Agentic commerce attribution means crediting revenue to the AI agent that completed the purchase, not to whatever channel happens to catch the leftover session. A protocol-checkout order arrives through an API call with no browser, no cookie, and no session for a client-side tool to join, so standard analytics either drops it or files it as unattributed direct traffic. A defensible model routes that revenue to a warehouse you control instead of asking a vendor tool to reconstruct a session that structurally does not exist.
Landscape as of September 2026: OpenAI and Stripe rolled back ACP's Instant Checkout in March 2026, so the live model is discovery-plus-merchant-redirect for that protocol today, with Google's AP2 payment-consent layer (three signed Mandates, donated to FIDO in April 2026) operating alongside it as a complementary layer, not a replacement. This is a fast-moving stack; treat the protocol specifics below as accurate to this month and re-check before quoting them in a year.
Why the number you are reporting today is probably wrong
Two independent 2026 research threads (commercetools and digitalapplied's agentic-commerce guides, cross-referenced against Adobe Analytics and Shopify's own published figures) converge on the same shape of finding. AI-driven visits to US retail grew 4,700% year over year in 2025, and by March 2026 that AI traffic was converting 42% better than non-AI traffic - a full reversal from March 2025, when the same comparison ran 38% worse. Shopify reports AI-attributed orders grew 11x between January 2025 and March 2026. Those are named, dated figures from the sources above, not our own measurement.
Set against that growth, the same research estimates that roughly 70.6% of AI referrals go invisible to standard GA4 instrumentation, and that AI referral traffic is undercounted by a factor of 3 to 4. [hypothesis, per the cited research - the exact share will vary by site and stack; treat it as a directional estimate, not a number to plug into your own model unmeasured].
The finding we did not have to estimate, because we ran it
The Webclat demo store runs a working ACP-shaped protocol checkout, and we sent a real order through it end to end. GA4's Measurement Protocol returned its own verdict on the resulting event: "no _ga cookie; sent under anonymousId, no session id; event will not join a session." That is GA4's answer, quoted, not our interpretation. A server-sent event with no session id also lands in GA4 Realtime and then fails to appear in the standard reports - a second, separate limitation of the same platform, also observed directly rather than assumed.
It rules out "send the event server-side and the gap closes" as a complete fix - we did exactly that and the session-level reporting still failed, because the gap is structural to GA4's session model, not a delivery problem. It does not rule out server-side collection generally: the same event landed correctly in both warehouse legs (ClickHouse and BigQuery), which is the argument for warehouse-native measurement of this channel rather than vendor-native. See the working exhibit at store.webclat.com/agentic-commerce for the full run, curl transcript included.
A model worth defending in a room, not just publishing
Per house method (three buckets, never one hero number - the same discipline the demo store's signal-recovery calculator applies to blocked-tag and ITP loss), agent-completed revenue splits into three populations that need three different treatments, not one blended figure:
Browser-driving agents
Tracked normally - your tags fire - but misattributed, since a competent agent session looks statistically unlike a human one and can distort behavioural metrics if left unsegmented.
Protocol checkout
Genuinely invisible to client-side tools by structure, not by a fixable bug. Needs server-side capture routed to a warehouse, tagged with an explicit agent-channel dimension.
Crawlers claiming agent identity
A meaningful share of traffic claiming to be ChatGPT, Perplexity or similar agents is spoofed. Attribute only what a signature or composite signal actually supports; an unverified claim in a header is not evidence.
Blending these three into a single "agent revenue" line is the fastest way this kind of model fails in front of a competent analyst, for the same reason blending blocked-tag recovery with ITP reattribution does on the calculator above: each bucket has a different mechanism, a different confidence level, and a different fix.