The monthly report shows more AI citations and fewer organic sessions. One team calls the programme a success because visibility has grown. Another questions the investment because traffic has fallen. The report contains two observations, but neither interpretation follows automatically.
AI search creates a measurement problem partly because influence and visits can occur in different places. A shopper can encounter a recommendation, buy through a retailer or search for the brand later. A citation can also be seen without affecting a decision. The reporting needs room for both possibilities.
What the click evidence tells us
In Pew Research Center's July 2025 analysis of browsing data from 900 US adults, traditional-result clicks occurred in 8% of visits with an AI summary and 15% of visits without one. This describes observed behaviour in that study. It does not show what would happen to a particular APAC brand's sales after a GEO programme.
The finding makes traffic an incomplete account of discovery. It does not make traffic irrelevant, and it does not turn every unclicked mention into an attributable conversion. An honest measurement plan accepts that some influence will remain unobserved.
Read three numbers without inventing a funnel
Imagine a monthly report with 18 accurate recommendations in 60 valid prompt observations, 35 identifiable AI-referred sessions and four purchases from those sessions. These numbers are hypothetical, but the interpretation problem is common.
18 out of 60 describes an answer sample
The 30% rate applies to a declared panel under recorded conditions. Its value depends on the commercial relevance of the prompts and the consistency of the testing. It is not an impression count and does not estimate how many customers saw the brand.
35 sessions describes observable visits
The session count depends on analytics configuration and referral classification. Visits with missing or ambiguous attribution require separate treatment. It would be wrong to assume that the 18 sampled recommendations produced these sessions, because the observations do not identify the same people or events.
Four purchases describes recorded behaviour
The orders can be analysed using the business's normal attribution and transaction rules. They still do not prove that the GEO changes caused four additional purchases. Some customers might have bought anyway. Others influenced by AI may have converted outside the measured route.
The dashboard should change the next decision
A visibility improvement with persistent factual errors points towards product information work. Better answers followed by poor landing-page performance points towards the offer, availability or purchase experience. A rise in AI referrals during a major promotion requires a discussion of both activities.
Microsoft's Citations Overview explicitly separates citation activity from traditional rankings, impressions and click-through rates. Its AI Performance reporting update adds further context within its reported experiences. A provider's data should retain its stated scope in the dashboard.
That scope discipline matters when combining tools. Two products may use different prompt panels, dates and definitions of a mention. Their scores cannot be averaged into a credible market-wide measure simply because both are labelled AI visibility.
When a stronger claim is worth testing
If investment depends on incremental commercial impact, design the comparison before the intervention. Comparable page groups or markets may help, where volumes and operations permit. Record pricing, media, stock and seasonality changes that could affect the result. A before-and-after chart can support monitoring, but cannot isolate those explanations on its own.
For an early GEO pilot, the defensible conclusion may be narrower: the team corrected important information gaps, observed persistent changes in a relevant answer sample and learned what to test next. That is useful evidence for investment without presenting uncertainty as revenue.