You delegate. It acts. Discovery becomes fully passive.

Which means, it stops being discovery at all.

The worship of signal has blinded the industry to the value of noise.

At Meta, we were steeped in the language of signal and noise. Signal was worshipped as the behavioral data that revealed intent. Noise was reviled as the random variation that obscured signal. Every optimization was meant to eliminate noise and promote signal.

But the science of exploration suggests this binary is false (as binaries typically are). Robert Wilson’s research shows humans don’t just use one mode of exploration, they use two.³

Directed exploration is what you do when you’re trying to learn something specific. You seek information about uncertain options because you want to make a better decision. This is the mode AI agents excel at supporting. This is signal-following.

Random exploration is different. It’s noise in the decision-making process, choices that aren’t perfectly optimized, and detours that don’t obviously serve the goal. It looks inefficient because it is. It’s also how we discover things we didn’t know we were looking for.

Agentic systems only support directed exploration. You tell the agent what you want. The agent finds options. There’s no mechanism for random exploration because random exploration looks like failure from an optimization standpoint.

But noise — what the industry treats as the enemy — is actually where serendipity lives.

Your Instagram algorithm can’t make the jeans-to-cords leap because cords don’t optimize for the engagement signal jeans created. An AI shopping agent won’t suggest cords because you didn’t ask for cords. The entire architecture is designed to eliminate the “noise” that random exploration depends on.

How did we get here, you ask?

Well, the people building the systems upon which commerce now flows are engineers. Their native language is binary — literally, zeros and ones. Signal or noise. Convert or bounce. Optimized or inefficient.

Their systems have no word for “valuable encounter that doesn’t convert for six months.” No metric for “planted a seed.” No category for the wander, the maybe, the interjection of kismet.

If your language can’t describe (or even acknowledge) serendipity, your system won’t deliver it. If your metrics are binary, your optimization will be too. And if you’re building the infrastructure the world runs on, you’ll build in black and white and there won’t even be room for grey.

Tech thinks in binaries and its systems are remaking everything in its image.

Every tech “advancement” has done the same thing: removed friction.

Catalogues to search. Search to algorithms. Algorithms to agents. Each step more efficient. Each step destroying a little more of what made discovery possible.

Friction is where discovery lives. The browse. The wander. The “while I’m here I might as well” moment. What tech calls friction is what most humans call experience.

Consider the good ol’ Sears Wish Book (Google it, kids) — the annual catalogue families used to spend hours with, dog-earing pages, circling items with pencil, negotiating wants around the kitchen table. That ritual was friction. It was also ownership, anticipation, and relationship. The act of circling made items yours in imagination before purchase.

Amazon’s infinite scroll is frictionless. It’s also meaningless. Nothing is ritualized. Nothing is anticipated. Nothing is yours until it arrives in a brown box you’ve already forgotten you ordered.

We’ve frictionlessed commerce to death. What we call convenience is actually the hollowing out of relationship and meaning.

The business case is just as uncomfortable.

Brand building creates durable memory structures. Performance marketing captures demand but decays quickly. The research says optimal investment is roughly 60/40 brand to performance.⁴ But for twenty years the industry has inverted this, seduced by attribution models that credit the last touch and hide the discovery that made conversion possible.

Performance marketing is a Ponzi scheme. It draws from a reservoir it doesn’t fill.

Nike just proved what happens when you drain it. They over-indexed on performance marketing and DTC, gutted their brand investment, and lost $70 billion in market value. The CEO was replaced. The company publicly reversed course back toward brand and the retailers they’d abandoned.[^5]

And yet consultancies and pundits are now telling everyone else to optimise for AI agents. Be “agent-preferred.” Make your catalogue machine-readable.

Brands are about to win every transaction and lose every relationship.

We’ve seen this movie before.

Hotels ceded customer relationships to online travel agencies. Now they pay the Discovery Tax: 15-30% commissions per booking, zero customer data, and more limited ability to build loyalty.

Artists gave discovery over to Spotify’s playlist algorithms. Now they earn less than a third of a penny per stream. 87% of tracks receive zero meaningful compensation. The platform decides who gets discovered. Artists get streamed…or they don’t.

Amazon third-party sellers receive orders from anonymous proxy emails. No direct customer contact. No relationship building. No loyalty. Amazon identifies bestsellers, then creates Amazon Basics to compete directly. The sellers built the demand. Amazon captured it.

Every wave of disintermediation follows the same sequence: new intermediary offers convenience, brands accept lower margins to maintain access, intermediary captures customer relationship and data, brand differentiation collapses into commoditized features.

Agentic commerce threatens to be more severe than any previous wave. Total autonomy capture. No direct customer contact. Price as likely primary differentiator. Brands won’t just be disintermediated, they’ll slip into Agentic Invisibility. Existing in datasets but not in human memory. Present in an catalogue of sorts, but absent from the consideration set. Technically available but functionally gone.

So what’s the answer?

While everyone’s current instinct is to ask: how do we optimize for AI agents? How do we become “discoverable” to the algorithms that will decide what gets bought?

This is the wrong question. It’s the question hotels asked about OTAs — and look where it got them.

The right questions are harder: How do we build discovery experiences that create loyalty agents can’t capture? How do we create customer relationships that exist outside the agentic layer? How do we make discovery the product, not just the pathway to purchase?

The brands that survive will be those who understand that discovery in and of its self is valuable — not as a means to transaction, but as the thing that creates meaning, memory, and attachment. They’ll build experiences that resist delegation. They’ll create friction that people actually want.

They’ll understand that efficiency isn’t the only thing consumers are optimizing for. That autonomy is a basic psychological need, not a luxury to be traded for convenience. What tech calls friction is often what humans call experience.

Some of this will start with the other d — delivery. Packaging, unboxing, and the moment something arrives: are touchpoints agents can't own, and most brands still treat as logistics. Some will mean making stores genuine destinations again — the Selfridges and Marshall Fields model, where shopping was never just buying.

But none of it works if you don’t understand the core distinction, discovery isn’t finding. The brands that can’t tell the difference are about to learn the cost.

Footnotes

1. Schultz, W. (2016). Dopamine reward prediction-error signalling. Nature Reviews Neuroscience.
   
   

2. Busch, C. (2024). Towards a Theory of Serendipity. Journal of Management Studies.
   
   

3. Wilson, R.C. et al. (2014). Humans use directed and random exploration. Journal of Experimental Psychology: General.
   
   

4. Binet, L. & Field, P. (2013). The Long and the Short of It. IPA.
   
   

5. Nike lost ~$70B market value 2024; CEO John Donahoe replaced by Elliott Hill, who committed to shifting from performance back to brand.