Measure visits and conversions that can be attributed to AI products, then state what the analytics cannot observe. A customer may read an answer and later arrive through another route. That possibility does not make every unattributed visit an AI referral.
Capture the referral traffic properly
Review source and referrer data for known assistant domains. If useful, create an analytics group for those sources and document its rules. Test the rules against actual visits, since hostnames and reporting behavior can change.Google Analytics Help, Custom channel groups, which walks through the regex-based channel definition and warns that channel order decides which rule captures the traffic. The pattern list needs updating as assistants launch and rename.
Missing referral information limits attribution. Native apps and privacy controls can affect it. Review landing pages and conversions for identified visits, while avoiding the assumption that a visit proves how a page appeared in the answer.Cloudflare, the crawl-to-refer analysis on Radar, including the caveat that app-originated visits without a Referer header are invisible to referral counting.
What click data cannot establish
Pew’s March 2025 browsing study found traditional-result clicks on 8% of Google visits with an AI summary and 15% without one. Source links within summaries were clicked on 1% of those visits. These observations describe that sample; they do not measure the resulting sales influence.Pew Research Center, Google users are less likely to click on links when an AI summary appears: browsing data from March 2025, 900 US adults.
Cloudflare’s crawl-to-referral ratios measure requests relative to attributed visits. A high ratio does not establish that every crawled page influenced an answer or a purchase.The same Cloudflare Radar analysis; ratios for the June 19-26, 2025 period, ranging from Anthropic's 70,900:1 down to Mistral's 0.1:1, with the app-referrer caveat meaning true ratios are somewhat lower.
Additional indicators
Use additional indicators with their limitations stated:
Compare branded search, direct visits, and customer-reported discovery sources with visibility measurements. Each can be affected by other marketing and customer behavior. A rise in direct visits alone cannot identify AI as the cause.
If estimating an unobserved contribution, state the assumptions and show how the estimate changes when they change. Multiple correlated indicators do not automatically provide independent confirmation.
Presenting the evidence
Separate the report into observed referrals and conversions, correlated changes, and estimates. Name the data source and limits for each. Connect spending to completed work without presenting an untested causal chain as fact.
For changes in answer visibility itself, use Auditing Visibility Problems.
Keep impressions, sampled answers, and visits separate
Google now documents a Generative AI performance report for AI Overviews and AI Mode. Use it for the impression data it exposes. Do not infer AI-specific clicks or conversions from an impression total.Google Generative AI performance report. Lists the included features and available report dimensions.
Maintain three records: provider-reported impressions, results from your repeated prompt tests, and visits or conversions with identifiable sources. Their totals describe different populations and should not be added together.
Check an analytics grouping before using it
- Export recent session sources and referrers. Keep the original values.
- Choose the assistant domains actually present in that export.
- Create a custom channel rule for those sources. Check rule order so a broader referral rule does not take precedence.
- Test a known matching source, an unrelated source, and a source with missing attribution.
- Compare the grouped total with the matching original rows. Record when the rule changed.
For example, 12 identified assistant referrals and 3 completed enquiries are observed outcomes in that group. They do not establish that every enquiry was new business caused by the assistant. This example illustrates reporting, not an OppAlerts result.