What does an AI impression tell you about an affiliate page? It records exposure to a link. To assess what happened afterward, you need evidence of a visit, a referral or a transaction.
Start by deciding which of those questions you are answering. Then keep the observation and the conclusion beside each other, so a visibility change does not quietly become a revenue claim.
What Google’s reporting covers
Google announced dedicated generative AI performance reports on 3 June 2026; its later note confirms worldwide rollout on 31 August. The Search report covers impressions in AI Overviews and AI Mode. Search and Discover have separate reports, with page, country and date views; device views apply to Search. This data already appears in overall performance reporting, so adding it to overall totals would double-count it.
The Search report’s help page documents impressions, not dedicated AI clicks, CTR, conversions or an AI Overviews-versus-AI Mode breakdown. Low impression counts can hide the report; Search Labs experiments are excluded. Page data generally follows canonical URLs, dates use Pacific Time, and page/property aggregation can produce different totals. Newest data may be preliminary. Export converts unavailable ~ or – values to zero. Preserve their original meaning before interpreting the file; an unavailable observation is not a measured zero.
These limits matter before you compare two periods. A calculation cannot repair observations that describe different things.
Build an observation-to-action record
The following ledger is Affrate’s recommended working method. Use one record for one question, such as whether a particular group of review pages warrants closer inspection.
Comparison scope
What to write: Report, property or page, periods, filters, time basis and collection time
Decision it supports: Whether the observations are comparable
Observation
What to write: Reported value and a separate availability/status field
Decision it supports: Whether there is usable evidence to interpret
Editorial context
What to write: Relevant page changes, their dates and the reason for each
Decision it supports: Which explanation deserves investigation
Outcome evidence
What to write: Separately observed visits, referrals or transactions, with their own scope
Decision it supports: What evidence exists beyond visibility
Next action
What to write: One question, an owner and a review point
Decision it supports: What to check before making a stronger claim
Choose the scope before looking for an interesting movement. Keep the same scope for the comparison and record any necessary change. Retain the original export; put your interpretation in a separate column.
If you cannot recover what an ambiguous exported value originally represented, mark that part of the comparison unresolved. Do not let spreadsheet arithmetic decide what the missing observation meant.
Also write down relevant page changes before deciding what caused a movement. A revised comparison table, a changed offer and a new title are different interventions. Recording them helps the next reviewer ask a specific question.
Read a hypothetical result carefully
Suppose a publisher compares two complete periods of equal length, using consistent filters and aggregation within each report. Recorded AI impressions rise from 1,000 to 1,300, while overall Search clicks fall from 90 to 70. These are invented figures for explanation, not Affrate results.
The defensible observation is that recorded AI visibility increased while overall Search clicks decreased. The figures do not identify which surface produced those clicks, explain why either measure changed, or prove that a page edit caused the movement.
Dividing overall clicks by dedicated AI impressions would mix populations and would not produce an AI-specific CTR.
Now suppose an affiliate dashboard records an approved transaction during the second period. A matching page and reporting period still do not identify that customer’s search journey. Any claim connecting the transaction to AI exposure needs evidence that supports that connection.
The useful next step is a narrower question: which page changed, what changed on it, and what separate evidence could help explain the observation?
Keep multimodal search labelled
Google’s 24 September multimodal reporting announcement covers Lens, Circle to Search on Android, uploaded images and Chrome image search. Global rollout began on that date; visible metrics depend on relevant activity. A multimodal search is not, by itself, proof of an AI citation.
If you analyse that filter, give it its own labelled comparison. Do not use a change of filter to make two otherwise incompatible periods look continuous.
Let the finding determine the next check
Choose a page-level question you can investigate: is an important condition hard to find, is the evidence current, or does the page answer the reader’s decision clearly? Record the proposed correction and its reason. Treat any later movement as another observation, not automatic proof of the correction’s effect.
If the next action is a content revision, Affrate’s source-first iGaming AI playbook provides the separate research and editorial-check workflow.
Google’s optimization guidance retains established SEO principles and requires no special AI markup or llms.txt for Google Search. Eligibility does not guarantee visibility. That guidance concerns Google, not every AI service.
A useful measurement record leaves you able to explain what was observed, what remains unknown and why the next check follows. That is enough to support a careful editorial decision without claiming a result the evidence cannot show.
Sources
Sources checked: 6 November 2026.
