Who Gets Credit When an AI Agent Buys? An Attribution Framework for Agentic Commerce: ChatGPT Instant Checkout died within six months, but agent-mediated buying is returning through Google and Shopify infrastructure. Here is a taxonomy for what attribution can and cannot claim when an agent touches the order.
Read the full article below for detailed insights and actionable strategies.
The attribution problem
One sale. Four channels. 400% credit claimed.
Reported revenue: €400 · Actual revenue: €100 · Gap: €300
When an AI agent buys on a customer's behalf, no attribution model on the market can cleanly assign credit, because the agent compresses discovery, comparison, and checkout into a machine session your tags never see. The workable answer is a taxonomy. Agent-initiated, agent-assisted, and agent-influenced orders each support different attribution claims, from aggregate causal measurement down to informed estimation. With agent infrastructure shipping from Google and Shopify in 2026, define the categories now, before the volume forces the question.
What happened to ChatGPT Instant Checkout?
The first attempt at agent-run checkout failed fast. OpenAI launched Instant Checkout on September 29, 2025 (OpenAI's announcement), letting shoppers buy inside ChatGPT without visiting the merchant's site. Adoption never arrived: fewer than 15 to 30 merchants ever went live, and Walmart's conversion through Instant Checkout ran at roughly one third of its site conversion.
The what is documented; the why is inference, so treat it as such. A conversion rate at one third of site suggests shoppers were comfortable researching in chat but not finishing there, and merchants saw little reason to integrate a checkout that shrank their funnel. The market learned an expensive lesson: consumers will let an agent advise a purchase long before they let it complete one.
By March 2026 the feature was effectively dead. On March 24, 2026, OpenAI pivoted to discovery-first: product research stays in the chat, but the transaction goes back to retailer sites.
Two lessons for measurement. First, even when the platform owned the entire transaction, merchant-grade reporting never materialized. Second, the discovery surface that survived is now monetized with ads, which carry their own GA4 measurement gap.
How is agentic commerce returning through infrastructure?
The second wave runs through pipes, not chat windows. On January 11, 2026, at NRF, Google and Shopify announced the Universal Commerce Protocol, a standard way for agents to discover products and transact across merchants. On May 20, 2026, at Google Marketing Live, Google announced Universal Cart, a cross-merchant cart that agents can act on.
Incumbents are defending the checkout while the pipes get built. On March 10, 2026, Amazon won an injunction against Perplexity's agent shopping feature. Expect agent access, and any reporting that comes with it, to be negotiated platform by platform rather than granted by default.
Universal Cart sharpens the credit problem in particular. One agent session can now assemble a basket across several merchants, which means no single platform sees any one merchant's full journey either.
Notice what the infrastructure model does to measurement. In the Instant Checkout model, the platform owned the session and the merchant got whatever reporting the platform chose to provide, which turned out to be very little. When transactions execute against merchant systems through a protocol, your order data stays yours. That is a genuine improvement, but only for agent-initiated volume, and only if you build the aggregate measurement to read it.
| Date | Event | Why it matters for attribution |
|---|---|---|
| Sep 29, 2025 | ChatGPT Instant Checkout launches | First platform-owned agent checkout; fewer than 15 to 30 merchants ever live |
| Jan 11, 2026 | Google and Shopify announce Universal Commerce Protocol at NRF | Agent transactions become shared infrastructure, not a chat feature |
| Mar 10, 2026 | Amazon wins injunction against Perplexity agent shopping | Platforms gate agent access; reporting will be negotiated |
| Mar 24, 2026 | OpenAI pivots to discovery-first; Instant Checkout effectively dead | Influence stays in chat, checkout leaves: referral-style measurement returns |
| May 20, 2026 | Google announces Universal Cart at Google Marketing Live | One agent session, many merchants, credit assignment unclear |
All of this lands on top of an already crowded year of measurement change, tracked in the 2026 attribution changelog.
What are agent-initiated, agent-assisted, and agent-influenced orders?
The taxonomy that matters, defined by who acts and what your stack can see:
- Agent-initiated: the agent researches, decides, and completes the purchase through a protocol like UCP. The human never visits your site.
- Agent-assisted: the agent researches and compares, then hands off. The human clicks through and buys in a normal session.
- Agent-influenced: the agent shaped the shortlist earlier in the journey. The human buys later through branded search or by typing your URL.
Concrete versions, using a generic skincare purchase. Agent-initiated: a shopper tells their assistant to reorder the usual moisturizer when it drops below €30, and the agent executes through UCP without a site visit. Agent-assisted: a shopper asks which retinol cream suits sensitive skin, clicks the assistant's product link, and buys in that session. Agent-influenced: a shopper reads an assistant's comparison in the morning, then buys that evening by typing your URL.
| Type | What your analytics see | What attribution can honestly claim |
|---|---|---|
| Agent-initiated | An order with no user session, or a machine session | Aggregate causality only: causal inference across regions or time periods, plus incrementality testing. No user-level path exists. |
| Agent-assisted | A referral session when the referrer survives; direct traffic when it does not | Touch-level credit inside a multi-touch attribution model, fragile because the referrer often drops |
| Agent-influenced | Direct or branded organic; the assistant touch is invisible | Almost nothing at user level. Estimation through surveys and aggregate models such as marketing mix modeling |
The middle category is where the volume sits in July 2026, and it is the one your current stack half-sees. When the referrer survives, the session looks like any referral. When it does not, the order lands in Direct and the assistant's role vanishes, which is why we wrote a diagnostic for finding AI referrals hiding in GA4.
What can attribution honestly claim when an agent buys?
For agent-initiated orders, user-level attribution is over, and any vendor promising per-user credit for machine orders is inventing data. The unit of analysis becomes the market, not the user. Compare regions or periods where agent access is live against those where it is not, and let causal inference estimate the counterfactual: what would sales have been without the agent channel. Incrementality becomes the only honest credit claim.
For agent-assisted orders, keep your tagging and referrer discipline, then accept the undercount. Last click attribution will systematically mis-credit Direct and branded search for journeys the assistant started. Multi-touch attribution does better but still breaks wherever the referrer drops.
For agent-influenced orders, no tag will ever see the touch. Triangulate instead: post-purchase surveys, aggregate reads, and honest ranges reported to your CFO as estimates, not facts.
Operationally, report agent-touched revenue in three lines. Measured: agent-assisted orders where the referrer survived, the smallest and hardest number. Estimated: the agent-influenced share from surveys and aggregate models, reported as a range. Unknown: everything else, stated plainly. A CFO who sees those three lines will trust your measurement more than one who receives a single precise figure nobody can defend.
One more warning: platform-reported ROAS for agent surfaces will carry the same seller-grading problem as every walled garden before it. Verify against your own order data before you scale spend toward any agent channel.
An illustrative case: an anonymized Dutch home goods brand with €2.4M in quarterly revenue estimates that 6% of orders, about €144K per quarter, are now agent-touched, mostly agent-assisted. Its causal read asks the only question that matters for that line: how much of the €144K would have arrived anyway. Re-run the read monthly as protocol volume grows, because the agent-touched share will move faster than any annual review cycle.
That is the read Causality Engine runs on your GA4 export: causal read out in 5 to 10 minutes, no pixel, no annual lock-in, €99 per read or €299 per month on Pro. Run a causal read on your own GA4 export and put a defensible number on the agent-touched share.
Key takeaways
- ChatGPT Instant Checkout (September 29, 2025) was effectively dead by March 2026, with fewer than 15 to 30 merchants live and Walmart converting at roughly one third of its site rate; OpenAI went discovery-first on March 24, 2026.
- Agentic commerce is returning through infrastructure: Universal Commerce Protocol (January 11, 2026) and Universal Cart (May 20, 2026), while Amazon's March 10, 2026 injunction against Perplexity shows checkout access will be fought over.
- Use the taxonomy: agent-initiated (aggregate causality only), agent-assisted (fragile touch-level credit), agent-influenced (estimation only).
- No attribution model can assign user-level credit for agent-initiated orders; the counterfactual is the only honest unit of credit.
- Build incrementality measurement now, before protocol-driven volume makes the question urgent.
Further reading
Get attribution insights in your inbox
One email per week. No spam. Unsubscribe anytime.
Key Terms in This Article
Agentic Commerce
Agentic commerce is shopping performed by AI agents on a customer's behalf: the agent finds, selects, and purchases products through APIs or automated checkout rather than a human browsing a storefront session.
Attribution Model
An Attribution Model defines how credit for conversions is assigned to marketing touchpoints. It dictates how marketing channels receive credit for sales.
Causal Inference
Causal Inference determines the independent, actual effect of a phenomenon within a system, identifying true cause-and-effect relationships.
Conversion rate
Conversion Rate is the percentage of website visitors who complete a desired action out of the total number of visitors.
Counterfactual
Counterfactual is a hypothetical outcome that would have occurred if a subject had received a different treatment.
Incrementality Testing
Incrementality Testing measures the additional impact of a marketing campaign. It compares exposed and control groups to determine causal effect.
Marketing Mix Modeling
Marketing Mix Modeling (MMM) is a statistical analysis that estimates the impact of marketing and advertising campaigns on sales. It quantifies each channel's contribution to sales.
Multi-Touch Attribution
Multi-Touch Attribution assigns credit to multiple marketing touchpoints across the customer journey. It provides a comprehensive view of channel impact on conversions.
Related Articles
Sixty-second versions of these ideas: Causality Engine on YouTube Shorts.
Ready to see your real numbers?
Own the budget? Upload your GA4 export and see which channels drive incremental sales, with confidence intervals, in minutes. Have to defend it? Start with the live demo and take the read to your CFO.
Full refund if you don't see value.
Stay ahead of the attribution curve
Weekly insights on marketing attribution, incrementality testing, and data-driven growth. Written for the person who owns the budget and the person who has to defend it.
No spam. Unsubscribe anytime. We respect your data.
Frequently Asked Questions
What is agentic commerce?
Agentic commerce is buying mediated by AI agents: software that researches, compares, and sometimes completes purchases on a shopper's behalf. In 2026 the infrastructure includes the Universal Commerce Protocol from Google and Shopify, announced January 11, 2026, and Google's Universal Cart, announced May 20, 2026, which let agents transact across merchants.
Did ChatGPT Instant Checkout fail?
Yes, by any adoption measure. Launched September 29, 2025, it never attracted more than 15 to 30 live merchants, and Walmart converted through it at about a third of its normal site rate. The feature was effectively dead by March 2026, and OpenAI pivoted to discovery-first on March 24, 2026.
What is the difference between agent-initiated and agent-assisted orders?
In an agent-initiated order, the agent researches and completes the transaction itself, usually through a protocol, and the shopper never visits your site. In an agent-assisted order, the agent handles research and comparison, then the shopper clicks through and completes the purchase in an ordinary, measurable session.
Can any attribution model credit an agent-initiated order?
Not at user level, because there is no user session to track. The honest approach is aggregate: causal inference comparing regions or periods with and without agent access, plus incrementality testing to estimate the counterfactual. Any vendor promising user-level credit for machine-placed orders is fabricating numbers.
Why did Amazon sue over Perplexity's shopping agent?
Because agent access to a storefront threatens the platform's control of checkout, ads, and customer data. Amazon won that injunction on March 10, 2026. Expect reporting access for agent traffic to be negotiated case by case, one platform at a time, instead of granted by default.
How should brands prepare attribution for agentic commerce?
Define your taxonomy now: initiated, assisted, influenced. Keep GA4 export discipline and referrer tagging for assisted journeys, add a post-purchase survey for influenced demand, and build aggregate causal reads for initiated volume. A causal read on your GA4 export takes 5 to 10 minutes and costs €99.