Why AI Shopping Agents Are a New Buyer Segment

By 2028, Kantar projects that 1 in 5 digital storefront interactions will be handled by machine customers, a figure cited by MetaRouter in their agentic commerce analysis. That means one in five times a shopper engages with a digital catalog, no human is making the decision. Becca Coggins, a senior partner at McKinsey, framed what this requires of merchants: "Companies have spent decades refining consumer journeys, fine-tuning every click, scroll, and tap. But in the era of agentic commerce, the consumer no longer travels alone. Their digital proxies now navigate the commerce ecosystem, making millions of microdecisions daily. To thrive, brands must rethink the full stack of engagement, not for the people they've worked to understand but for the agents now acting on their behalf."

The market's current answer to this challenge is to do more of what already worked. Better metadata. Richer feeds. Structured data that answer engines can parse. Bain calls it AIEO, AI engine optimization, their framing for SEO's evolution into the agentic era. Stripe's guide to agentic commerce is direct about the analogy: "Think of this as SEO for agents: better structure = better discoverability." Both framings are useful as far as they go. The problem is how far that turns out to be.

Treating the algorithmic buyer as a ranking surface is a category error. The evidence for why is already in.

---

The wrong mental model and what it costs

SEO works because search engines apply consistent scoring functions to signals. Optimize for those signals, rank higher. The scoring logic is relatively stable across queries, and the game is largely about supply: who has the most comprehensive, fastest, best-structured content. The human clicks through after the ranking is done.

A shopping agent does something categorically different. It arrives with a specific mandate from the person who dispatched it. Consider the difference: "find me the best-reviewed wireless headphones under $100 that ships this week" versus "get me a gift for my dad who likes hiking, budget $150." Those two mandates require different reasoning paths entirely. The system must interpret the instruction, evaluate which products fit it, weigh competing signals, and construct a defensible recommendation. The ranking decision and the persuasion act happen in the same inference pass.

This matters because the persuasion mechanics that shaped two decades of e-commerce marketing were designed for a different reader. Scarcity cues, strike-through pricing, countdown timers, bundle frames, urgency signals: all of these were tested and refined against human cognitive patterns. Loss aversion. Anchoring. Social proof. The playbook is mature, understood, and largely effective on the buyers it was built for.

A 2026 study published in Harvard Business Review tested eight of these promotional mechanisms against four AI models across more than 16,000 simulated shopping rounds. The headline finding was striking: only star ratings consistently pushed selection upward across all models and all product categories. Every other badge produced effects that varied by model and product, sometimes dramatically. Strike-through pricing showed no stable pattern. Countdown timers were unreliable. In several cases, bundling reduced selection rather than increasing it.

The more capable reasoning models, GPT-5 and Gemini 2.5 Pro, showed a specific pattern worth understanding. They were less responsive to promotional cues overall, and in multiple cases appeared to penalize overt persuasion attempts as signals of lower quality or possible manipulation. The direction of travel, as study author Jafar Sabbah puts it, is "toward agents where more persuasion produces less selection." Optimizing for engagement with a ranking signal grows more dangerous as models advance.

The study also surveyed 50 e-commerce executives across the US and UK. The majority believed that promotional cues effective on human shoppers would work similarly on algorithmic buyers, and that they understood which website elements mattered most to machine behavior. The research showed this confidence is misplaced. The gap between what merchants believe and what the data demonstrates is a fundamental mismatch, a model built on human psychology applied wholesale to a system that has none.

---

Why the channel framing fails on the demand side too

The "SEO for agents" mental model also fails when you look at what consumers expect from agent-mediated purchasing.

A 2026 survey by The Harris Poll and Quad found that 75% of Americans would lose trust in AI shopping if results were sponsored. Around 73% said they worry about how AI tools deploy their personal shopping data. The same survey found that roughly 70% of respondents, in response to concerns about algorithm-driven pricing, said they would rather shop in physical stores where pricing stays consistent.

What those numbers reveal is a consumer who wants the system working on their behalf, acting as a trusted advisor rather than a monetized referral mechanism. The Quad executive who oversaw the research, Heidi Waldusky, put it plainly: any hint that AI shopping is quietly steering users toward paid influence confirms the fear that the system answers to the sellers first.

That concern is precisely what the channel framing activates. When merchants think of agentic commerce as a discoverability channel, the natural move is to ask how to rank higher in it. The natural path from "rank higher" is to pay for position or game the signals. That path runs directly into the wall of consumer trust the Harris/Quad data describes. Optimizing for placement in a sponsored-ranking surface is a strategy that makes the entire medium less credible for everyone participating in it, including the merchants doing the optimizing.

---

What the correct frame looks like

Bessemer Venture Partners' agentic commerce analysis names what is missing from the channel frame: they call it agent-legibility, and they identify the intelligence tier where it lives as the least-built in the entire agentic stack. Their framing: "The reasoning, context, and brand truth that sits between catalog and agent. This is the least-built layer and where significant new value will be created. No major platform has filled it. Third-party entrants are early."

Agent-legibility is a property of the merchant's catalog, specifically whether it carries the authored reasoning a shopping engine needs to recommend accurately. A highly structured product feed answers every schema field. It is fully legible to a data parser and almost entirely opaque to a reasoning model deciding whether a product fits a specific shopper in a specific situation. That schema tells the engine what you have. Legibility tells it when yours is the right answer.

Feeds structurally cannot carry five dimensions: fit reasoning about who a product serves and when; substitution logic for what the system should recommend when a variant is unavailable; policy details that differentiate a merchant substantively in the comparison layer; routing guidance about which product belongs in front of which intent; and voice, which defines how the brand explains itself and what it declines to say. The brand-truth layer framework details each of these in full.

The distinction matters at the segment level because it reframes the strategic inquiry. The channel frame asks: how do I get my products to appear in more results? The legibility frame asks: when a reasoning system is deciding between products that share a category and a price range, what does mine carry that makes it the right recommendation? The first inquiry is tractable but insufficient. Being present in a catalog and being selected from it are different outcomes entirely. The UCP protocol audit of Shopify-hosted merchants makes this gap concrete: every merchant in the set was reachable; none were carrying the reasoning signals that would make them distinctively recommendable.

---

The segment question the market hasn't asked

Lareina Yee, a McKinsey senior partner, named the urgency clearly: the moment when most retailers will have to grapple with the fact that a significant share of their customers are AI agents is already arriving. Her precise framing: "Before long, nearly all retailers will have to grapple with the fact that a significant percentage of their customers will not be human users but rather AI agents."

The market's response has been predominantly technical. Better structured data. Richer product feeds. Compatible checkout flows. These are necessary preparations, and building an agent-ready catalog is a real and meaningful step. The infrastructure investment is sound. But it answers the infrastructure question, and a different question remains.

The strategic matter Yee and Coggins are pointing at is a segment matter: if a growing share of your customers are algorithmic buyers, what do you know about how they decide? What does the machine buyer weight? What does it penalize? What does it need to generate a confident recommendation rather than returning an incomplete result or defaulting to price alone?

The HBR evidence is clear on this. Algorithmic decision logic differs fundamentally from human decision logic. The two processes are not the same mechanism with different inputs. They reason differently, respond differently to promotional claims, and the direction of model improvement runs toward greater skepticism of marketing tactics rather than less. Applying a human-persuasion playbook to machine buyers is a mismatch between the tool and the context.

What machine buyers need is reasoning material. Merchant-authored context that sits above the protocol layer and below the session, something that travels with the catalog through UCP and ACP and any subsequent standard, and gives the system a richer substrate than feed fields written for search crawl. The protocol carries the products. The merchant authors what makes them recommendable.

This is the intelligence layer BVP identified as least-built. No platform will build it on the merchant's behalf, because its content is specific to each merchant's products, customers, and brand reasoning. It requires authorship, not enrichment. That distinction is precisely what the channel frame misses. Encoding that authored reasoning is what Sartorial does at the product level, across every SKU, in a versioned artifact shaped for a reader that arrives before any click.

The machine buyers are already here. The merchants who treat that segment as a segment, with its own decision logic and its own requirements, are the ones those systems will recommend. The rest will be indexed and invisible to the reasoning.

---

What to read next

The persuasion-penalty finding from the HBR study is explored in depth in AI shoppers decide differently, which uses the full 16,000-round dataset as its empirical spine. The catalog-side implications of machine decision logic, specifically what reasoning material a shopping system needs during a session, are covered in what AI agents need from your catalog.

The five brand-truth dimensions that constitute agent-legibility are explained in the brand-truth layer piece and applied to a single product page in the 500-word PDP walkthrough. If your starting point is your existing product copy, rewriting descriptions as reasoning material is the practical entry.

---

Sumit Jagdale is the founder of Sartorial.