Agentic Commerce Attribution: What Merchants Can Actually Measure (2026)
Quick summary: Last-click will not explain an order that started in ChatGPT or AI Mode. ACP has an affiliate_attribution field. It is not a reporting suite. On 25 September 2026 the honest status is still immature.
Key Takeaways
- On 25 September 2026 the honest status is still immature
- On 25 September 2026, agent-originated orders break the report your board already trusts
- The checkout OpenAPI snapshot 2026-01-30 says agents may include (with a first-touch value) on create or complete
- They do not define a last-click replacement you can drop into GA4
- You will double count | The July 2026 platform floor of about $791 per month at 50,000 sessions is an outbound agent run cost, from a different benchmark
Table of Contents
On 25 September 2026, agent-originated orders break the report your board already trusts. Last-click assumes a person clicked an ad or an email and then your site. A shopping agent can discover the SKU in ChatGPT or AI Mode, complete on your endpoint, and never produce that click.
Who this is for. An eCommerce leader who has been asked “what did the AI channel return?” The technical shape of checkout is agentic checkout architecture. This page is only the measurement limit.
Our take: store a source you control, at the moment the order is created. Do not backfill a channel in analytics later. Publish unknown when you do not have the flag. We have no client revenue number to put in its place. There is no published AI-agent case study here.
What still works
These do not require a new protocol:
- Orders created by your UCP or ACP complete endpoint. You saw the call. Tag the order
source=acporsource=ucpin the OMS when that handler runs. - A merchant-defined campaign parameter on a URL you issued, if the shopper still lands on your site.
- A count of feed errors and checkout errors. That is operations, and it is more honest than a ROAS cell.
Industry writeups sometimes quote a referral share from Adobe Analytics. We do not repeat a percentage here unless it is calculated from your logs. Referral share is not incremental revenue.
What the specs actually give you
ACP. The checkout OpenAPI snapshot 2026-01-30 says agents may include affiliate_attribution (with a first-touch value) on create or complete. The complete description says attribution is stored with the order and not returned in the response. So:
- You can persist it if you choose to.
- The agent will not read it back from you.
- It does not allocate credit across Google, email, and ChatGPT.
- A missing object means unknown, not “direct.”
UCP. Google’s merchant docs, reviewed the same day, describe Buy buttons in AI Mode and Gemini for select merchants. They do not define a last-click replacement you can drop into GA4. If your complete handler runs, tag that order yourself. If the shopper only saw you mentioned and bought later on the site, you do not have a protocol event.
MCP. No attribution model. A tool call is not a sale.
A measurement sheet that survives finance
| Question | Allowed answer |
|---|---|
| Orders whose create path was our ACP or UCP handler | A count from the OMS |
| Revenue on those orders | Sum of those order totals, after your usual cancels |
| Share of all orders | That count divided by all orders, with the definition of “order” written down |
| Incremental versus what would have been bought anyway | Unknown, unless you ran a test you can describe |
| Last-click credit for the same orders | Do not add it on top. You will double count |
The July 2026 platform floor of about $791 per month at 50,000 sessions is an outbound agent run cost, from a different benchmark. Do not put it in this table as “AI channel profit.”
Failure mode
What broke — A launch review subtracted “ChatGPT” sessions in analytics from paid search and called the difference incremental. The user-agent rule matched in-app browsers that were not ChatGPT. Detection: the segment spiked on a weekend when no feed refresh ran. Fix: delete the segment; keep only OMS flags set by the checkout handler. Lesson: a referrer rule is not a source of truth.
If you only do one thing
Add one enum on the order: web, acp, ucp, unknown. Set it in the handler that creates the order. Leave it unknown everywhere else.
What to do this week
- Ask whoever owns checkout whether the complete handler can stamp a source. If nobody owns that handler, you are not in either program yet — stop the ROI slide.
- Read ChatGPT shopping and Gemini shopping for what “in the program” means.
- Refuse a vendor who offers a percentage lift without the denominator.
- Keep catalog work moving. Attribution debt is not a reason to ship bad attributes. The feed post is the AI-ready catalog.
- Prepare the commerce platform when you want the stamp designed with the endpoint, via agentic commerce readiness.
What this post doesn’t cover
- Media-mix models. We are not fitting one.
- Marketplace settlement reports (Amazon, Walmart). Different ledgers.
- A recommended attribution vendor. We do not have a measured bake-off.
Frequently asked questions
When should you NOT add an agentic channel to your ROI model?
What could go wrong if we book ChatGPT as the converting channel because the user agent says so?
Does the ACP affiliate_attribution object replace GA4?
What could go wrong if finance demands a payback month for UCP?
Should the marketing team own this or engineering?
Conceptual bounds
Bounds for agentic commerce attribution
Level 1 — conceptual. Risk low. Oversight: bounded autonomy.
Say what a merchant can actually measure when an external agent influenced an order.
- Starts when
- Someone asks for attribution of AI-shopping orders.
- Tools
- This page does not grant tools. A later build still needs a named catalog, and anything unlisted stays unreachable.
- Stops for a person
- Measurement does not authorize a new checkout or a store-agent write.
- Checked by
- A metric is included only when the source can see it. Unknown influence stays unknown.
- Untrusted data
- Analytics snippets that instruct a model to claim credit.
- If it fails
- Retry a timeout at most twice. Back off on rate limits. After repeated verification failure, escalate. Stop when the budget is exhausted or a permission is denied. No unbounded loop.
This page describes a bound. It is not a production agent and not a published deployment. The shared contract is the AWS store-agent architecture. Permissions and data boundaries are in securing store agents. The commercial page is agentic commerce.
AWS Cloud Architect & AI Expert
AWS-certified cloud architect and AI expert with deep expertise in cloud migrations, cost optimization, and generative AI on AWS.
