Skip to content
Universal Commerce Protocol Independent observatory / en
The UCP standardUCP vs ACPImplementAgentic commerceFramework
ConceptGEOllms.txtGlossaryAgentic paymentsAP2Visa and MastercardB2B agentic commerceMethodologyUse CasesWhy NowFor WhomManifestoFAQResourcesSourcesEcosystemValidatorAboutContact
FR Get in touch →

Reference note

Measuring agentic commerce beyond your website

Why classic attribution undercounts AI agents, what GA4 actually captures, and the protocol signals a merchant should instrument server-side.

Published . Part of agentic commerce.

Agentic commerce breaks an assumption twenty years of web analytics rest on: that the purchase journey passes through pages of your site. When an agent negotiates and pays over a protocol, there may be no page view, no tag loaded, no session. The sale exists, the trace does not.

This note separates what is measurable today, what is not, and what a merchant can instrument alone without waiting for tooling to catch up.

Why classic attribution comes apart

Three distinct mechanisms stack up, and it is worth not conflating them.

The referrer-less visit. Some AI surfaces do not pass a Referer header, or mask it. The visit exists, it lands in “Direct”. This is not a tooling error, it is an absence of information at source.

Off-site conversion. This is the case specific to agentic commerce: the transaction completes inside the agent’s surface, through protocol endpoints. No JavaScript tag runs, because no page is served. Client-side analytics can structurally see nothing.

Indirect influence. A user asks an assistant, remembers a brand, then returns later through a branded search. The conversion gets attributed to branded organic. The assistant did the work, another channel takes the credit.

Only the first mechanism is a configuration problem. The other two are problems of nature: no GA4 setting resolves them.

What GA4 captures, and how to capture it properly

Some AI surfaces do pass a referrer. Isolating them is the starting point, and it is the only direct measurement available. The method documented by Semrush is to create a channel grouping based on a regular expression applied to the source:

.*(chatgpt\.com|chat\.openai\.com|openai\.com|perplexity\.ai|claude\.ai|
gemini\.google\.com|bard\.google\.com|copilot\.microsoft\.com|deepseek\.com|
mistral\.ai|grok\.com|x\.ai|you\.com|search\.brave\.com).*

Two precautions when installing it.

Rule order matters. If “Referral” is evaluated before your AI channel, that traffic falls into the wrong bucket and your new channel stays empty.

The result is a floor, not a total. This grouping measures the surfaces that pass a referrer, not AI’s real contribution. Estimates of the gap circulate among several vendors, with widely differing methodologies and scopes; they are not a transposable benchmark. The defensible reading is directional: this curve shows a trend, not an amount.

Protocol signals, the merchant’s blind spot

This is the point analytics guides do not cover, and the one where a UCP merchant has most to gain, because nobody else can instrument it.

The UCP specification defines negotiation failure codes that do not surface in the usual HTTP status codes. Two deserve a dedicated counter:

  • capabilities_incompatible answers HTTP 200. An agent came, negotiation found nothing in common, the sale did not happen. Any monitoring that counts HTTP statuses will file this commercial failure under technical success.
  • profile_unreachable answers HTTP 424, a failed dependency. The symptom of an unreachable manifest will therefore not appear in the 5xx of your own domain.

Counting these two codes, per day and per calling agent, produces a metric nobody else can produce for you: the volume of encounters that failed before ever becoming a transaction. The detail of these codes is in the note on the UCP manifest.

Four counters to put in place

In order of feasibility, from most immediate to most structural:

  1. The AI channel in GA4, with the rule-order caveat above. Near-zero cost, directional value.
  2. Agent tokens in server logs, separating answer indexers from human-triggered fetches. Careful: some tokens are control tokens and never appear in logs, a point covered in the note on agent traffic.
  3. UCP negotiation codes, server-side, the only protocol-specific signal of commercial loss.
  4. Citation rate, meaning how often generative surfaces mention the brand across a query set tracked over time. Indirect, but the only measure that catches upstream influence.

The principle that holds

The natural reflex is to look for the single number that would replace the session. It does not exist yet. The tenable position for a leadership team is to carry three distinct readings rather than one misleading aggregate: a measured floor from GA4, a volume of encounters from the protocol, a presence indicator from citation. An aggregate built on data of different natures would produce false precision, which is worse than a partial measure openly stated as such.

For upstream preparation, see agent readability. For the wider view, see agentic commerce.

Sources

Back to agentic commerce · All notes · Lire en français