A growing share of purchases now start and finish without a human clicking through a page. Protocols like OpenAI’s Agentic Commerce Protocol, built with Stripe, and Google’s Universal Commerce Protocol, built with Shopify, let an AI agent discover a product, add it to cart, and complete payment on a user’s behalf in a single connected flow. Retailers already integrated with these protocols are seeing real order volume come through them. The problem is that most analytics stacks were built assuming a human generates the session: they click a link, land on a page, browse, and convert, all of it leaving the trail conventional tracking depends on. An agent-mediated purchase can skip nearly all of that. There’s no page view funnel to analyse if the agent never rendered a page.
That’s a measurement gap, not just a UX one. If agent-originated orders get attributed to “direct” or lumped in with API traffic because there’s no field distinguishing them, you lose the ability to answer basic questions: which channel is agentic commerce actually replacing, is it incremental revenue or cannibalised from existing paid channels, and which product categories are agents selecting versus which ones need a human deciding. Teams that treat agent checkouts as an edge case now will have a much harder time retrofitting attribution once the volume is large enough to matter, because the data simply won’t exist for the period before tracking was added.
Data Points to Track
- Checkout protocol and initiating agent identity — whether an order came through ACP, UCP, a direct API integration, or a conventional browser session, and which agent platform initiated it where that’s exposed
- Product discovery-to-purchase latency for agent-mediated orders — how long between an agent’s catalog query and a completed transaction, since agent flows compress a journey that used to take a human multiple sessions
- Cart modification and cancellation rate by agent vs. human origin — whether agents complete purchases at a materially different rate than human sessions, and where in the flow agent-initiated carts drop off
- Category and price-point distribution of agent-selected purchases — which parts of the catalog agents gravitate toward, as a proxy for what’s easy for an agent to evaluate versus what still needs human judgement
- Post-purchase event continuity — whether return, refund, and support events for agent-originated orders are linked back to the original transaction, since a broken link here fragments the customer record
Setup Steps
- Add a checkout-origin field to every order record that captures protocol (ACP, UCP, direct API, conventional web/app) rather than inferring it from user-agent strings, which agent traffic won’t reliably populate.
- Instrument the agent-facing catalog and checkout endpoints separately from the human-facing ones, even if they share backend logic, so discovery and conversion events can be measured on their own funnel.
- Propagate the checkout-origin field through to post-purchase systems — returns, refunds, support tickets — so the full order lifecycle stays attributable, not just the initial transaction.
- Build a dashboard comparing agent-originated and human-originated conversion, cancellation, and return rates side by side, refreshed frequently enough to catch a shift in volume before it’s a quarter old.
- Review category and price-point distribution for agent orders monthly against the same breakdown for human orders, to spot where agentic demand is concentrating.
Actionable Insights
The channel attribution split is the number that prevents a false read on growth — a revenue bump that’s actually agentic commerce eating into an existing paid channel looks identical to genuine incrementality until the checkout-origin field exists to separate them. A materially different cancellation rate between agent and human checkouts is worth investigating immediately, since it usually points to a catalog data quality issue an agent can’t work around the way a human can. And keeping post-purchase events linked to checkout origin matters most when volume is still small, because that’s when it’s cheapest to build the pipeline correctly — waiting until agentic orders are a significant share of revenue to add attribution means reconstructing history that was never captured.
Related Resources
Need help tracking this in your app?
Our team sets up analytics pipelines for mobile and web teams every day. Talk to us and get your first events flowing in under an hour.
Talk to an expert