Quick answer: Agentic commerce is shopping carried out by an AI agent on a person’s behalf — the agent researches, compares, selects and increasingly pays, without the shopper ever loading your product page. In 2026 it is small but growing faster than any channel in retail history: AI-driven visits to US retail sites rose roughly 4,700% year over year, yet AI sessions are still under 0.2% of total ecommerce traffic. The work retailers need to do now is technical, not marketing: structured product data, an agent-readable catalogue, and a checkout that can accept a delegated, verifiable purchase.
Every few years a channel appears that rewards the retailers who prepared early and punishes the ones who waited for proof. Mobile was one. Marketplaces were another. Agentic commerce is the 2026 version, with one difference: this time the channel does not render your website at all. It reads your data.
This guide is for CTOs, heads of ecommerce and founders in the US and UK. It covers what agentic commerce actually is, what the credible numbers say (including where they contradict each other), the protocol stack you will be asked about in your next board meeting, the seven things that break when an agent shops your store, and the order in which to fix them.
What is agentic commerce?
Agentic commerce is a transaction where an AI agent acts on a shopper’s behalf across the full purchase journey — discovery, comparison, selection, and in some implementations payment and post-purchase service. The shopper states an intent (“replace my running shoes, same model, size 10, under £120, delivered before Friday”) and the agent executes against it.
It sits on a spectrum, and conflating the levels is the most common mistake in boardroom conversations:
- AI-assisted research. The shopper asks ChatGPT, Gemini or Perplexity for recommendations, then clicks through to your site and buys normally. This is the overwhelming majority of agentic activity today.
- Agent-supported actions. The agent fills the basket, applies the promotion, or completes a return, with the human approving each step.
- Delegated purchase. The agent transacts inside the AI surface, using a delegated payment credential and a spending mandate. Small today, and the part every protocol war is about.
Your readiness work is the same for all three, which is the good news. A catalogue an agent can read accurately is what wins level one, and level one is where the revenue is right now.
What the 2026 numbers actually say
Be careful with agentic commerce statistics. A lot of the figures circulating are aggregator blog posts citing each other. Here is what holds up, with the original source named.
Growth is real and very fast from a very small base
- Adobe Analytics recorded roughly 4,700% year-over-year growth in AI-driven visits to US retail sites, and around 3,300% growth during Prime Day.
- Despite that, AI-driven sessions remain below 0.2% of total ecommerce traffic. Both facts are true at once.
- EMARKETER forecasts US agentic commerce at $20.57bn in 2026 — about 1.5% of total US retail ecommerce.
The 2030 forecasts disagree by an order of magnitude
- Bain & Company: $300–500bn in the US by 2030, or 15–25% of ecommerce sales.
- Morgan Stanley: 10–20% of US ecommerce agent-driven by 2030, adding around $115bn.
- Gartner: 20% of digital commerce transactions executed through AI platforms by 2030.
- J.P. Morgan: up to 25% of US online sales by 2030, concentrated in groceries and subscriptions.
- McKinsey: around $1tn in orchestrated US retail revenue, and $3–5tn globally.
When four credible houses span $300bn to $1tn for the same market and year, the honest read is that nobody knows the size — but everyone agrees on the direction. Plan for the channel to exist, not for a specific number.
Conversion is the genuinely contested figure
You will see “AI traffic converts 42% better” quoted widely. You will also see Adobe data showing AI-referred traffic converting about 23% below non-AI sources, narrowing from a 49% gap earlier in the year. Peer-reviewed work from the University of Hamburg and Frankfurt School (Kaiser & Schulze) places ChatGPT referral conversion above paid social but below direct, organic search and email.
The reconciliation: AI referrals are high-intent but early-journey, and the gap is closing quickly. Walmart’s published experience adds a useful wrinkle — in-chat purchases converted roughly 3x worse than redirects to the site, but ChatGPT traffic brought in about twice as many new customers as traditional search. Treat agentic traffic as an acquisition channel first and a conversion channel second.
The trust gap is the real ceiling
YouGov data puts US consumer trust in AI to compare prices at 65%, but trust in AI to place an order at 14% (rising to 29–30% among Gen Z and Millennials). That roughly 50-point gap, not the technology, is what caps fully autonomous checkout today. Kearney separately finds 60% of shoppers expect to use AI agents within twelve months — expectation is running well ahead of authorization.
The strategic implication is unglamorous and important: optimize for the agent that recommends you, not the agent that buys for you. The first is worth money this quarter.
The agentic commerce protocol stack, explained for CTOs
Five acronyms will come up. They are not competing versions of the same thing; they sit at different layers.
- MCP (Model Context Protocol) — the connection layer between an AI model and your systems. Became a founding project of the Linux Foundation’s agentic AI foundation in December 2025, with more than 10,000 published servers at announcement. This is the one your engineering team will touch first, and it is not retail-specific.
- ACP (Agentic Commerce Protocol) — a commerce-specific specification for how an agent discovers products and places an order with a merchant. Five specification versions shipped between September 2025 and April 2026; expect continued churn.
- UCP (Universal Commerce Protocol) — launched January 2026, co-developed by Google, Shopify, Etsy, Wayfair, Target and Walmart, with 20+ further endorsers. The presence of both a hyperscaler and the major platforms makes this the one most likely to reach your stack via a plugin rather than a build.
- AP2 (Agents to Payments Protocol) — announced by Google Cloud in September 2025 with 60+ collaborating organizations, covering the payment mandate layer: how an agent proves it is authorized to spend a specific amount on a specific person’s behalf.
- x402 and A2A — blockchain-settled agent payments (100m+ cumulative transactions through Q1 2026) and agent-to-agent coordination respectively. Watch, do not build, unless crypto settlement is already in your business model.
The practical CTO position in 2026: do not pick a winner. Build your product, inventory and pricing data into a single clean, API-addressable source of truth, and treat each protocol as an adapter over it. Retailers who hard-code to one specification will rewrite within eighteen months. This is ordinary platform engineering discipline, applied to a new consumer.
7 things that break when an AI agent shops your store
Agents do not experience your site the way humans do. They do not see your hero image, your brand film or your carefully art-directed PDP. They read text and structured data, and they leave the moment they cannot resolve a fact.
1. Your product data is not machine-resolvable
Size, material, compatibility, warranty and dimensions live in marketing prose or in an image, not in fields. An agent asked for “waterproof, under 2kg, fits a 15-inch laptop” cannot verify any of that and will recommend a competitor who published it as data. Rithum’s Commerce Readiness Index found 45% of retailers “sometimes” and 36% “often” face data-quality problems affecting decisions — this is the single largest blocker.
2. You have no structured data, so you are not cited
Pages carrying structured data are cited roughly 3.1x more often in Google AI Overviews, and structured data appears in about 71% of ChatGPT citations and 65% of Google AI Mode citations. Product, Offer, AggregateRating, FAQPage and Organization schema are no longer an SEO nicety; they are the API through which an agent reads your store.
3. Real-time price and stock are not exposed
An agent that recommends an out-of-stock item damages its own user’s trust, so agents systematically down-rank sources they cannot verify. If your availability lives only behind a JavaScript-rendered widget, you are invisible at exactly the moment of decision.
4. Your bot defences block the agents you want
The awkward one. Visa reported a 25% rise in malicious bot-initiated transactions over six months, 40% in the US, and Accenture found 78% of financial institutions expect AI-agent-linked fraud to increase. Blanket bot blocking is now a revenue decision, not just a security one. You need agent identity and allow-listing, not a wall.
5. Checkout assumes a human with a session
CAPTCHAs, multi-step funnels, mandatory account creation and session-bound carts all fail a delegated purchase. Accenture found 87% of CTOs and payment leads believe trust — not capability — is the barrier to agentic payments. Your checkout needs a path that accepts a verifiable mandate with spending limits and instant revocation.
6. Returns and service have no machine interface
Post-purchase is where agentic commerce quietly saves the most money. If an agent can initiate a return, track a shipment or reschedule a delivery through an API, your contact-centre volume falls. If it can only send an email, it does not. This is natural territory for virtual assistant agents wired into real fulfilment systems rather than an FAQ.
7. You cannot attribute any of it
Most analytics stacks classify AI referrals as direct traffic or lump them into “other.” You cannot justify investment in a channel you cannot measure. Segmenting AI referrers is a one-sprint job and should be done before anything else on this list, so that every subsequent change has a baseline. Pair it with proper analytics dashboards so the trend is visible to the people funding the work.
The readiness order: what to do, in what sequence
Sequenced so that each phase pays for the next. Most mid-size retailers can complete the first two inside a quarter.
Phase 1 — Measure (1–2 weeks)
Segment AI referrers in analytics as their own channel. Record sessions, assisted revenue, new-customer share and conversion rate separately. You now have the baseline that makes every later argument winnable with finance.
Phase 2 — Become readable (4–8 weeks)
Audit the product catalogue for machine-resolvable attributes and fill the gaps — this is unglamorous data work and it is where the return is. Ship Product, Offer, AggregateRating and FAQPage schema across templates. Make price and availability server-rendered. Publish clear shipping, returns and warranty terms as text, not PDFs or images. Given the 3.1x citation lift on structured data, this phase alone usually moves the needle.
Phase 3 — Become reachable (6–10 weeks)
Expose a clean catalogue and inventory API over your single source of truth. Review bot policy and introduce agent allow-listing with identity verification rather than blanket blocking. If you are on Shopify or a platform backing UCP, most of this arrives as configuration rather than code — check before you build. Legacy stacks usually need targeted application modernization at the integration surface, not a replatform.
Phase 4 — Become transactable (quarter 2 onward)
Add a delegated-checkout path with mandate verification, spending caps and instant revocation, wired to whichever protocol your platform and payment provider adopt. Keep it behind a flag and a limit. Given that only 14% of consumers currently trust an agent to place an order, this is capability you want ready, not capability you want to promote.
Phase 5 — Run your own agents
The inward-facing half, and the one most retailers skip: agents for merchandising, pricing, demand forecasting, content generation at catalogue scale, and post-purchase service. The same structured data that makes you legible to external agents makes your own AI agents useful. We covered the equivalent build discipline for industrial settings in our guide to agentic AI in manufacturing — the governance pattern is identical.
Build, buy, or partner
- Buy the protocol adapters. They will change repeatedly; let your platform or payment provider absorb that churn.
- Build the data layer. Your catalogue, pricing logic and inventory truth are the asset, and no vendor’s fixed schema will model your merchandising rules. This is where ecommerce engineering and custom model work earn their keep.
- Partner when the constraint is capacity. Agentic readiness needs catalogue data engineering, schema and technical SEO, API design, payments and fraud, plus MLOps — an uncommon combination to hold in-house. Staff augmentation or a delivery team covers the peak without permanent hiring.
The metrics to track from day one
- AI-referred sessions, split by source (ChatGPT, Gemini, Perplexity, Copilot, shopping agents)
- Conversion rate of AI referrals versus organic search — and the trend, which matters more than the level
- New-customer share of AI-referred revenue
- Citation rate: how often your products appear in AI answers for your priority queries
- Catalogue completeness: percentage of SKUs with every decision-critical attribute populated
- Schema coverage: percentage of PDPs with valid Product and Offer markup
- Agent request success rate and latency on your catalogue API
- Blocked-agent rate — legitimate agents your bot rules are turning away
The last one is the metric nobody has, and the one most likely to be quietly costing money right now.
Frequently asked questions
What is agentic commerce in simple terms?
Agentic commerce is when an AI agent shops on a person’s behalf — researching options, comparing them, selecting a product and, in some cases, completing the payment. The shopper gives an instruction and a budget; the agent does the work. Most activity today stops at research and recommendation rather than autonomous purchase.
How big is agentic commerce in 2026?
Small but growing very fast. EMARKETER forecasts roughly $20.57bn in US agentic commerce in 2026, about 1.5% of US retail ecommerce, and AI-driven sessions remain under 0.2% of total ecommerce traffic. Adobe Analytics measured roughly 4,700% year-over-year growth in AI-driven visits to US retail sites, so the base is tiny and the slope is steep.
Does AI traffic convert better or worse than search?
The evidence is genuinely mixed. Adobe data has shown AI-referred traffic converting around 23% below non-AI sources, narrowing from a 49% gap earlier in the year, while peer-reviewed research places ChatGPT referrals above paid social but below organic search, direct and email. The consistent finding is that AI traffic brings a higher share of new customers, so treat it as acquisition first.
Which agentic commerce protocol should we implement?
None exclusively, yet. MCP handles model-to-system connection, ACP and UCP handle merchant-side commerce, and AP2 handles payment mandates; specifications are still changing rapidly. The durable investment is a clean, API-addressable source of truth for products, pricing and inventory, with each protocol implemented as a thin adapter over it.
How do we get our products recommended by ChatGPT and AI Overviews?
Publish decision-critical attributes as structured data rather than prose or images, mark up products with Product, Offer and AggregateRating schema, server-render price and availability, and state shipping, returns and warranty terms in plain text. Pages with structured data are cited roughly 3.1x more often in Google AI Overviews, and structured data appears in about 71% of ChatGPT citations.
Should we block AI bots from crawling our store?
Blanket blocking is now a revenue decision, not only a security one, because the same rules that stop scrapers also stop the agents that recommend you. The workable approach is agent identity and allow-listing: permit verified shopping and answer agents, rate-limit unknown ones, and keep aggressive blocking for traffic showing fraud signals.
Is agentic checkout safe from fraud?
It is manageable but not solved. Visa reported a 25% rise in malicious bot-initiated transactions over six months, and Accenture found 78% of financial institutions expect AI-agent-linked fraud to increase. Safe implementations depend on verifiable payment mandates, per-transaction spending caps, cryptographic agent identity and instant revocation rather than on trusting the agent.
How long does agentic commerce readiness take?
Analytics segmentation takes one to two weeks. Catalogue attribute work and schema rollout typically take four to eight weeks for a mid-size retailer, and exposing a clean catalogue and inventory API a further six to ten. Delegated checkout is a second-quarter item, and is best kept behind a flag until consumer trust in agent-placed orders rises above its current 14%.
Where to start this month
Segment your AI referral traffic, then pull fifty of your best-selling SKUs and ask a plain question: could an agent answer every attribute a buyer cares about using only your published data? Wherever the answer is no, you are already losing the recommendation. That audit takes an afternoon and usually settles the business case on its own.
DRC Infotech builds agentic-ready commerce platforms for retailers in the US, UK and Europe — catalogue and data engineering, schema and technical SEO, ecommerce development, agentic framework development and MLOps. See how we work with retail and eCommerce clients, or start with an AI strategy and roadmap engagement.
Want to know how your catalogue looks to an AI agent? Book a 30-minute agentic readiness review and we will audit a sample of your product data, schema coverage and bot policy — no obligation.
Sources
- Adobe Analytics — AI-driven traffic to US retail sites (2025–2026)
- EMARKETER — US agentic commerce forecast, 2026
- Bain & Company, Morgan Stanley, Gartner, J.P. Morgan and McKinsey — 2030 agentic commerce forecasts
- YouGov — US consumer trust in AI for price comparison and order placement
- Kearney — shopper expectations for AI agent use
- Kaiser & Schulze, University of Hamburg / Frankfurt School — peer-reviewed analysis of ChatGPT referral conversion
- Accenture and Visa — AI-agent-linked fraud and bot-initiated transaction data
- Rithum — Commerce Readiness Index on product data quality
- Linux Foundation, Google Cloud, Google/Shopify/Etsy/Wayfair/Target/Walmart — MCP, AP2 and UCP protocol announcements
