Google announced its dedicated generative AI performance reports on June 3, 2026, and marked the worldwide rollout complete on August 31. The Search report gives you a separate view of impressions in Google’s AI features. Those impressions remain part of overall Search performance. Google’s announcement.
For a service business, the useful question is which page deserves attention and why. A rising headline total can hide falling exposure in the market you serve. A download can also turn an unavailable value into a misleading zero.
On October 8, we opened the report in SEO Informatica’s Search Console property and downloaded its CSV export. That download arrived as a ZIP containing Chart, Pages, Countries, Devices and Filters files. It contained no clicks column or query table. The interface still displayed a Beta label. This is a bounded observation of our property on that date, not a promise about every account or future export.
This guide also follows documentation checked on October 8, 2026. The exercise uses invented figures, separate from our account and provider facts. We have not tested aggregation through controlled search activity or identified which AI answer produced an enquiry.
Know which report you are reading
The Generative AI performance report for Search covers AI Overviews and AI Mode. Its documented metric is impressions, grouped by page, country, device and date. It excludes Search Labs experiments; Discover has a separate report. Google’s report definition.
An impression concerns a displayed link. It is not proof that someone visited, read the page or chose your business. The report also does not expose a transcript of the answer surrounding that link.
| Question | What this record can support |
|---|---|
| Which pages received reported AI Search exposure? | Inspect the Pages view under a recorded scope. |
| Did exposure change in our target market? | Compare the same country and period settings. |
| Which exact question triggered the appearance? | Not established by the report’s documented dimensions. |
| Did AI generate a visit or suitable enquiry? | Requires separate evidence; an impression is insufficient. |
A normal Web query report can suggest questions to investigate, but joining it to AI impressions by page and date does not recover the missing AI query. The shared page is not an event identifier.
Preserve the view before interpreting the number
Open the report for the intended Search Console property. Save the report name, dates, country, device, search type and any URL filter with the export. Keep the unfiltered property view separate from subsequent page investigations.
Use complete, comparable periods. The daily reporting basis is Pacific Time. Newest data can be preliminary, so record that state before treating a change as final. A selected URL changes the AI chart’s aggregation from property to URL. Report configuration and counting.
Our worksheet asks for the visible value, exported value and accepted working value separately. That distinction matters because Google converts ~ and - to zero in exports. A bare exported zero cannot tell you which condition produced it. Retain the original view or a note from it. Google’s export guidance.
Record an unknown as unknown. The worksheet is a local record you fill in; it does not connect to Search Console, reconstruct missing values or validate your account.
Download the blank interpretation record, the worked fictional records and the exercise and answer key.
Work through a misleading headline
Synthetic exercise. Every business detail, URL, figure and interface observation below is invented. These are not real account exports, a Google sample dataset or measured results.
Imagine a payroll consultancy serving US employers. Its fictional property is payroll.example. Period A is September 7–20; Period B is September 21–October 4, 2026. Both contain 14 days with the same weekdays. For this exercise, the observations are complete, use Pacific Time, cover all devices and use Web text-based search.
The first saved view has no URL or country filter:
| Property-level view | Period A | Period B |
|---|---|---|
| Generative AI Search impressions, all countries | 120 | 180 |
| Same report, United States | 90 | 60 |
| Same report, United Kingdom | 20 | 70 |
| Same report, Canada | 10 | 50 |
| Ordinary Web impressions, all countries | 2,000 | 2,400 |
The headline AI total rises by 60, or 50%. US impressions fall by 30, or approximately 33.3%. In this constructed record, the country rows cover the whole total: the gain comes from outside the business’s stated market.
The defensible reading is that reported exposure increased overall while decreasing in the United States. It is not evidence that more suitable employers discovered the consultancy. Country describes where the search originated, not whether the person was a qualified buyer.
The reviewer should inspect the US page records before recommending more international content or declaring the existing work successful. A smaller commercially relevant segment can change the decision behind a larger positive total.
Do not repair the chart by adding page rows
The same fictional property has these global page rows:
| Page | Period A | Period B |
|---|---|---|
/services/payroll/ |
70 | 40 |
/guides/payroll-checklist/ |
80 | 140 |
/case-studies/manufacturer/ |
20 | 30 |
The shown rows sum to 170 and 210. The property chart shows 120 and 180. That difference is compatible with distinct aggregation, not proof that one view is broken. Multiple pages can receive impressions when the property is counted once. Table coverage limits are another reason not to substitute a row sum for a chart total. Google’s aggregation explanation.
Keep both records with their labels. Do not force them to agree, and do not infer the exact number of co-appearances from their difference. These aggregate views do not identify which pages appeared together in individual answers.
Nor should you add the AI total of 180 to the ordinary Web total of 2,400. That would report 2,580 from overlapping measures. The ordinary Web total remains 2,400.
Narrow the decision to a service page
The reviewer now selects the payroll service URL and United States in both periods. This is a new saved scope, not a replacement for the property record.
| US payroll-page record | Period A | Period B |
|---|---|---|
| Generative AI Search impressions | 60 | 30 |
| Ordinary Web clicks | 8 | 9 |
AI impressions for the selected service page fell 50%. Ordinary Web clicks increased by one. Neither finding cancels the other, and dividing nine clicks by 30 AI impressions would not establish a 30% AI CTR.
Google counts external clicks from AI Overviews and AI Mode under its Search measurement rules. That does not make this page’s ordinary Web clicks exclusively AI clicks. Google’s click definitions.
The proposed action is a focused review: confirm the reported canonical URL, inspect ordinary page-filtered query evidence and check whether the page still answers the intended payroll-buying question. Most Search Console performance data is assigned to Google’s canonical URL, which may differ from a duplicate or the URL an editor expected. Page and query grouping.
The reviewer can then identify a specific issue worth correcting, or retain the cause as unresolved. The numbers alone do not justify rewriting the service page, creating another page for the same intent or blaming an update.
Keep an unavailable value out of the calculation
One additional fictional page record displays ~, while its downloaded cell contains 0. The worksheet retains both and leaves the accepted numeric value blank with status unavailable.
This row cannot establish zero exposure or a 100% decline. If the original view was not retained, mark the meaning of the zero unresolved and recheck the same view. Do not quietly convert missing evidence into a negative result.
The same principle applies when the entire report is absent. Google’s current explanation points to insufficient impressions and possible exclusion from the covered features. Record availability before investigating performance. Report availability.
The separate Search generative AI control has inclusion and exclusion settings, with possible inheritance from a parent property. Inspect its effective state if relevant; changing it merely to test reporting could change eligibility. Google’s control documentation.
Finish with a decision someone can review
The completed exercise produces three useful conclusions: the global increase does not establish growth in the target market; the US payroll page deserves investigation; and the unavailable row supplies no measurable result.
The workbook asks you to state the accepted observation, rejected inference, next check and owner. It also keeps missing clicks and queries explicitly unsupported rather than filling them with estimates. This is a report-reading exercise, not an ongoing citation monitor or lead-attribution system.
For repeated answer observations, use our monthly citation log. For received enquiries and business outcomes, use the GSC, GA4 and CRM measurement guide. For a wider decline, follow the traffic-drop diagnostic.
If the unresolved question is which page to improve, our AI SEO work connects the evidence to the service and market you want to reach. Start a review with the saved view and the decision you need to make.