AI search attribution in GA4: connect referrals to outcomes

What GA4 can measure about AI search

GA4 measures activity collected by your site or application. An answer shown inside an assistant is outside that site's event stream unless a supported integration supplies additional data. A visitor who clicks a source link can produce a measurable session; someone who only reads an answer does not produce a site visit for you to count.

Google's current default-channel documentation^[1] lists AI Assistant and describes classification using the ai-assistant medium for recognized assistant referrers. The same documentation explicitly excludes Google AI Overviews and AI Mode from that channel. This is a current platform definition, not evidence that every AI-related journey is identifiable.

Begin with the definition in your report. Do not take an old chart labeled “AI traffic,” add a new channel and compare the totals as if the collection rules had never changed. Record the dimensions, channel definition, reporting period and any changes in tracking.

Start with the question the report should answer

Write one question before selecting dimensions. “Did AI influence our market?” is too broad for a session-acquisition report. “Which identifiable AI Assistant visits reached our integration documentation and completed an account registration?” has observable components.

Our recommended first report asks:

During this period, which identifiable assistant-origin sessions reached our site, where did they land, and what useful actions did those sessions include?

That question has boundaries. It does not cover every customer influenced by AI, every answer containing your brand or every later purchase by someone who once visited.

Select a period long enough to inspect the volume you actually receive, rather than promising a universal weekly benchmark. State the property timezone. Exclude known internal and test activity using your maintained analytics configuration, and record relevant collection changes. Avoid comparing a partial current day with a complete prior day.

Build a report for assistant sessions and completed registrations

Before opening GA4, confirm you have access to the correct property, know its timezone, and have a reliable completion event. The example below uses the recommended sign_up event for a successful account registration; a button click should not trigger it. A trial that begins at another stage needs its own precisely defined event. Google documents recommended events, but your implementation must send them.^[7]

1. Identify the acquisition cohort

Open Reports → Acquisition → Traffic acquisition. Choose a complete reporting period and the Session default channel group dimension. Filter it to AI Assistant, then add Session source / medium as a secondary dimension to inspect the underlying sources. Record the property, dates and filter with the export.^[2]

If your navigation differs or the report is missing, use the report search or ask an Editor to restore it. Do not substitute User acquisition: that changes the question to how users were first acquired. Keep the word “Session” in the dimensions you select.

2. Find the entry pages

Open Reports → Engagement → Landing page and apply the same period and session-channel filter. Keep Landing page as the row dimension and Sessions as the count. This identifies entry pages rather than all pages viewed later in the visit. If you add a source breakdown, retain its session scope.^[5]

Keep query-string variants separate until you understand them. An application callback and a marketing page may share part of a path while serving different purposes. Document any grouping applied in the exported report.

3. Select the completion you actually mean

Have the property owner mark the validated sign_up event as a key event if registration is the business outcome. This affects shared measurement, so use the existing event definition rather than adding a duplicate for this report. Record when the definition became active and compare periods using the same definition. In the acquisition report, select that event for Session key event rate, rather than leaving the metric on all key events. Select the same event for any Key events count you display. Google defines that rate using sessions in which the selected key event occurred divided by all sessions in scope.^[2] Its key-event guidance explains event selection and reporting.^[8]

Label the result “sessions with a completed registration.” It does not establish qualified trials, activated accounts or won revenue. If the interface cannot expose the selected-event rate in the required breakdown, use a separately reviewed exploration or analytics model and record its definition. Do not divide a repeated event count by sessions as a substitute.

4. Reconcile and save the report definition

Save two exports: acquisition by source and entry pages under the same filter. Confirm their scope before comparing totals. Record any row limits, (not set) values, sampling or thresholding notices. Keep incomplete or restricted data visible in the report note. Check whether later data processing changes the totals before freezing a management report.

The attribution worksheet holds the dimensions, completion event and reporting rules. Someone else should be able to rebuild the report from that record.

When a custom channel is useful

Start with Google’s default channel. A custom group is appropriate when you need a narrower business definition, such as an explicitly maintained subset of assistants, or want to analyze historical source rows classified elsewhere.

An Editor can open Admin → Data display → Channel groups, copy the default group and add a named channel. Define it from inspected source values, place it before every broader channel that could match those values, then validate the affected rows. Google applies the first matching channel and supports retroactive custom reporting.^[3]

For a deliberately narrow example, exact source conditions for chatgpt.com OR perplexity.ai could define “Selected assistants v1” if those are the values your property actually records. Keep other sources in their existing categories. This is a reporting subset, not a complete directory of AI services.

Inspect examples that should match and should not match before saving the definition. A broad rule containing “ai” could include unrelated traffic. A source-based rule also cannot recover an assistant visit whose source was never collected. Retain the default channel alongside custom results and avoid splicing the two series into one unlabeled trend.

Validate the path from the click to the landing page

Test collection before interpreting a low assistant-session count. Use a link from a real assistant response, follow it through the actual redirect and signup path, and keep the test conditions: surface, browser, device, consent state and destination.

Use Tag Assistant to inspect the configured Google tag and GA4 DebugView to confirm the expected page and completion events. Follow with a processed-report check for the session source and classification. DebugView confirms collection; it does not establish the final channel assignment. Google’s troubleshooting guidance distinguishes the tools and says standard reports can take 24 to 48 hours to process.^[6]

Label and exclude test activity through the property’s maintained configuration. Do not weaken consent settings or disable production filters merely to make a test appear. If debug traffic is intentionally excluded from reporting, use a controlled test property or an approved validation method for the classification check.

Google describes traffic-source handling as collection, processing and reporting.^[4] Preserve that separation in the incident record. Avoid putting personal information in campaign URLs or publishing individual customer journeys.

Join acquisition to the task, not only to a page view

A page view establishes that a page was viewed in the collected data. It does not establish that the visitor learned the answer, started a valid trial or became a qualified opportunity.

Choose an outcome the site can verify. For a developer product, that might be a completed account registration, a successfully created workspace or a qualified demo request. Define it precisely enough to distinguish a real action from a button click.

For example, “demo form submitted” and “demo accepted by sales” are separate stages. Record which system establishes each stage before presenting an acquisition-to-pipeline chart.

A useful reporting table might look like this:

Keep the acquisition definition and event selection with the table. Downstream lead and deal rows require the separate matching method described below.

Use explicit arithmetic when reporting a conversion rate

Consider a hypothetical cohort of 500 identifiable assistant sessions. Suppose 20 distinct sessions include at least one completed account registration.

Under the session-based definition above, the observed registration-session rate is 20 divided by 500, or 4%. It is not the number of registration events divided by sessions if repeated events inflate the numerator.

The counts are deliberately hypothetical. They illustrate the denominator problem; they are not EarShots performance data or an industry conversion benchmark.

Worked monthly report: what 500 sessions can tell you

This is a fictional, internally consistent report, not an EarShots result or a conversion benchmark. The property uses UTC, the period is a completed 30-day window, internal activity is excluded, and the selected completion is a validated registration event. Counts are distinct sessions under one saved channel definition.

The total rate is 20/500. Do not average the three row percentages, and do not silently remove the 50 login sessions to improve the result. A separate acquisition report that excludes existing-user entry pages needs an explicit rule and a new denominator. Round only after calculating from counts.

The executive note could read:

We recorded 500 sessions in GA4’s AI Assistant channel during the completed period. Twenty sessions included a completed registration, a 4.0% session rate. Documentation was the largest entry group. Login traffic is included and may reflect existing users. This report excludes unidentifiable AI influence and Google AI Overviews/AI Mode. Qualification and revenue remain separate downstream measures.

That note supports investigating documentation-led onboarding and the login cohort. It does not show that the documents caused registration or that assistant visitors outperform another channel. A comparison needs the same completion definition, period and treatment of existing users; small counts should remain visible alongside rates.

Connect downstream outcomes with an explicit matching rule

For customer relationship management (CRM) reporting, choose the matching method before calculating pipeline. A reviewed implementation might connect a consented first-party registration identifier to a lead, then link the lead to an opportunity. Record whether the rule uses the registration session, first known touch or another defined attribution window. GA4 does not automatically establish this business relationship.

Count each lead and deal once under the stated rule. Keep unmatched records and the reason they could not be linked. If six leads are qualified but only five can be linked to this cohort, report five linked qualified leads and one unmatched lead. Do not attach the unmatched record because its timestamps look close.

For multi-person buying accounts, keep lead-level and account-level tables separate. Define how reopened opportunities, multiple contacts and later channel visits are handled. A list of matched deals is a descriptive cohort outcome; it is not a causal estimate of revenue created by AI search.

Keep cited URLs and landing URLs in separate tables

An assistant might cite a technical reference and link a product page as the next action. A visitor may also arrive on a login page to use a product they already know. Neither scenario should be reported as proof that the landing page acquired a new buyer.

Maintain an answer-observation table containing prompts, test conditions and displayed source URLs. Maintain an acquisition table containing observed source dimensions, landing pages and outcomes. Compare patterns at the appropriate level, but do not join an anonymous session to a specific prompt just because the timestamps are nearby.

For technical teams, this comparison can reveal useful work. A heavily visited setup guide with many failed onboarding attempts deserves an implementation review. A frequently cited pricing reference with inaccurate plan language deserves a factual correction. Those decisions do not require an invented one-to-one answer-to-session link.

Our citation-source guide explains why cited domains and business results should retain separate meanings.

Ask about discovery without counting it as tracked acquisition

An optional discovery question can reveal experiences the analytics record does not capture. Ask broadly enough to avoid steering the response: “Where did you first hear about us?” Offer an open response or clear choices, including “Not sure.” If you need an AI-specific follow-up, ask which experience the respondent remembers.

Report the question, response options, number of respondents and eligible population. If eight of 40 respondents name an assistant, the result is eight of 40 respondents. It does not establish that 20% of all customers were acquired through AI.

Do not add those eight people to a session-attributed lead total without checking for overlap. A person can both click an assistant link and report discovering the product there. Two measurements of the same person are not two acquisitions.

Self-report can be informative even when it conflicts with tracking. Preserve the disagreement and investigate it. Memory, different interpretations of “first,” and unavailable source data can all affect the comparison.

What should an executive report say?

Separate what was recorded from what you think may explain it. A concise report can use four lines:

• Acquisition: identifiable assistant sessions, the exact classification, landing pages and period.
• Outcomes: distinct completed actions and downstream records linked under the documented method.
• Customer discovery: survey answers with respondent counts and the actual question.
• Answer observations: named brands, citations and correctness in the saved test set.

If you model an additional value estimate, show the assumptions and a range separately. Do not relabel the unexplained part of direct traffic as AI revenue. A before-and-after increase can coincide with email, product launches, press, seasonality or tracking changes.

A 30-day launch can establish whether distribution and collection work. Its duration and sample may still be insufficient to estimate a revenue effect.

A monthly attribution review template

Use answer monitoring alongside your analytics

Explore EarShots Prompt Universe for the questions and answer evidence your team wants to investigate. Keep GA4 and your downstream business records responsible for the acquisition and outcome measurements described here. The combination is useful when the team can explain what each source actually measures.

Method and updates

Documentation checked October 29, 2026. This article reports no original traffic study. All numeric examples are hypothetical. The report design is an EarShots editorial recommendation, not an automated EarShots attribution feature. Navigation and classification were checked against Google’s documentation; no authenticated customer property was used to validate this walkthrough. Menus and available breakdowns can vary by property and access level.

Sources

1. Default channel group. Google Analytics. The AI Assistant definition, classification and stated Google AI-feature exclusions.
2. Traffic acquisition report. Google Analytics. Session-acquisition reporting and its distinction from user acquisition.
3. Custom channel groups. Google Analytics. Custom classifications, source rules and channel ordering.
4. Campaigns and traffic sources. Google Analytics. Collection, processing and reporting of traffic-source data.

1. Landing page report. Google Analytics. Entry-page dimensions, navigation and report scope.
2. Verify and troubleshoot your Google Analytics setup. Google for Developers. Tag Assistant, DebugView and processing expectations.
3. Recommended events. Google Analytics. Account registration and other explicit event meanings; implementation is required.

1. About key events. Google Analytics. Key-event definitions, event selection and reporting.

Download the blank working records for all four guides.