How to See Customers Who Found You Through AI in Your Reports

Measuring AI-referred customers is the practice of tracking interest from AI assistants through separate records of referrals, branded searches and customers’ own accounts. Someone may discover your brand in ChatGPT, then search Google the next day. An Organic Search record identifies that visit’s channel, but cannot establish where the person first encountered your business.

Summary: Check GA4’s AI Assistant channel, monitor branded queries in Search Console and add an optional discovery question to your forms. Review these records together each month, keeping unknown sources separate. This distinguishes measured referrals, remembered discovery and search trends without treating them as interchangeable evidence.

What the study found, and what we conclude

The study observed more branded searches and visits following AI recommendations, but did not observe sales. Michael Iannelli and Alan Ai, both affiliated with Scrunch AI, matched consenting panellists’ browsing records with their ChatGPT, Claude and Gemini conversations. See Do AI brand recommendations bring visitors? for the detailed review.

The reviewed arXiv preprint is v2, dated 31 August 2026; v1 appeared on 9 June 2026. Table 3 compares historical windows with the following 7 days for recommendations where no brand interaction was observed in the preceding 7 days:

• Google searches containing the brand: 2.9% versus 7.2%, a difference of 4.3 percentage points.
• Own-site visits: 3.1% versus 5.5%, a difference of 2.4 percentage points.
• Brand-specific retailer-page visits: 0.8% versus 1.8%, a difference of 1.0 percentage point.

Denominators are eligible user-response-brand records, not unique customers, orders or all assistant users. These are percentage-point differences, not relative increases or successive conversion stages.

Data came from Scrunch AI’s commercial measurement pipeline; panel and subgroup sizes were withheld. The study was observational, without random assignment. Brand interest arising at the same time could influence both the recommendation and the visit, leaving causality unresolved. Despite “purchase” in the title, no transactions, revenue or sales were observed. Consumer brands in 2 English-speaking markets provide no ready forecast for Turkish local services or B2B.

DijitalPi’s interpretation: Referral clicks may miss part of the journey. Assigning every branded search to AI would introduce another measurement error.

Four layers of evidence: link, search, account and unknown

Keep AI referrals, branded search trends, customers’ accounts and unknown traffic separate. Each answers a different question and has different limitations. Preserving these distinctions in your monthly report helps prevent counting the same person repeatedly across records or mistaking a sign of interest for evidence of a sale.

Read these layers alongside each other; do not add them together. Direct is a channel classification, whereas (not set) can indicate a missing value in the selected dimension.

Check the AI Assistant channel in GA4

Start with AI Assistant in GA4’s default channel group. Google announced this channel on 13 May 2026, but you should still inspect how your property classifies individual sources. Read channel totals alongside source, landing page and form submissions to understand recorded visits and the measurable actions that followed.

1. Find the default channel. Check AI Assistant in a report using the default channel group. When referrers match Google’s assistant list, medium becomes ai-assistant and campaign becomes (ai-assistant). Google controls the list; do not assume universal coverage.
2. Inspect sources. Check Source values and their assigned channels, including Perplexity. Google’s announcement does not state whether the change applies historically. Before comparing periods preceding May 2026, check whether older sessions remain under Referral.
3. Create a custom group if needed. Go to Admin > Data display > Channel groups. “Create new channel group” starts from a copy of the default group. Select “Add new channel”, choose the “Source” dimension and “matches regex” operator.
4. Check matching and order. DijitalPi’s suggested narrower starting regex is chatgpt\.com|chat\.openai\.com|perplexity|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai. Test it against your Source list. Avoid broad patterns matching anything containing “ai”. Use “Reorder” to place your AI channel above Referral: traffic goes to the first matching channel. Select “Save group”. Standard properties allow 2 custom groups alongside the predefined group; custom groups can apply retroactively.
5. Follow the visit. Review landing pages and form submissions. Google Tag Manager setup can support form submission event measurement.

What referrer data carries, and what it does not

Referrer data can identify where a visit came from, but does not reveal the question someone asked an AI assistant. Source parameters on ChatGPT search links do not remove this limitation. Some transitions transmit no referrer at all, so source fields cannot stand in for conversation history or first discovery.

OpenAI’s publisher FAQ states that referral links in ChatGPT search include utm_source=chatgpt.com. Do not extend this statement to every ChatGPT click.

The default browser policy, strict-origin-when-cross-origin, sends only the origin, meaning protocol and domain, across sites, excluding path and query. With no-referrer, no referrer information is sent.

Apps or settings may prevent referrer transmission. Google AI Overviews and AI Mode were also excluded from the study.

Track branded search trends in Search Console

Tracking branded queries separately shows how interest in searching for your business by name changes over time. It does not prove an AI recommendation prompted those searches. Use the branded query filter or a brand-name regex, compare comparable periods and record campaigns, press coverage and seasonal influences alongside the results.

• Select the filter. In the Performance report, open the Query filter and choose “Branded” or “Non-branded”. AI-assisted classification may mislabel queries.
• Check availability. Sites with low impressions cannot access the filter. Google specifies a 16-month history starting March 2025.
• Prepare an alternative. If unavailable, filter using a regex covering your brand and common misspellings. Review matches.
• Explain comparisons. Use the comparison feature and Search Console annotations to record campaign and publication dates.

If branded clicks rise during a campaign, record that overlap rather than automatically attributing growth to AI.

Add a discovery question to forms and your CRM

An optional discovery question captures a customer’s account of a beginning that technical records may miss. Keep “Where did you first hear about us?” short and store the answer separately from measured source data. A Google enquiry can then retain an earlier AI discovery as a customer statement.

Use a free-text field or optional choices: search engine, AI assistant, social media, recommendation and other. You can ask the same question during sales conversations.

Keep these CRM fields separate:

• Last measured source: UTM or channel record.
• Reported first discovery: Where the person remembers first hearing about you.
• Reported AI tool and topic: Details they choose to share.

Memory may be incomplete. Keep the question optional and brief, and never replace technical records with the answer. CRM setup supports field organisation; offline conversion tracking setup supports matching relevant records when sales close away from your website.

An illustrative monthly report page

Your monthly report should show AI referrals, branded searches, customer statements and uncertain sources as separate rows on one page. Include a period value, comparison and interpretation for each. The purpose is to identify observed changes and further checks, while preserving the distinctions between these different kinds of evidence.

Do not fill the unknown with guesses

Visits with unknown sources should remain explicitly uncertain in your report. Without a verifiable record, do not allocate Direct traffic or branded searches to AI. Identify which evidence layer lacks information, then check technical source records and customer statements separately when you next review and update the report.

Ask during reviews:

• Did source classification or form measurement change?
• Were the same filters used in both periods?
• Is there a customer statement, or only a trend?

“Source unknown” preserves the limits of your evidence and makes the next check clearer.

A four-week setup plan

A weekly sequence lets you review existing measurement before organising search trends and customer statements. Bring the records together in a shared report at the final stage. This illustrative plan assumes access and necessary data are available; check each stage’s output and adapt progress to your website’s circumstances.

Use the first meeting to confirm consistent reporting and identify sources, pages and measurement gaps to investigate. Avoid premature conclusions about which source brings more customers.

Sources

The channel rules, filter features and referrer limitations described here draw on the official documentation below. The research link identifies the reviewed preprint version. Use product documentation for setup checks and the study’s scope for interpretation. DijitalPi’s regex, reporting template and weekly plan are illustrative implementation designs, not research findings.

• Research: From Prompt to Purchase, v2
• Google Analytics: What’s new
• GA4 default channel definitions
• GA4 custom channel groups
• Search Console branded query filter
• Search Console Performance report: branded queries and comparisons
• Google AI features and website measurement
• MDN: Referrer-Policy
• OpenAI: Publishers and developers FAQ