When AI Becomes the Buyer: What Agentic Commerce Means for SMBs

From AI Advisor to AI Shopper

A few months ago, I wrote about how conversational AI chat ads would reshape the customer journey by monetising the middle of the funnel – the moment when customers compare, question, and decide.

With the rise of agentic AI, we are now moving from a world where AI advises the buyer to a world where AI is the buyer.

Autonomous shopping agents built on top of models such as ChatGPT, Gemini, and Claude are already being used to search catalogues, compare products, and complete checkouts on behalf of consumers. Their growing role at different stages of the customer journey – especially at the buying stage – forces a new question: what happens when the entity making the final buying decision is no longer human?

A recent working paper by Allouah and colleagues, What Is Your AI Agent Buying? Evaluation, Biases, Model Dependence, Emerging Implications for Agentic E‑Commerce, provides the first rigorous look at how AI shopping agents behave in controlled, randomised experiments.

In my December 2025 article, the 3C Framework (Content, Conversation, Commerce) already pointed in this direction: the Commerce pillar argued that product and pricing data needed to become structured and machine‑readable so AI assistants could use them. This new research lets us go deeper and see what specifically to optimise in the Commerce pillar, how strongly those changes can shift AI agent choices, and which new risks come with it.

The Core Idea: AI Agents Are Biased, Not Fully Rational Buyers

We might assume that an AI shopping agent is a perfectly rational economic optimiser – a tireless analyst that always picks the cheapest, best‑rated, most‑reviewed product every time.

It isn’t.

Using a provider‑agnostic sandbox called ACES (Agentic e‑CommercE Simulator), the researchers ran thousands of randomised controlled trials across eight product categories and multiple frontier models. They causally isolated how AI agents respond to attributes like price, ratings, reviews, badges, and position on the page. Three findings stand out in their study:

1. AI agents have strong, model‑specific position biases. Agents systematically favour products based on where they appear in the grid or ranking, and these biases persist even when the agent is explicitly told to ignore position (including in headless, JSON‑only environments without any visual rendering).
2. AI agents reward and punish platform signals in distinctive ways. They consistently discount products tagged "Sponsored" and strongly reward organic endorsements like an "Overall Pick" badge. These effects hold up in randomised experiments even when position and attributes are held constant, this means they're causal, not coincidental.
3. AI demand is highly volatile. A model update like moving from one version of GPT to the next can significantly alter these biases and materially reshuffle which products agents prefer, even when the underlying product catalogue is unchanged.

The key insight is that AI agents are not just a faster, more rational version of your existing customer. They are a new type of decision‑maker, with their own systematic preferences, blind spots, and triggers that only partially overlap with human shoppers.

Why This Matters Now for Marketers

If AI agents play a growing role at the buying stage of the journey, three consequences follow for marketers and SMB leaders.

1. A new “winner‑take‑more” risk for niche brands

Human consumers have dispersed, idiosyncratic preferences. AI agents less so. In ACES experiments, AI agents often collapse demand onto a small set of “modal” products (the one or two an agent keeps picking) in each category, leaving some products and brands effectively never chosen.

For large brands with broad presence, that can be good news. For niche and SMB brands, it’s a structural risk: if a dominant AI agent doesn’t surface you, a disproportionate share of referred demand simply never sees your offer, even when you would have been competitive in a human‑driven market.

2. Demand becomes exposed to exogenous model shocks

In traditional e‑commerce, demand shocks come from competitors, seasonality, or shifts in consumer taste. In agentic commerce, a new kind of shock emerges: the model update.

When the AI agent your customers rely on is upgraded, its choice patterns can shift sharply. Products that were previously the modal choice may lose that position in the new version, while others gain share, with no change at all on your side. This is not a hypothetical effect: across model generations in the ACES experiments, selection shares for the same catalogue moved by tens of percentage points purely because the underlying model changed (Allouah et al., 2025).

3. The format of "being chosen" is changing

In the human‑driven funnel, being chosen meant winning attention, clicks, and conversions. In the agentic funnel, being chosen means being parsed, ranked, and selected by an algorithm that favours structured information, clear feature descriptions, and platform‑level endorsements over persuasive copy or emotional appeals.

In this new agentic commerce environment, it is no longer enough to write a great product page for humans. You must also write one that an AI agent can confidently parse, compare, and use to justify its choice – exactly what the Commerce pillar of the 3C Framework was designed to address.

Real-Life Application: The Agentic Readiness Audit

So how do you prepare without falling into hype‑driven over‑engineering?

When I work with clients on this, I focus on three practical aspects that operationalise the Commerce pillar of the 3C Framework. Think of it as a lightweight Agentic Readiness Audit, grounded in the most recent evidence.

The Agentic Readiness Audit

1. Product data: Catalogue is structured, complete, and machine‑readable.
2. Listing strategy: Titles and copy are optimised for AI parsing.
3. Platform signals: You actively earn the organic badges and reviews agents trust most.

1) Product data: structured, complete, unambiguous

AI agents do not read your brand story. They evaluate a feature space: price, key specifications, delivery time, warranty, ratings, reviews, and category fit.

Which means SMBs need to:

• Audit core attributes (price, specs, dimensions, compatibility, use case) for accuracy and completeness across every channel.
• Use structured data and schema markup so AI agents can confidently extract and compare your products against alternatives.
• Eliminate ambiguity in titles and descriptions. Vague phrasing that humans interpret generously will be penalised by an AI parser.

2) Listing strategy: from SEO to GEO

The ACES experiments show that simple text edits, particularly front‑loading the right keywords in product titles, can drive significant, rapid shifts in selection share among AI agents, sometimes lifting a product from near‑zero visibility to category leader.

For example, the product description of an office lamp that begins its title with “Office Desk Lamp…” will be selected far more often when an AI is shopping for office supplies than the same lamp with a purely feature‑led title.

This is the SMB version of GEO (Generative Engine Optimisation):

• Front-load the most query-relevant terms in titles and first descriptions.
• Match the language of likely AI queries, not just human search habits: study how assistants currently describe and recommend products in your category, and mirror that phrasing in titles and first sentences.
• Add concrete trust signals in copy: statistics, named sources, expert quotes, and specific outcomes. Early experiments with LLM‑based evaluators suggest these evidence‑style elements increase the likelihood of being selected or cited.

3) Platform signals: chase endorsements, not just ads

Here is the most counter‑intuitive finding from ACES: AI agents penalise products tagged as “Sponsored” and reward organic endorsements like “Overall Pick”, even when product attributes and position are held constant in randomised experiments.

For SMBs with limited budgets, this is good news. It means that the path to AI visibility runs less through paid placement and more through:

• Earning organic platform endorsements (best‑seller badges, editor’s picks, category leaders).
• Accumulating high‑quality reviews with substantive text, not just star ratings.
• Maintaining strong fundamentals (fulfilment, returns, customer service) that platforms use to assign those organic signals.

The key message remains clear: organic credibility now compounds twice – once with humans, and once with the AI agents acting on their behalf.

Conclusion: From Customer Journey to Agent Journey

The shift from AI as an advisor to AI as a buyer is likely to be one of the most consequential changes in digital marketing in the last decade. It won't replace human customers – in one 2026 Omnisend study, 80% of U.S. shoppers were open to AI handling checkouts, but only 29% were mostly or fully comfortable letting AI complete purchases independently.

Even so, agentic commerce already inserts a new, non‑human decision‑maker at critical points in the journey, especially for low‑consideration, replenishment‑style purchases.

The 3C Framework hasn’t aged out; if anything, it has become more relevant. What changes is the depth at which we can talk about each pillar. This article deepens Commerce, the buying stage, using fresh empirical grounding. Future articles will keep pulling on the remaining parts of the journey as research catches up.

The enduring lesson is the same one I keep returning to: frameworks only matter when they help you act. You don’t need a perfect agentic strategy before you start. You need a small set of experiments.

If you only do three things in the next 30 days, do these:

1. Run a product listing upgrade. Pick your ten most important SKUs and rewrite their titles and first descriptions with AI parsing in mind: query‑relevant keywords first, clear specs, no fluff. Track any changes in human metrics (CTR, conversion) and in AI‑driven visibility when you run your quarterly assistant tests.
2. Map and prioritise your platform signals. Identify which marketplaces and platforms your category lives on, list the organic badges and endorsements that exist there (best seller, “overall pick”, editor’s choice, category leader), and rank them by how much weight AI agents are likely to give them. As a rule, organic, editorial endorsements tend to count for more than paid placements. Then pick one or two badges and define concrete steps to earn them.
3. Build a lightweight AI‑shopping audit ritual. Once a quarter, ask the main AI assistants (ChatGPT, Gemini, Claude, Perplexity) to recommend the best product in your category for a specific use case. Record whether you appear, where you appear on their shortlists, and what reasoning they give. Track this over time as your “agent visibility log” – your SMB‑scale analogue of the algorithmic auditing ACES provides for platforms.

Do that, and you’ll be building resilience into your customer journey long before agentic commerce is fully democratised.

Thanks for reading

Marc

If you found this useful, you can subscribe to the Monthly Marketing Insight Newsletter where I break down the latest research in marketing and translate it into practical, evidence‑based frameworks for SMBs.

I’d love to hear from you:

• Have you started seeing AI agents influence your category yet?
• Which of the three pillars (product data, listing strategy, platform signals) feels most urgent for your business?

Share your thoughts in the comments👇