For a century, marketing has been the art and science of influencing human decisions. Every framework, every theory, and every billion-pound media plan has assumed a person at the end of the funnel: someone who sees, feels, remembers, desires, and chooses. Attention, emotion, memory, trust. Those are the currencies the discipline was built to trade in.
That assumption is breaking.
AI agents now make purchasing decisions on behalf of humans. An algorithm compares your product with seventeen alternatives in milliseconds, weighing price, ratings, delivery speed, and live inventory. Your brand’s century of emotional storytelling does not register. The agent feels no loyalty. It has no memory of your Super Bowl advert. It reads structured data, applies its principal’s preferences, and transacts.
What makes this concrete rather than theoretical is that the infrastructure is already running. OpenAI’s Agentic Commerce Protocol (ACP) is an open standard for agent-to-merchant communication covering product discovery, checkout, and fulfilment. Google’s Agent Payments Protocol (AP2) does the equivalent job for agent-facilitated transactions. When the two largest AI companies in the world both publish commerce protocols for machines, the signal is not subtle. This is not a white paper. It is shipping code.
McKinsey projects that agentic commerce will generate three to five trillion dollars globally by 2030, with agents researching, negotiating, and completing purchases often without direct human intervention. The infrastructure for automated agent-to-consumer and agent-to-merchant transactions is live now. Production, not preparation.
What follows is an examination of what happens to marketing when the customer is a machine, and of the dual-track operating model that meets the new reality.
The Scale of Agentic Commerce
McKinsey’s three to five trillion dollars by 2030 rests on three capabilities agents already have: autonomy (acting without constant input), reasoning (adapting to changing conditions), and interoperability (working across platforms via APIs).
The traffic data is the part that made me sit up. Adobe Analytics reports that traffic to US retail sites from generative-AI browsers and chat services rose 4,700 per cent year on year in July 2025. Users arriving through AI agents spend 32 per cent more time on site, browse 10 per cent more pages, and bounce 27 per cent less often than those arriving through traditional channels.
Treat the headline figure with the caution any four-figure percentage deserves. It is growing from a small base, and AI-referred traffic remains a fraction of total retail visits. Growth of that magnitude still represents the early slope of an adoption curve, and it points to customers who are demonstrably more engaged than the ones arriving by the usual routes.
Gartner sharpens the picture. By 2026, 30 per cent of large companies will have a dedicated business unit or sales channel to reach machine-customer markets. By 2028, 60 per cent of brands will use agentic AI to deliver one-to-one interactions, and 90 per cent of B2B buying will be intermediated by AI agents, pushing $15 trillion through agent exchanges.
The channel architecture is being redefined now. The share of revenue actually routed through agentic shopping tools will move more slowly. Marketers who ignore agentic commerce will be blindsided; marketers who abandon human-focused strategies to chase it will be premature. The correct posture is preparation without panic: build the capabilities and the infrastructure now, so the organisation is ready when the inflection point arrives.
Discovery, Evaluation, Transaction: Each Rewritten
Marketing has always operated across three layers: discovery, evaluation, and transaction. Every one of them changes when the customer is an AI agent.
Discovery: from SEO to GEO. Search engine optimisation has been the discipline of being findable for twenty-five years. Google’s algorithm weighed backlinks, content quality, page speed, and hundreds of other signals, and an industry worth tens of billions grew up around reading those signals.
Generative Engine Optimisation is the next turn: making your brand discoverable, citable, and selectable by AI systems. The mechanisms differ sharply, because the searcher is no longer a human scanning a results page. It is a model constructing an answer from training data, retrieved context, and real-time information.
Each platform weights the signals differently. ChatGPT favours domain reputation and authoritative sourcing. Perplexity rewards verifiable citations and structured data. Gemini inherits much of Google’s traditional ranking apparatus and adds entity-level understanding. Claude prioritises multi-source verification and calibrated confidence.
Research from COSEOM shows that adding statistics, source citations, and structured formatting to content lifts AI visibility by up to 40 per cent. Models that generate answers from retrieved content cite material that is structured, verifiable, and data-rich.
The measurement tools have arrived too. Otterly.ai, Semrush AI Toolkit, and Ahrefs Brand Radar track AI citation frequency: how often AI systems mention your brand when answering a relevant query. GEO is now a standard marketing practice rather than an experimental one.
Evaluation: from emotion to evidence. This is where the inversion is starkest.
A human evaluating a product may be moved by brand heritage, packaging design, a celebrity endorsement, or the memory of an advert from childhood. An agent evaluating the same product prioritises price, ratings, delivery speed, and live inventory over brand familiarity or loyalty.
Brand does not become irrelevant. It becomes a proxy signal. Agents trained on human-generated reviews carry the brand preferences embedded in that data, and a brand with consistently high ratings across multiple platforms will be recommended more often than an unknown competitor, all else being equal. What has changed is the mechanism of influence. You are no longer persuading a person to feel something; you are making sure the data about your product is thorough, accurate, and favourable across every source an agent consults.
Four requirements follow for machine-readable evaluation.
Structured product data. Machine-readable specifications in standard formats: schema.org markup, product feeds, structured APIs. If an agent cannot parse your product attributes programmatically, you do not exist in its evaluation set.
Real-time pricing and availability. Agents operate in milliseconds. If your inventory system takes thirty seconds to answer, the agent has already gone. API response time becomes a competitive factor.
Verifiable claims. Agents weight verified third-party data (reviews, independent tests, certifications) above self-reported marketing claims. Trust but verify, enforced algorithmically.
Comparative accessibility. If your product data sits inside a walled garden that requires a human to browse it, agents will compare you unfavourably with competitors whose data is openly accessible. Transparency becomes a competitive advantage.
Transaction: from checkout to protocol. In traditional commerce the transaction happens on your property: your website, your app, your store. You control the experience, capture the data, and own the relationship.
In agentic commerce the transaction may happen through a protocol. The agent talks to your commerce API, never visits your website, and never sees the brand experience you spent a year building. The product is discovered, evaluated, and purchased without a single human impression of your brand.
The risk here is real: less direct access to customers, weaker brand loyalty, and growing dependence on intermediary platforms. Refusing to participate in ACP or AP2 would be like refusing to list on Google in 2005. Technically possible, strategically suicidal.
How AI Agents Actually Evaluate Products
An agent operating under agentic commerce protocols usually follows a structured evaluation sequence.
Step 1: constraint filtering. The agent removes products that fall outside its principal’s stated requirements: budget, category, specifications, delivery timeline. Any hard constraint cuts the candidate set immediately, and products without machine-readable constraint data are excluded by default.
Step 2: feature matching. Among the survivors, the agent scores products against the principal’s preferences. Structured data matters most here. If the principal prefers lightweight running shoes, the agent has to find and compare weight data across products. A brand that publishes weight in machine-readable form gets evaluated. A brand that buries it in unstructured description text may not.
Step 3: social proof integration. The agent consults aggregated review data, expert evaluations, and third-party testing results: social proof weighted algorithmically, through rating distributions, review volume, sentiment analysis, and source credibility. Products with thin review profiles are disadvantaged whatever their quality.
Step 4: availability and logistics. Real-time inventory, delivery options, and fulfilment reliability enter the final evaluation. An agent optimising for speed will penalise a product that takes seven days to arrive, even when it scores higher on other dimensions. API performance has a direct effect on marketing outcomes.
Step 5: comparative ranking. The agent produces an ordered list from the weighted evaluation, then either presents the top three for human review or buys the highest-ranked option outright, depending on how the consumer has configured it. Conversion happens here, and every factor influencing it is algorithmic.
Read that sequence back and the diminished influence of familiar tactics becomes obvious. A beautifully designed product page does nothing at Step 1. An evocative brand story does nothing at Step 2. A celebrity endorsement may nudge Step 3, and only if it shows up in structured review data.
The new imperative is to perform well at each algorithmic step, which is a categorically different problem from performing well in human cognition.
The Dual-Track Marketer
What emerges is the professional who works fluently across two entirely different paradigms at once: the dual-track marketer.
Track one is human marketing: brand, emotion, storytelling, creativity, trust, community. Everything the discipline has done for a century, refined and amplified by AI tools but still aimed at human cognition. This track does not disappear. It remains the source of the brand equity that shapes both human purchasing and the AI training data behind algorithmic recommendation.
Track two is machine marketing: structured data, API accessibility, entity authority, GEO, agentic commerce protocols. A new discipline aimed at algorithmic cognition, making the brand discoverable, evaluable, and selectable by AI agents.
The best marketers of 2028 to 2030 will be bilingual, fluent in human persuasion and machine optimisation alike. The two tracks demand different skills, different metrics, different tools, and in many cases different people. Asking a brand storyteller to manage API response times is no more reasonable than asking a data engineer to write emotional advertising copy.
Two genuinely new roles fall out of this. The M2M Marketing Strategist designs marketing strategies for AI agent audiences rather than human ones, making sure product catalogues are machine-readable, API-accessible, and structured for agent evaluation. The Agent Commerce Operations Lead builds and maintains the technical infrastructure for agent-to-brand interaction: APIs, product feeds, structured data, and the agentic commerce protocols themselves.
Three Diagnostics for the Next Channel Strategy Review
The agent is already looking at your catalogue. Find out what it sees.
Audit your product catalogue for machine readability. For your top-revenue SKUs, are all the critical attributes available as structured data through an API that responds in under 500 milliseconds? Most enterprise catalogues fail that test. Fixing it matters more than the next CRM upgrade.
Track your AI citation rate. Use Otterly.ai, Semrush AI Toolkit, or Ahrefs Brand Radar to baseline how often AI systems mention your brand when answering relevant queries. The baseline is rarely flattering. The trajectory is the number that matters.
Identify your first GEO specialist hire. If the role does not exist in your organisation today, you are eighteen months behind. The skill is teachable, the salary is competitive, and the capability gap closes faster with one dedicated hire than with five training programmes.
The infrastructure for agent-to-merchant commerce is already shipping. Marketers who absorb it inside their current planning cycle will own the channel by 2028. Marketers who file it under next year’s roadmap will arrive after the protocols have hardened, the metrics have set, and the leaders have compounded their advantage.
The customer who is an algorithm is not a future. It is the present, growing from a small base, at a slope that ought to change the agenda of your next executive offsite.
Keep Reading
That’s all for this week book chapter summary, come back next Monday for the next chapter summary.
The Agentic CMO - Second Edition is available today in hardcover, paperback and ebook.
Disclaimer: The views and opinions expressed in The Agentic CMO, Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.