Agentic commerce: what it means, and what to do about it now
A growing number of shoppers no longer browse your site before they buy from it.
They ask an AI assistant to find them running shoes under $150, or a birthday gift for someone who likes cooking, and the assistant does the comparing, the filtering and increasingly the buying on their behalf.
This isn't a future scenario anyone needs to prepare for eventually.
Shopify and Google's own commerce protocol went live in January 2026, OpenAI's checkout tooling has been processing live transactions since late 2025, and Adobe Analytics has tracked AI-referred traffic to US retail sites growing by roughly 4,700% year on year. The shift has already started, and most ecommerce brands aren't ready for it.
What is agentic commerce?
The word "agent" gets used loosely, so it's worth being precise.
A chatbot answers a question when you ask it one.
An agent does something with the answer.
Salesforce draws the distinction cleanly: chatbots are reactive, agents are proactive, retrieving data, planning a sequence of steps and carrying them out without needing a human to approve each move.
In an ecommerce setting, that plays out in four stages.
A shopper states an intent, specific or vague. The agent searches across catalogues, checks specs and stock, and narrows the field. It builds a basket, applies any relevant loyalty or discount logic, and completes payment through a tokenised system that keeps the shopper in control of spending limits. Then it keeps an eye on delivery and can start a return or exchange without the shopper lifting a finger. None of this requires the shopper to ever land on your homepage.
Our client Perxify, an Estonian commerce platform, describes the same shift in starker terms in its own writing: a move from a web of pages to what it calls the agentic web, where the traditional sales funnel stops being about guiding a human through a website and becomes a series of algorithmic evaluations instead. Whether you buy the full framing or not, the underlying point holds. Success increasingly depends on a brand being discoverable and evaluable by machines, not just persuasive to people.
Whether you buy the full framing or not, the underlying point holds. Success increasingly depends on a brand being discoverable and evaluable by machines, not just persuasive to people.
How agentic commerce protocols work: UCP, ACP and MCP
Three technical standards are doing most of the work here, and it's worth knowing their names even if you never touch the implementation yourself.
The Universal Commerce Protocol, built by Google and Shopify, covers the whole journey from discovery through to post-purchase support, and its backers now include Amazon, Meta, Microsoft, Salesforce, Visa and Mastercard. The Agentic Commerce Protocol, built by OpenAI and Stripe, is narrower and handles checkout specifically. It's already live for Etsy and rolling out across Shopify's merchant base.
The Model Context Protocol, originally built by Anthropic, gives AI systems a standard way to pull real-time data from a catalogue or inventory system, and it's becoming the connective tissue enterprise commerce platforms use to feed agents accurate information.
You don't need to choose between them.
Most brands end up running two or three at once, and if you're on Shopify, a lot of this is already switched on for you by default.
Perxify has written about what building for this shift looks like from the engineering side, in a piece on an API-first approach to the AI-native web, worth a read if you want the infrastructure view rather than the marketing one. Being built agent-ready from day one, as that post argues, is a genuinely different starting position than retrofitting an older platform once agent traffic already matters.
Why product data is the new battleground in agentic commerce
Here's the part that should worry marketers more than the protocol names do.
Agents don't respond to persuasive copy, hero images or a clever discount pop-up.
They read structured data: attributes, materials, dimensions, stock status, price. Channable's breakdown of what it calls the three pillars of agentic readiness comes down to the same point in three different ways: your feed has to be complete, it has to be consistent, and it has to be accurate in close to real time, because an agent that recommends an out-of-stock item loses trust in both the retailer and the assistant that suggested it.
Forbes contributor Catherine Erdly puts the practical version of this well: before chasing clever prompts or new distribution channels, brands need their catalogue to be legible to a machine, with structured data such as JSON-LD, an AI-readable sitemap and a working product catalogue API. Skip that step, and it doesn't matter how good your positioning is, because the agent never sees it.
This is a different job to the one most ecommerce marketing teams are set up to do, and it overlaps heavily with the work we've already been writing about on getting your brand cited inside LLM answers and tracking whether those citations are actually happening.
The tools are similar. The stakes are higher, because a missed citation costs you a mention, while a missed feed costs you the sale.
Generative engine optimisation: the SEO discipline built for AI agents
If structured data is the foundation, then generative engine optimisation, or GEO, sits on top of it. Perxify's own framing of GEO in its engineering writing is a useful way to break the discipline down into three parts.
The first is semantic understanding rather than keyword matching.
An AI system connects a search for a specific concept to products described in entirely different language, provided the underlying meaning lines up. This means product copy needs to be written for how a model interprets meaning rather than for the exact words a shopper might type.
The second is what gets called query fan-out: a single request gets broken down into dozens of related searches behind the scenes, so a brand that only optimises for the query it expects is missing the adjacent ones the agent is quietly running as well.
The third is individual-level personalisation, where recommendations are tailored to one shopper's specific history rather than a broad customer segment, which raises the bar for how rich and current your customer data needs to be.
None of this replaces conventional SEO.
It sits alongside it, and the overlap with what we've already written about brand visibility inside LLMs is significant, because the same structured, entity-rich content that gets you cited in a chat answer is largely the same content that gets you selected by a shopping agent.
AI product recommendation engines and on-site personalisation
External agents solve discovery.
They don't solve what happens once a shopper actually lands on your site, or inside your own AI-powered storefront. This is where on-site personalisation and an AI product recommendation engine earn their keep, and it's the argument Perxify makes in its own writing on rebuilding personalisation around an AI product recommendation engine.
The illustrative case Perxify uses is a useful one, even treated as a scenario rather than a reported result: a shopper who buys a record and a turntable gets shown a listening lamp on a completely different store within the same commerce network, because the recommendation engine is working from a shared profile rather than one store's isolated data.
That's a meaningfully different proposition to the single-store recommendation widgets most small and medium ecommerce brands are used to, and it's the same argument Salesforce makes when it describes ecommerce agents as needing trusted, structured data before they can act well on a shopper's behalf, one of the five attributes Salesforce sets out for a working commerce agent, alongside a clearly defined role, a set of actions it can take, guardrails on what it shouldn't do, and the channels it operates across.
The practical takeaway for a smaller brand isn't that you need a shared commerce network to compete. It's that on-site personalisation is no longer optional polish.
A recommendation engine that only ever shows a shopper what's already popular is leaving conversion on the table against a competitor whose agent, on-site or external, actually understands what that specific shopper wants.
The agentic commerce market opportunity, by the numbers
The scale of this shift is why the big platform vendors are moving so fast.
Gartner's research, cited by Salesforce, puts enterprise adoption of agentic AI at 33% by 2028, up from under 1% today. McKinsey has estimated the wider agentic commerce opportunity at $3 to $5 trillion by 2030.
Gartner's 33% by 2028 figure originally comes from Gartner's Intelligent Agents in AI research, and was picked up and cited on Salesforce's agentic commerce page, which is where we pulled it from. McKinsey's $3 to $5 trillion by 2030 estimate comes from McKinsey's own article, The agentic commerce opportunity, which we found referenced in Fin.ai's agentic commerce guide.
McKinsey's projected agentic commerce opportunity by 2030
Source: McKinsey, "The agentic commerce opportunity" (2026)
The categories most exposed to this shift are also among the largest and already moving fastest.
The global secondhand apparel market, for instance, grew 18% to $197 billion in 2023 according to resale platform ThredUp's annual report, cited by GlobalData, roughly fifteen times faster than general retail. That's exactly the kind of high-volume, high-friction category where structured listings and AI-assisted pricing genuinely change how much supply reaches the market, and that agentic commerce infrastructure is being built for first.
It's also worth noting why some of this infrastructure is being built specifically in Estonia.
StartupBlink ranked Estonia the number one AI startup ecosystem in the EU, and sixth globally, a result of strong government digital infrastructure and a concentrated, well-connected local talent pool. That's part of the pitch behind Perxify's own positioning, and it's a genuinely credible one rather than marketing colour, since Estonia's broader digital-government track record long predates the current AI wave.
Whatever the exact size of the opportunity turns out to be, the direction isn't in question.
Where humans and on-site AI still matter in ecommerce
None of this replaces the need for a good on-site experience.
External shopping agents are good at macro discovery, working out which merchants carry the right product and comparing prices across them. They're much weaker at the micro moment where a shopper has narrowed things down to three options and needs a specific question answered, or needs a return sorted quickly enough that they'll come back and buy again.
Brands that optimise only for external agent visibility and neglect their own on-site assistance are solving only half the problem.
The ones getting this right are pairing feed-level discoverability with a genuinely useful assistant on their own storefront, whether that's a dedicated AI shopping tool or a well-built site search layer purpose-built for this kind of query, of the sort vendors like Athos Commerce are now building specifically for.
Bluecore puts it well: agents don't browse a catalogue the way a human does; they find, filtering out anything that doesn't match a shopper's established pattern before it's ever shown. That makes the on-site layer, not just the external one, the place where a brand's actual differentiation still matters.
How to prepare your ecommerce store for agentic AI
None of this requires ripping out your commerce stack. It requires a shift in priority.
Audit your product data before you touch anything else. Complete attributes, consistent formatting, accurate stock and pricing in close to real time. This is unglamorous work and it matters more than almost anything else on this list.
Add structured data and an AI-readable sitemap so agents can actually parse what you sell, not just what a human visitor would see.
Invest in your on-site recommendation and search layer alongside external visibility. A great feed with a poor on-site experience still loses the sale once the shopper arrives.
Treat brand mentions as a KPI alongside clicks. Getting cited accurately inside an AI answer is now part of the funnel, and it needs the same attention as paid search and social spend already gets.
Build the business case properly. This is a genuinely hard thing to justify to a finance team on a hunch, and the same discipline that applies to proving SEO ROI applies here: measure what you can, and be honest about what you can't yet.
The uncomfortable bit
A lot of what gets published on this topic right now is vendor marketing dressed up as trend analysis, ourselves included from time to time.
Not every retailer needs to solve for all three protocols this quarter.
What every retailer needs is a product feed that doesn't quietly disqualify them from a channel that's growing faster than almost anything else in commerce right now. That's a smaller, more tractable problem than the hype suggests, and it's exactly the kind of groundwork worth getting right before the bigger decisions arrive.
If you're trying to work out whether this is worth prioritising this quarter or next, it's the same conversation we tend to have early with any client we take on, and one covered from a different angle in our piece on getting a business noticed in a crowded market.
-
No. Most brands don't need to touch their underlying platform. If you're on Shopify, Agentic Storefronts already handles the UCP and ACP protocols by default. The work that actually matters is cleaning up your product data and adding structured markup, not migrating platforms.As a small, time-boxed experiment, yes, worth testing. As a meaningful budget shift away from proven channels, not yet. The reach and targeting limitations are real and structural, not something a well-optimised campaign can work around today.
-
No. Smaller ecommerce brands are affected just as much, often more, because agents penalise incomplete or inconsistent product data regardless of company size, and a small catalogue with clean, complete attributes can outperform a large one with messy data.
-
It depends on your existing stack. If you're on Shopify, both UCP and ACP are largely handled for you. If you already use Stripe for checkout, ACP is the natural starting point. If you have a strong Google Merchant Center presence, prioritise UCP. Most brands end up needing at least two of the three eventually.
-
Check three things: completeness (every product has full attributes, not just the ones that look good in marketing copy), consistency (formatting, units and naming are uniform across the catalogue), and accuracy (stock and pricing are correct in close to real time). Missing any one of these means agents are likely to skip your products in favour of a competitor's.
