When Your Customer Is a Machine

Marketing to AI Agents in the Age of Agentic Commerce

How autonomous buying agents are rewiring discovery, evaluation, and purchase — and what it means for every brand that sells anything to anyone.

The Two Revolutions Happening Simultaneously

For thirty years, digital marketing operated on a single foundational assumption: a human being would see, evaluate, and act on a brand’s message. Every framework — from the sales funnel to programmatic display to search engine optimization — was designed around a person sitting in front of a screen, making decisions shaped by emotion, social proof, and cognitive shortcuts.

That assumption is now dissolving on two fronts at once. The first disruption is already well underway: conversational AI platforms like ChatGPT, Perplexity, Gemini, and others are replacing traditional search as the primary mechanism for product discovery. Users no longer scroll through ten blue links; they receive synthesized answers. The second disruption is just beginning, and it is far more consequential: AI agents are starting to make purchasing decisions on behalf of humans, evaluating options and executing transactions with minimal human oversight.

A February 2026 study from Wharton, published in Harvard Business Review, framed these as concurrent revolutions that will fundamentally reshape how companies compete for customers. One is about how consumers search for information. The other is about who — or what — makes purchasing decisions. Together, they represent the most structural shift in marketing since the invention of the search engine.

Marketing has always been a discipline built on understanding human decision-making — heuristics, biases, emotional triggers. But when the decision-maker is an algorithm with its own evaluation logic, the entire persuasion model breaks down. You can’t anchor a machine. You can’t create urgency for a system that doesn’t experience time pressure. This isn’t an optimization problem. It’s a foundational rethink.

What AI Agents Actually Are — And Why They’re Not Chatbots

The term “agent” has been applied loosely across the technology landscape, but in the context of commerce, it carries a specific and technically meaningful definition. An AI shopping agent is a system capable of understanding a consumer’s preferences, constraints, and intent; searching across multiple brands, platforms, and marketplaces; evaluating products using structured and unstructured data; and making decisions or recommendations with minimal human involvement.

This is not a chatbot that responds to prompts. It is an autonomous system that initiates action. The distinction matters enormously. A chatbot waits for instructions. An agent observes context, forms a plan, executes against objectives, and learns from outcomes. In Forrester’s framing, agentic AI combines memory, reasoning, and tool use to act semi- or fully autonomously across the shopping journey.

The infrastructure is maturing quickly. Perplexity launched a shopping experience with conversational product discovery and instant checkout powered by PayPal. OpenAI introduced shopping research capabilities in ChatGPT using reinforcement learning for comparative product analysis. Google’s shopping graph now powers AI-driven purchase recommendations across its ecosystem. These are not experimental features. They are production systems processing millions of transactions.

How an Agent Actually Evaluates a Product

Understanding what happens inside these systems is essential for anyone trying to optimize for them. A shopping agent operates on a retrieval-augmented generation (RAG) pipeline — a multi-stage architecture that separates knowledge retrieval from response generation.

Here’s the simplified sequence. When a user says “find me a mid-range espresso machine that’s easy to clean,” the agent first decomposes that query into an embedding — a high-dimensional vector representation of the semantic intent. That embedding is compared against a pre-indexed corpus of product data using cosine similarity or approximate nearest neighbor search. The retrieval stage pulls the top-k candidate documents: product descriptions, spec sheets, review aggregations, comparison articles, and structured catalog entries.

But retrieval alone doesn’t determine the final recommendation. A re-ranking layer scores each retrieved document on relevance, recency, source authority, and information density. This is where structured data creates an enormous advantage. A product page with clean JSON-LD schema — explicitly declaring price, dimensions, material, compatibility, warranty terms, and average review score — gives the re-ranker dense, unambiguous signal. A product page that buries the same information inside lifestyle photography, JavaScript-rendered tabs, or PDF spec sheets is partially or entirely invisible at this stage.

After re-ranking, the top candidates are passed to the language model as context for generation. The model synthesizes a response, but it’s constrained by what the retrieval layer surfaced. If your product wasn’t retrieved — because your data wasn’t structured for embedding, or your entity signals were inconsistent across sources — it doesn’t matter how good your product is. You were never in the candidate set.

The critical insight for marketers: the agent doesn’t read your brand story. It parses your schema markup. It doesn’t feel your brand’s emotional resonance. It measures whether your structured attributes match the user’s decomposed intent vector. This is a fundamentally different optimization surface than anything traditional marketing was built to address.

The Data: Consumer Adoption Is Accelerating Faster Than Expected

The shift toward AI-mediated purchasing is no longer theoretical. A January 2026 IBM and National Retail Federation global study surveying more than 18,000 consumers across 23 countries found that nearly half of all shoppers now turn to AI during their buying journeys. Forty-one percent use AI to research products, thirty-three percent use it to interpret reviews, and thirty-one percent use it to find deals.

Net Conversion’s March 2026 consumer research, surveying 900 U.S. adults, found that sixty-four percent of consumers now use AI tools to research purchases, up from forty-five percent just one year earlier. Perhaps more significant: fifty-seven percent of those consumers report that AI has increased the number of sources they consult, suggesting that these tools are expanding rather than narrowing the consideration set.

On the B2B side, the trajectory is even steeper. Eighty-four percent of B2B buyers using AI tools report faster research and decision-making cycles. Among Gen Z professionals, eighty-six percent use AI daily at work, primarily for product research. Forrester projects that ninety percent of B2B purchasing will be AI agent-intermediated by 2028, driving more than fifteen trillion dollars in spend through agent-to-agent exchanges.

The velocity of this adoption curve should alarm any marketer who hasn’t started adapting. We went from forty-five to sixty-four percent AI-assisted purchase research in twelve months. That’s not incremental adoption — that’s a phase transition. The B2B projections suggest we’re approaching a world where the majority of commercial transactions have an AI intermediary somewhere in the decision chain. The brands that aren’t legible to these systems simply won’t be considered.

From SEO to GEO: The Technical Infrastructure of AI Visibility

The practical consequence of this shift has given rise to an entirely new discipline: Generative Engine Optimization, or GEO. First formalized by researchers at Princeton and Georgia Tech, GEO is the practice of structuring content and digital presence so that AI-powered platforms can retrieve, cite, and recommend a brand when answering user queries.

If traditional SEO aimed to rank on page one of Google, GEO aims to become part of the AI-generated answer itself. The difference is not cosmetic. AI systems that use real-time retrieval evaluate a page’s relevance primarily on its opening content, entity clarity, structured data consistency, and third-party citation authority. The competitive set is no longer ten organic links; it is the two to seven sources a large language model typically cites in a single response.

The Four Pillars of GEO

Entity Clarity. AI systems cross-reference signals from multiple sources and formats. A brand’s description on LinkedIn must align with what appears on its website, on Crunchbase, on review platforms, and in industry directories. When these signals are consistent, AI systems categorize and reference the brand with greater confidence. When they conflict, confidence drops, and the brand is less likely to surface in generated responses.

This is technically an entity resolution problem — the same class of challenge that knowledge graph systems solve when disambiguating whether “Apple” refers to the company, the fruit, or the record label. AI models resolve brand entities by triangulating structured signals: your Organization schema on your website, your Google Business Profile attributes, your LinkedIn company page metadata, and your presence in third-party directories and databases. Every inconsistency — a different founding year here, a slightly different service description there — degrades the confidence score the model assigns to your entity node. At scale, this is the difference between being cited and being ignored.

Content Extractability. AI models reward content that is structured for machine parsing rather than human browsing. This means explicit data surfacing — no hidden accordions, no specifications buried in images. JSON-LD schema markup, HTML tables, and structured attribute fields become essential infrastructure.

In practical terms, this means your product pages need to declare attributes the way a database would, not the way a copywriter would. Consider the difference:

A marketing-optimized product description might read: “Our premium espresso machine delivers barista-quality results with an intuitive one-touch interface.” An agent can extract almost nothing actionable from that sentence.

A GEO-optimized version pairs that narrative with structured markup declaring: price: $449, pumpPressure: 15 bar, waterReservoir: 67oz, cleaningSystem: automatic, warrantyMonths: 24, avgRating: 4.6, reviewCount: 2,847. The narrative keeps the human engaged. The schema feeds the machine. You need both — but if forced to choose, the schema is what gets you into the agent's candidate set.

Trust Signal Management. Research from Princeton demonstrates that AI engines strongly favor earned media — authoritative third-party sources — over brand-owned content. Large language models weight user-generated content highest, followed by influencer and media content, with brand-owned content at the bottom of the citation hierarchy. The content a brand controls least has the most influence on AI recommendations.

This is counterintuitive but structurally logical. Language models are trained to assess source reliability, and they’ve internalized a version of the same heuristic that journalists use: independent verification carries more weight than self-reported claims. A brand saying “we’re the best” is noise. A Wirecutter review, a Reddit thread with authentic user feedback, or an industry analyst report saying the same thing is signal. The practical implication is that PR, review generation, and thought leadership publishing — activities many brands have de-prioritized in favor of performance marketing — are now core GEO infrastructure.

Freshness Signals. AI engines weigh recency when selecting sources. A guide published in 2024 with no updates will lose ground to a 2026 article on the same topic, even if the underlying information hasn’t changed. Regular content refresh with updated data, new insights, and clear timestamps is no longer best practice — it is table stakes for AI visibility.

GEO represents the most significant structural change in digital marketing since Google introduced PageRank. But unlike SEO, which evolved over two decades of incremental refinement, GEO is being built in compressed time against a rapidly shifting technical landscape. The brands that treat this as another channel optimization will lose. The ones that understand it as an infrastructure problem — data architecture, entity management, cross-platform consistency — will build durable competitive advantage.

Marketing to Machines: What Changes When the Buyer Has No Emotions

The most disruptive implication of agentic commerce is not technological but conceptual. Marketing, as a discipline, was built on a theory of human persuasion. Every framework assumes an audience susceptible to framing effects, social proof, scarcity, emotional resonance, and narrative. When the buyer is a software agent executing against a preference function, these levers become irrelevant.

AI agents do not experience brand affinity. They do not respond to urgency. They are not influenced by celebrity endorsements or aspirational imagery. They evaluate products against structured criteria: specifications, pricing, availability, review sentiment, return policies, and compatibility with the user’s existing preference graph. Emotional language and marketing copy still matter for the humans who set agent preferences, but the agent itself prioritizes machine-readable information it can evaluate and compare.

The competitive advantage shifts from persuasion to legibility. Can the machine read your value proposition? Can it verify your claims? If not, you don’t exist in its decision space.

This creates a two-audience problem that has no precedent in marketing history. Brands must simultaneously optimize for human emotional engagement (to be chosen as a preference input) and machine-readable precision (to survive the agent’s evaluation logic). The creative brief and the schema markup are now equally important documents.

This is where economic thinking becomes essential. An AI agent is essentially a rational actor in the classical sense — maximizing utility against a defined objective function with near-complete information access. Marketing to a rational actor with perfect information is a fundamentally different problem than marketing to a bounded-rational human with selective attention.

Consider what this means in practice. Traditional pricing strategy relies heavily on anchoring, decoy effects, and context-dependent framing — all of which exploit cognitive biases that agents don’t have. A human shopper might perceive a $200 product as a good deal when it sits next to a $400 option. An agent with access to the full price distribution across every retailer will evaluate that $200 price against the true market clearing price, historical price trajectories, and predicted future price movements. Anchoring doesn’t work when the buyer has already computed the efficient frontier.

The same principle extends to product differentiation. Humans can be persuaded that two functionally identical products are meaningfully different through branding, packaging, and narrative. An agent evaluating structured attributes will see identical specs and route to the lower-cost option unless the differentiation is reflected in measurable attributes — better warranty terms, higher review scores, faster shipping — that the agent’s objective function can weigh. Brands that have relied on perceived differentiation rather than structural differentiation are exposed in an agent-mediated market.

The Trust Architecture: Ads Inside the Oracle

The trust dynamics of this new ecosystem became sharply visible in February 2026 when OpenAI began testing advertisements inside ChatGPT. The ads appear at the bottom of conversational responses for free-tier users, matched to the topic of the user’s conversation. OpenAI states clearly that ads do not influence the answers ChatGPT provides — a principle the company calls “Answer Independence.”

The introduction of advertising into conversational AI mirrors historical precedent: radio, the early web, and social media each evolved from trusted information spaces into monetized platforms. The critical difference is that conversational AI occupies a uniquely intimate position in the user’s decision-making process. Users treat these systems more like advisors than media channels. Research suggests that user trust in AI-generated content drops sharply when users suspect commercial influence, making the separation between organic answers and paid placement not just an ethical imperative but a structural requirement for the platform’s long-term viability.

For marketers, this creates a new strategic layer. Gartner forecasts that traditional search engine volume will decline by twenty-five percent by 2026 as AI chatbots replace conventional queries. The brands that build organic authority within AI recommendation systems will compound advantage over time. Those that rely solely on paid placement within these systems face the same trust erosion that degraded the value of sponsored search results a decade ago.

The advertising question inside AI platforms is actually a market design problem — a classic two-sided market where the platform’s value to users depends on perceived neutrality, but the platform’s revenue model depends on advertiser participation. Classical economics tells us these incentives will eventually conflict. This isn’t speculation; it’s the same dynamic that degraded the informational value of Google Search results over the past decade as paid placements expanded and organic results compressed.

The strategic hedge is straightforward: build organic citation authority now, while the AI recommendation layer still operates primarily on merit-based retrieval rather than pay-to-play placement. The brands that establish strong entity signals, deep third-party citation networks, and high-quality structured data will have a defensible position regardless of how the advertising layer evolves. Those that wait for the ad platform to mature before engaging are making the same mistake brands made in 2005 when they ignored organic SEO in favor of waiting for Google Ads to “figure itself out.”

The Measurement Gap: Why Traditional Marketing Metrics Break

The shift to AI-mediated commerce exposes a fundamental inadequacy in existing marketing measurement frameworks. Traditional metrics — click-through rates, page views, conversion attribution, cost per acquisition — were designed for a world where humans navigated linear paths from impression to purchase. In an agent-mediated world, there may be no click, no page view, and no attributable touchpoint.

When an AI agent evaluates a product on behalf of a consumer, the “impression” occurs inside a retrieval pipeline that the brand cannot observe. The “evaluation” happens across a synthesized comparison that the brand did not create. The “conversion” may execute through an API transaction that bypasses the brand’s owned digital properties entirely. The entire purchase funnel — from awareness through consideration to transaction — can occur without the consumer or the brand touching a browser.

Emerging Metrics: What to Track

New measurement frameworks are emerging to address this gap. AI citation frequency — how often a brand appears in AI-generated answers — is becoming a primary visibility metric. Share of voice across AI platforms measures competitive positioning in generated responses. Specialized tools now track brand mentions, citation sources, sentiment, and recommendation rank across ChatGPT, Google AI Overviews, Perplexity, Claude, Gemini, and others. But the discipline is nascent, and Jasper’s 2026 State of AI in Marketing report found that only forty-one percent of marketers can currently prove AI ROI, down from forty-nine percent the prior year.

The Causal Inference Layer: Measuring What You Can’t Directly Observe

This measurement crisis is structurally identical to the attribution crisis that hit digital advertising a decade ago, but more severe. At least with digital ads, you had observable events — impressions, clicks, conversions. With agent-mediated commerce, the decision process is opaque by design. The agent’s evaluation logic is a black box. This is where causal inference methods move from academic niceties to operational necessities.

Consider a concrete scenario. A brand invests in GEO — restructuring product data, building schema markup, improving entity consistency, earning third-party citations. Three months later, revenue from a product category increases. Did GEO drive that? Or was it seasonal demand, a competitor’s stockout, a pricing change, or some combination?

Traditional A/B testing can’t answer this cleanly because you can’t randomize which AI models see your structured data. Instead, you need quasi-experimental methods designed for observational settings.

Synthetic control methods offer one approach. You construct a weighted combination of comparable product categories or brand segments that didn’t receive the GEO treatment, creating a counterfactual “synthetic” version of the treated unit. The difference between the synthetic control’s trajectory and the treated unit’s actual performance post-intervention is your estimated causal effect. This is the same methodology economists use to evaluate policy interventions where randomization isn’t possible.

Bayesian structural time series (BSTS) models provide another. A BSTS model builds a probabilistic forecast of what would have happened absent the GEO intervention by fitting a state-space model to pre-intervention data, incorporating seasonality, trend, and regression components from control series. The posterior distribution of the counterfactual gives you not just a point estimate of impact but a credible interval — a probabilistic range that accounts for model uncertainty. Google’s CausalImpact package in R implements this approach and is accessible enough for most analytics teams.

Difference-in-differences works when you have geographic or segment-level variation in rollout timing. If you implemented GEO for one product line in Q1 and another in Q2, you can use the staggered rollout as a natural experiment, comparing treated and untreated units before and after intervention while controlling for parallel trends.

None of these methods are new. But in the context of agent-mediated commerce — where the treatment effect is obscured by an opaque intermediary layer — they shift from “advanced analytics” to baseline measurement infrastructure. Organizations that lack causal inference capability will be unable to distinguish genuine GEO impact from noise, and will inevitably either over-invest or under-invest as a result.

The Organizational Reckoning: Humans Supervise, Agents Operate

The operational implications of agentic commerce extend deep into organizational structure. Adweek’s March 2026 analysis observes that AI is eroding the middle layers of marketing faster than most leaders acknowledge, manifesting not as mass layoffs but as role confusion, eroding confidence, and quiet disengagement among product marketers, strategists, creatives, media planners, and analysts.

The emerging operating model is one where humans supervise and agents operate. Marketing organizations are increasingly moving toward pod-based execution models that bring strategy, creative, analytics, and technical execution into single units, reducing the handoff friction that traditional departmental structures create. Strategy and execution operate in parallel rather than sequence. Campaigns evolve during delivery rather than in post-campaign review.

Jasper’s research found that sixty-five percent of marketing teams now have designated AI roles, often focused on AI operations, workflows, or strategy. Ninety-one percent of marketing teams use AI, up from sixty-three percent the prior year. But a stark confidence gap has emerged between leadership and frontline execution: sixty-one percent of CMOs report confidence in AI ROI, compared to just twelve percent of individual contributors.

The organizational design challenge here is fundamentally an economic coordination problem — principal-agent misalignment where leadership sets AI objectives that frontline teams don’t have the training, tooling, or role clarity to execute against. This isn’t a communication problem that gets fixed with better town halls. It’s a structural problem rooted in misaligned incentive systems. When a CMO’s performance is measured on AI adoption metrics and an individual contributor’s performance is measured on campaign throughput, the objectives are directionally contradictory. The CMO wants transformation; the IC needs to ship work.

The organizations that will thrive are those redesigning incentive structures, decision rights, and workflow architecture simultaneously. You cannot bolt agent-mediated marketing onto a departmental org chart built for the campaign era. The pod model works because it collapses the feedback loop — the person interpreting agent-driven performance data sits next to the person adjusting the schema markup, who sits next to the person adapting the creative. When your marketing operates inside a system where AI agents are constantly re-evaluating your brand, the speed of that internal feedback loop becomes a direct competitive variable.

What to Do Now: A Technical Roadmap for Agent Readiness

The window for proactive positioning in agentic commerce is open but narrowing. Citation authority, like domain authority before it, compounds over time. The brands that invest now will earn compounding advantages as AI becomes the primary channel for discovery and purchase. The following technical priorities represent the minimum viable infrastructure for agent readiness.

Audit your machine readability. Run your product pages, service descriptions, and key content through an AI crawler analysis. Tools like Scrunch and Profound now offer agent-perspective audits that show how AI crawlers actually navigate and interpret your site — which pages they hit, which they skip, and where your content is unreadable due to JavaScript rendering, image-embedded text, or missing structured markup. If specifications live inside images, if pricing is buried in JavaScript-rendered accordions, or if structured data markup is incomplete, AI agents cannot evaluate your offering. Surface all data explicitly using JSON-LD schema, HTML tables, and structured attribute fields.

Build entity consistency across platforms. Your brand description on LinkedIn, your Google Business Profile, your Crunchbase listing, your industry directory entries, and your owned website must present consistent identity signals. AI systems cross-reference these sources. Think of it as entity resolution hygiene — the same discipline that data engineers apply to customer record deduplication, applied to your brand’s distributed identity across the web. Inconsistency lowers citation confidence.

Invest in earned media and third-party validation. AI models weight independent sources far more heavily than brand-owned content. Customer reviews, industry analyst citations, media coverage, and expert endorsements are the primary trust signals that drive AI recommendations. This is where the economics of GEO diverge sharply from traditional SEO. In SEO, you could build authority primarily through owned content and link building. In GEO, the highest-weight signals come from sources you don’t control — which means PR strategy, review generation programs, and analyst relations are no longer “nice to have” brand activities. They’re core technical infrastructure for AI visibility.

Develop a GEO measurement stack. Begin tracking AI citation frequency, share of voice across AI platforms, and the specific prompts that trigger competitor mentions. Build a prompt universe of fifty to one hundred fifty queries segmented by intent and funnel stage. Monitor these systematically across ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews. Establish a causal inference baseline — whether through synthetic controls, BSTS models, or staggered rollout designs — so that when you invest in GEO improvements, you can isolate the impact from background noise.

Redesign content for dual audiences. Every piece of content must serve both the human who sets preferences and the agent that evaluates options. This means combining narrative storytelling with structured, machine-parsable data. The creative brief and the schema markup are now companion documents. Practically, this requires content teams and data engineering teams to collaborate on every major page — a workflow that almost no marketing organization has today.

Prepare for agent-to-agent commerce. Ensure your product data is available through APIs and structured feeds, not just crawlable web pages. As agent ecosystems mature, brands that offer direct data integration through standardized APIs and product data feeds will have a structural advantage over those that rely solely on web-crawlable presence. This is especially critical in B2B, where Forrester’s projections suggest agent-to-agent transactions will become the dominant purchasing modality within two years.

Think of AI agents as a new market participant with distinct information processing characteristics. You wouldn’t market to institutional investors the same way you market to retail consumers, even though both buy the same securities.

Agent-mediated commerce requires the same kind of audience-specific strategy — different value propositions, different content formats, different success metrics. The brands that build this capability now are building structural competitive advantage that compounds quarterly.

The Bottom Line

The most consequential shift in the history of commercial marketing is not happening in the future. It is happening now, in production systems processing real transactions at scale. AI agents are not a speculative technology or an innovation-cycle talking point. They are active market participants that are already influencing how products are discovered, evaluated, and purchased across both consumer and enterprise contexts.

The brands that recognize this moment for what it is — a structural transformation of market architecture, not an incremental channel addition — will build the infrastructure, measurement systems, and organizational capabilities required to compete. Those that treat agentic commerce as another line item in a digital strategy deck will find themselves progressively invisible to the systems making an increasing share of purchasing decisions.

In 2026, one of your most important customers may not be a person at all. The question is whether your brand is ready to be evaluated by a machine that doesn’t care about your logo, your tagline, or your brand story — but cares very much about whether your data is clean, your claims are verifiable, and your value proposition is machine-readable.

Brian James Curry is an Enterprise AI Strategist, Marketing Economist, and Applied Researcher. He is the founder of Vector1 Research, an independent research lab focused on marketing economics, causal systems, cognitive AI architecture, and the economic implications of intelligent systems. He also founded KC AI Lab, a data science community and incubator for applied machine learning education.

Sources: Harvard Business Review (Puntoni, Feb 2026), IBM/NRF Global Consumer Study (Jan 2026), Net Conversion Consumer Survey (Mar 2026), Jasper State of AI in Marketing 2026, Forrester Research, Gartner Marketing Predictions 2026, Adweek, Princeton/Georgia Tech GEO Research, Search Engine Land, OpenAI, and BCG.