Reference
Agentic commerce is online retail in which an AI agent, acting for a person, carries out the shopping itself. It discovers products, reads their data, compares options and attempts the purchase, rather than presenting a list of links for the person to click. The distinction that matters to a merchant is that the storefront is now consumed by software reading structured data and driving a browser, not by a person looking at a page.
Last verified: 28 August 2026. Every claim below names its source in the sentence.
An agent follows the same four steps a person does, in the same order, and fails in different places. Each step below is a different kind of reading.
The agent has to find the product before it can consider it. That happens through search results, through a syndicated catalogue, or through a feed the platform supplies on the merchant’s behalf. It does not happen through a carousel on your homepage.
Shopify says that AI searches powered by Shopify Catalog convert at twice the rate of those using scraped data, which is a vendor claim about its own product and is worth reading as one. The underlying point stands regardless of the multiple: an agent given structured, authoritative product data behaves differently from one scraping a page.
What an agent consumes is not what a shopper sees. It reads structured product data and JSON-LD, product feeds and catalogue APIs, policy pages as plain text, and your robots.txt and bot rules to work out whether it is welcome at all. Some sites also publish an llms.txt index. Its value is genuinely contested and adoption figures for it are unreliable, so treat it as cheap to add and unproven rather than as a requirement.
On one major platform this reading step is a formal interface. Every Shopify store has its own Storefront MCP endpoint, which Shopify documents as connecting an AI agent to that store’s catalog, shopping cart, and policies.
Comparison is where a storefront built for a human eye quietly stops working. Text baked into an image is invisible. A specification that lives only in a hover tooltip or a drag interaction cannot be read. A shipping cost or a returns window revealed only in the last step of checkout arrives too late to be compared against a competitor who published it on the product page.
The last step is the one that produces a sale or produces nothing. An agent driving a browser meets interstitials, cookie walls, forced account creation, a size picker it cannot operate and a validation error it cannot read. None of these announce themselves. The agent simply stops.
Why AI agents abandon carts covers the failure shapes in this step in detail, and the agentic readiness checklist puts them in the order an agent hits them.
Four names come up constantly and they do different jobs. Each one below opens with what it is and then, more usefully, what it is not.
ACP
The Agentic Commerce Protocol is maintained by OpenAI and Stripe, is published under Apache 2.0, and covers the checkout step: cart, fulfilment options and delegated payment, with the business remaining merchant of record.
What it is not: a discovery standard. It says nothing about how the agent found the product.
UCP
Google announced the Universal Commerce Protocol at NRF on 11 January 2026, describing it as built together with Shopify, Etsy, Wayfair, Target and Walmart and endorsed by more than twenty others.
Shopify describes UCP as the standard it co-developed with Google, covering, in its words, the full commerce journey from discovery to carts to checkout.
What it is not: a competitor to the others. Google states that UCP is built to work with existing protocols including Agent2Agent, the Agent Payments Protocol and the Model Context Protocol, so these are layers rather than rivals.
AP2
Google announced the Agent Payments Protocol on 16 September 2025 with more than sixty organisations. AP2 does not move money. It produces signed mandates, an Intent Mandate for the delegation and a Cart Mandate that Google describes as a secure, unchangeable record of the exact items and price.
What it is not: a payment rail. It is proof that a person authorised this cart at this price, which is a different problem from settling the transaction.
MCP
MCP is the general interface by which a model calls tools, and in commerce it shows up as a per-store endpoint: a catalogue, a cart and a set of policies an agent can query directly rather than scrape.
What it is not: an industry commerce standard. It is plumbing, and each store exposes its own.
The point most protocol coverage misses is that none of them is load-bearing for most stores today. When no protocol is present, the agent does not give up. It falls back to driving an ordinary browser against an ordinary storefront, which is exactly the path nobody is testing. ACP and UCP compared goes further into where the two overlap.
Agentic commerce collects forecasts. This section deliberately carries none of them. Every item below is something that exists now, with the source named in the sentence.
Signature-Agent header for in-band key discovery and a JWKS key directory served from a well-known URI, on top of HTTP Message Signatures.Two of those belong together. The category of traffic now has a name in the policy layer at Cloudflare, and a draft standard at the IETF for proving which agent is which. Being identifiable is becoming the price of being admitted, which is why our own agents publish their user agent and IP addresses.
Short section, and short on purpose. The only well sourced trajectory figures come from one platform reporting on itself, and dressing a single company’s numbers as an industry trend is how this subject gets oversold.
On Shopify’s Q1 2026 earnings call in May 2026, president Harley Finkelstein said AI-driven traffic to Shopify stores had grown 8x year over year, and that orders from AI-powered searches had increased nearly 13x.
The one demand-side number here that is not a platform reporting on its own performance is the Semrush survey above, and its population is narrow by design: 1,030 US shoppers who had already used AI tools. Read it as evidence that in-tool purchasing is real, not as a share of all retail.
The forecasts, adoption figures and agent success rates that do exist, each with its primary source and date, are collected in agentic commerce statistics.
Here is the asymmetry that makes this worth attention. A shopper who cannot complete a purchase leaves a trace: a bounce in your analytics, an abandoned cart email, occasionally a support ticket. An agent that stalls at your size picker does none of those things. It does not bounce in a way you can see, it does not email support, and it does not appear as a lost sale anywhere in your reporting. It goes to the next tab, and the merchant on the other side of that tab gets an order.
Two data points suggest where the attempt will happen, and both point at your own checkout rather than someone else’s.
Walmart’s Daniel Danker, its EVP for product and design, told WIRED in March 2026 that products sold directly inside ChatGPT converted about three times worse than when shoppers clicked through to Walmart’s own site, calling the in-chat experience “unsatisfying.”
OpenAI’s retirement of Instant Checkout, quoted above, moved in the same direction: merchants use their own checkout experiences, and the assistant concentrates on product discovery. If the agent is going to end up on your storefront to finish the job, your storefront is the thing under test. Can ChatGPT buy from your store? walks through that attempt on a real shop.
The most useful sentence on this page for understanding the gap is Cloudflare’s own. Its Agent Readiness score checks for agentic commerce standards and explicitly does not count them towards the score. That is the honest boundary of every signal check: it reads a storefront and reports what it can see from outside.
Ottom does the other thing. A real AI agent attempts to discover, evaluate, compare and buy, and the report says at which stage it stopped and what it hit. A signal check reads a storefront. An attempt drives it. Both are useful and they answer different questions.
Ottom publishes what its agents meet on the stores it scans. Across 317 stores Ottom measured at every stage in the 90 days to 19 October 2026, in every category and including stores whose category could not be determined, the median agentic readiness score was 97 out of 100. 87% of them let an AI shopping agent reach the payment step.
A store enters Ottom’s survey only if a catalogue of at least ten products can be read from it in one request, which in practice means Shopify and WooCommerce storefronts. These figures therefore describe the storefronts Ottom can read in bulk rather than online retail as a whole. The stores are a convenience sample, not a random or representative sample of online retail: a store is counted because someone scanned it or because its catalogue could be read in bulk, so these figures describe what Ottom has seen and nothing wider.
The all stores benchmark has every stage and the stalls that stop agents most often, and agentic commerce statistics sets it beside the third-party figures. Still, the only number worth having about your store is your own.
Or estimate the exposure first with the AI commerce ROI calculator, see what a full report costs, or start with what is agentic commerce if this is new ground.
Agentic commerce is online retail in which an AI agent, acting for a person, carries out the shopping itself. It discovers products, reads their data, compares options and attempts the purchase, rather than presenting a list of links for the person to click. The merchant surface being consumed is the same storefront, read by software.
Probably none of them yet, and that is the honest answer. The Agentic Commerce Protocol covers checkout, the Universal Commerce Protocol covers discovery through checkout, the Agent Payments Protocol covers proof of authorisation, and the Model Context Protocol is a per-store interface rather than an industry standard. None is load-bearing for most stores today, because when no protocol is present an agent falls back to driving an ordinary browser against an ordinary storefront. That fallback path is the one worth testing first.
Not by default, and the widely repeated version of this is wider than the source. Cloudflare’s changelog of 1 July 2026 says that from 15 September 2026, new domains onboarding to Cloudflare default to blocking Training and Agent traffic on pages that display ads, that Search traffic remains allowed, and that customers can opt out of the new defaults at any time before that date. Most storefronts do not display ads and are not new to Cloudflare. The real change is narrower and more durable: agent traffic is now a policy category of its own, so the question of whether you admit it is now a setting somebody has to answer.
Ottom measures it on real stores and publishes the result, recomputed nightly. Across 317 stores Ottom measured at every stage in the 90 days to 19 October 2026, in every category and including stores whose category could not be determined, the median agentic readiness score was 97 out of 100. 87% of them let an AI shopping agent reach the payment step. A store enters Ottom’s survey only if a catalogue of at least ten products can be read from it in one request, which in practice means Shopify and WooCommerce storefronts. These figures therefore describe the storefronts Ottom can read in bulk rather than online retail as a whole. The stores are a convenience sample, not a random or representative sample of online retail: a store is counted because someone scanned it or because its catalogue could be read in bulk, so these figures describe what Ottom has seen and nothing wider.
No. Shopify is simply the platform that has published the most about it, which is why so many verifiable facts on this page carry a Shopify source. The mechanics are platform independent: structured product data, readable policy pages, a checkout an automated client can complete, and a bot policy that admits agents you want. A store on any platform, or on none, meets an agent through the same four steps.