Six Things to Do Now Before Agentic Commerce Leaves You Playing Catch-Up

Over the past few months I've been getting variations of the same question from brands I work with. Something along the lines of: we keep hearing about agentic commerce, what do we actually need to do about it?

The honest answer is that most brands are somewhere between mildly aware and actively ignoring it, and the gap between the ones that have started and the ones that haven't is already meaningful and widening. ACP is processing real transactions. Universal Cart is rolling out across Google Search and Gemini this summer. ChatGPT has somewhere north of 400 million weekly active users and a live purchasing agent. This isn't a forecast anymore.

What follows is the most direct version of what I'd tell a head of ecommerce who has limited time and wants to start in the right place. These aren't in order of difficulty. They're in order of impact and logical dependency.

The first one unlocks everything else.

Step 1: Audit your product data

Everything else in this list depends on this being done properly. And in my experience it's the step where most brands discover the gap between where they think they are and where they actually are.

GTIN coverage is the starting point. Pull your full catalogue and not just the hero products. Audit everything. Check what percentage of active SKUs have a valid, GS1-registered GTIN in your product schema. Then go to the Verified by GS1 tool and spot-check a sample of those GTINs. Confirm they resolve to your brand entity and not a supplier, factory, or unrelated company. I've seen brands with GTINs in their feed that resolve to a completely different fashion label. That's not a minor data quality issue. That's the agent being told your product belongs to someone else.

Beyond GTINs, audit your Schema.org product markup for attribute completeness. Brand field populated on every PDP. Material and fabric attributes present. Sizing information present in the schema and not just displayed visually on the page. Availability status accurate and consistent between your on-site schema and your Merchant Center feed. These fields are what an agent reads when it's evaluating whether your products are recommendation-worthy. Blank fields don't trigger errors. They just reduce the agent's confidence in your data and by extension in recommending you.

The tool I use for this is Screaming Frog. Pull a full crawl, extract the structured data, and build a simple spreadsheet showing coverage by field across the catalogue. For most brands I audit the results are uncomfortable. It's not unusual to find that GTINs are present on maybe 60 or 70 percent of active SKUs and that brand fields are empty on a third of the ones that do have schema. None of that shows up in your weekly performance report. It just sits there.

This single step affects both AI-mediated discovery and agent-led purchasing simultaneously. Everything else builds on top of it.

Step 2: Structure your policies in schema

This is the step most brands haven't thought about at all and it's one of the more commercially significant gaps I find.

When an AI shopping agent is evaluating whether to route a transaction through your store, it doesn't read your returns page the way a human does. It looks for structured policy declarations in your schema markup.

Specifically, MerchantReturnPolicy and OfferShippingDetails in Schema.org. These fields tell the agent programmatically what your return window is, what the conditions are, what shipping options are available, what the delivery timeframes look like, and what it costs.

An agent that can't evaluate your policies clearly has two options. Default to a competitor whose policies are declared in schema, or surface your products with a hedged, low-confidence characterisation that makes a buyer less likely to commit. Neither is a good outcome.

The implementation isn't complicated. It's a few hours of developer time to add the markup correctly. The fields are well-documented in the Schema.org specification. The harder part is usually getting the policy content itself into a consistent, structured format that maps cleanly to the schema fields. A lot of returns policies are written in natural language that's hard to parse programmatically even once the markup exists.

Check your current markup with Google's Rich Results Test and validator.schema.org. If MerchantReturnPolicy and OfferShippingDetails aren't returning clean results, that's a gap worth closing before you start worrying about the more complex steps.

Step 3: Enable agent channels on your platform

The infrastructure for agent-led commerce is native to the major ecommerce platforms now. The question is whether it's been switched on.

For Shopify brands, the ChatGPT sales channel is available in the admin. That's your ACP entry point. The Universal Commerce Protocol is native to Shopify stores, which means UCP-based purchasing through Google's infrastructure is already technically available to you. The question is whether your product data and policy schema are complete enough to make it work reliably, which is why steps one and two come first.

The temptation is to enable the channel first and assume the data will follow. I'd do it the other way around. Get the product data and policy schema right first, then turn on the channel. An agent that finds your store through an active channel but hits incomplete data or missing policies will drop off and that drop-off is harder to diagnose than the equivalent in a human checkout flow.

Step 4: Register with ACP

The Agentic Commerce Protocol is open source and any merchant can implement it. But appearing in ChatGPT's Instant Checkout, the consumer-facing purchasing agent that puts your products in front of ChatGPT's users,  requires registering specifically for that.

For Shopify merchants the process runs through the ChatGPT sales channel in your admin, which handles the connection to Stripe and the ACP merchant registration in one flow. If you're not on Shopify, the starting point is agenticcommerce.dev, which has the current documentation and merchant onboarding process.

The fee structure is worth knowing before you register. OpenAI charges a 4% transaction fee on every completed Instant Checkout purchase. That's on top of standard Stripe processing fees. This is roughly 2.9% plus 30p per transaction in the UK. On a £120 order you're looking at about £8.30 in combined fees. Not a reason to avoid it for most brands, but a reason to model it properly before you sign up and a reason to make sure your margin structure accommodates it.

The payments mechanism through ACP uses a Shared Payment Token. The buyer completes the transaction using their existing preferred payment method inline in the chat interface rather than being redirected to your site. 

Human approval is required at the point of payment. This isn't fully autonomous purchasing. The buyer is in the loop. What changes is where the checkout happens and how much friction sits between consideration and payment.

Step 5: Test agent purchasing yourself

This is the step that most quickly shows you where the gaps are in everything you've done in steps one through four.

Open ChatGPT and Perplexity. Search for your category the way a customer with purchase intent would. See whether your brand appears in the results and what the agent says about you. Then try to complete an actual test purchase through the agent interface. Watch where it drops off. Note what data it couldn't find. Look at which competitor it recommended instead when it couldn't complete your transaction.

The drop-off points are diagnostic. An agent that surfaces your product but can't confirm sizing is telling you your schema attribute coverage is thin. An agent that finds your product but won't complete the transaction is telling you your policy schema is missing or your ACP registration isn't active. An agent that routes to a competitor is telling you their entity data is cleaner than yours.

The most useful version of this test is running it on your own store and on your two or three closest competitors in the same session. The comparison is more instructive than the absolute result. I've done this with several brands and the pattern is consistent. The brand that wins the agent recommendation in a given category is almost never the brand with the best site or the strongest creative. It's the brand with the cleanest, most complete structured data and the clearest policy schema.

Run this test before the next drop. Run it again after. The delta tells you whether the preparation work is landing.

Step 6: Measure agent traffic separately from day one

This is the step that nobody has set up yet and will wish they had six months from now.

Agent-referred sessions have different characteristics from human sessions. Conversion rate is typically higher because the consideration work has already happened before the agent sends the user to your site or completes the transaction autonomously. Average order value can be different because agents may be optimising for specific criteria the user set rather than browsing organically. Returns rate may differ. The behavioural pattern is distinct enough that mixing agent traffic into your human traffic reporting will distort both.

Set up a dedicated segment in GA4 now. Google has a dedicated crawler identity for when its AI agents are visiting sites. There's an AI Assistant channel in GA4 you can use to filter and track agent-referred sessions separately. Do the equivalent in whatever other analytics tools you're using. Create a dedicated view or dashboard that tracks agent traffic conversion rate, AOV, and returns rate alongside but separate from the human equivalent.

The reason to do this from day one rather than retroactively is that the comparison data becomes valuable very quickly as agent traffic grows. If you start segmenting in six months you'll have six months of undifferentiated data that you can't go back and separate cleanly. The setup takes a couple of hours. The reporting value compounds from the moment you turn it on.

The order matters and the time pressure is real

I've laid these out sequentially because the dependencies are real. Turning on agent channels before your product data is clean is like opening a shop before you've stocked the shelves. Registering with ACP before you've structured your policies gives agents a reason to route transactions to competitors who've done that work. 

Testing agent purchasing before you've enabled the channels gives you an incomplete picture of where you actually stand.

Do them in order. Don't skip the data work to get to the more exciting channel enablement steps.

On timing: the brands I work with that have started this work are already in a meaningfully different position from the ones that haven't. The entity signals that determine agent recommendation confidence build slowly and compound over time. A brand that starts the GTIN audit and policy schema work now has a six-month head start on a brand that decides to think about this after the next planning cycle.

Most of the work in steps one and two is unglamorous and doesn't show up in any dashboard anyone is currently looking at. That's why it doesn't get done. That's also why doing it now creates a genuine advantage rather than just meeting a baseline.

The technology is moving faster than most ecommerce teams are planning for. Starting now doesn't guarantee you'll be ahead. Not starting now almost certainly means playing catch-up in a context that won't wait for you.