Google Search Console’s AI Report Is Global. Build the Measurement Baseline First
Google now reports generative-AI impressions worldwide. Use this field-by-field worksheet to measure visibility without mistaking it for traffic.
On August 31, Google updated its Search announcement to say that Search Console’s generative-AI performance reports and control had reached all websites worldwide. Open the official help pages, however, and they still offer “not all properties have access” as one reason a report may be missing. Too few impressions—or an exclusion setting—can also leave a property without a chart.
That tension makes access status the first data point. An absent report is not a zero, and a visible report does not reveal visits. The current Search report counts impressions from AI Overviews and AI Mode; the separate Discover report counts impressions from generative features in Discover. Neither report publishes clicks.
A defensible plan establishes a baseline before anyone edits pages or changes the property control. The worksheet below aligns the current fields with ordinary Search Console and site analytics, while preserving the conclusions those joins still cannot support.
Worldwide rollout still has an access threshold
Google’s announcement was published on June 3 and updated on August 31 with a worldwide-rollout note. The Search report help page and Discover report help page carry matching worldwide notes.
Those same report pages retain rollout language under “Not seeing the report?” They also say a site may not have enough generative-AI impressions to qualify. Search Engine Journal independently confirmed the worldwide update and this unresolved documentation mismatch on August 31.
Record one of four states rather than filling a missing chart with zero:
- Available with data: export the chart and table, then record the latest complete date.
- Visible without enough data: record the interface state and wait; do not treat a blank table as measured absence.
- Unavailable: record the date, property type, and account checked. The documentation does not let an operator distinguish rollout lag from an eligibility threshold.
- Excluded: verify the property and inherited control before interpreting the missing impressions.
Small sites are most exposed to this ambiguity. “No report” and “the report measured zero impressions” are different observations, and the current interface may supply only the first.
Search and Discover count different events
The word impression hides two measurement systems. In the Search generative-AI report, an impression means a link to the site was shown in AI Overviews or AI Mode. The property chart counts one impression when multiple results from the same property appear in one generative result. A table grouped by page instead aggregates each page separately, so adding page rows may not reproduce the chart total.
Discover has its own generative-AI report. There, a link must be scrolled into view, and only one impression is counted for a result in a session even if the person scrolls away and back. Every Discover impression is aggregated by page. Daily values below a threshold can disappear from the chart while remaining in the selected period’s total.
| Report | Included surfaces | Available dimensions | Counting detail that changes the analysis |
|---|---|---|---|
| Search generative AI | AI Overviews and AI Mode; Search Labs experiments are excluded | Page, country, date, device | The chart aggregates by property; a page-grouped table aggregates by page. Newest dotted-line data can still change. |
| Discover generative AI | Generative features in Google Discover | Page, country, date | The link must enter view; one impression is counted per result per session. Daily threshold omissions remain in the period total. |
Keep these exports separate. A Search impression and a Discover impression do not describe the same viewing condition, and neither supplies a click denominator. Combining them into an “AI click-through rate” would attach ordinary-site clicks to exposures that the reports cannot connect.
Give every field one permitted question
The current reports are most useful as a page-discovery diagnostic. Each field can answer a narrow question, provided the worksheet also states what remains unknown.
| Field or view | Question it can answer | Decision it can support | It cannot establish |
|---|---|---|---|
| Access state | Did this property expose the report on the observation date? | Start a baseline, wait, or inspect the inclusion control | That a missing report means zero visibility |
| Property chart | Did total reported exposure rise or fall over a complete period? | Select a period for page-level inspection | Which page, query, or link placement caused the movement |
| Page | Which canonical URLs accumulated reported impressions? | Freeze a cohort for comparison; inspect outliers | Why Google selected the page, which passage grounded an answer, or whether anyone visited |
| Country | Where the Search request originated, or where Discover content was viewed | Check whether movement is concentrated geographically | The user’s language, identity, or commercial value |
| Device in Search | Whether reported Search exposure came from desktop, tablet, or mobile | Check for a device-skewed change | Layout quality, click behavior, or the device mix in Discover |
| Date | When the report recorded exposure, in Pacific Time | Align complete before-and-after windows | That a page edit caused a same-day change |
| Property inclusion control | Whether the property is eligible for the covered AI surfaces | Preserve a stable baseline or document an intentional exclusion | AI-training permission or ranking in other parts of Search |
The page field is the practical join key, with care. Search generally assigns data to the final linked URL after redirects and to its canonical URL. Discover identifies the canonical source page, which may differ from the URL a user would land on after clicking. Normalize each export to the canonical page identity before comparing it with analytics routes; preserve the original URLs so redirects and locale variants remain auditable.
Freeze the measurement plan before the page changes
A useful baseline states the hypothesis, page cohort, clocks, and failure condition before anyone sees the result. Twenty-eight complete days before and after an edit can balance weekday effects for many publishers, but it is a planning default, not a universal minimum. Low-volume sites may need a longer window, and time-sensitive pages may need a shorter one with an explicit limitation.
Use this worksheet once for each planned intervention:
| Worksheet field | What to record before the intervention | Why it belongs in the record |
|---|---|---|
| Question | One claim, such as “adding primary evidence and a concise definition will increase reported AI Search exposure for this fixed page cohort” | Prevents a general traffic story from being written after the numbers arrive |
| Property and control | Exact Search Console property, property type, include/exclude/inherited state, and observation timestamp | Exposes a control or inheritance change that would invalidate the baseline |
| Frozen page cohort | Canonical URLs and a stable content identifier such as post_slug; include locale routes deliberately | Stops high-performing pages from being added after the result is known |
| Intervention | Exact pages, material change, publication timestamp, and deployment identifier | Separates an evidence or structure change from unrelated releases |
| Windows | Complete start/end instants, Search Console’s Pacific Time dates, analytics property timezone, and excluded washout days | Prevents partial days and mismatched clocks from manufacturing movement |
| AI Search exposure | Property-chart total plus page, country, device, and date exports | Preserves both property and page aggregation instead of summing incompatible totals |
| AI Discover exposure | Period total plus page, country, and date exports | Keeps Discover’s viewability and threshold rules separate from Search |
| Ordinary Search | Page-level impressions, clicks, click-through rate, position, and relevant query mix from the standard Performance report | Shows whether conventional discovery moved during the same window |
| On-site behavior | Landing-page sessions, page_view, a defined meaningful-read event, engagement duration, continuation, and the traffic source available in site analytics | Tests whether the page earned and rewarded visits without labeling them AI-referred |
| Decision threshold | Minimum meaningful change, acceptable downside, and the result that would falsify the hypothesis | Makes “no decision” an allowed outcome when the sample is too small or mixed |
| Confounders | Indexing changes, seasonality, campaigns, template releases, outages, news spikes, and other edits | Prevents a coincident movement from being assigned to one page change |
post_slug and article_read in this worksheet are example publisher-defined fields, not built-in Google metrics. A site could instead use a canonical URL and its own meaningful-read event. Google Analytics’ traffic-acquisition report can provide source/medium, sessions, event counts, and average engagement time, but its units and property timezone still need to be documented.
For the comparison itself, export each source independently. Join the fixed cohort by canonical page identity and align the same complete instants. Do not perform a row-level join between an AI impression and an analytics session: the current generative-AI report provides no click or shared request identifier that could make that link.
A worked result can stay useful without claiming attribution
Suppose a publisher freezes four evergreen guides, adds primary citations and clearer definitions on September 1, and compares two complete 28-day windows. The following numbers are fictional and show the reasoning, not a benchmark:
| Measure for the frozen cohort | Baseline | Observation | Defensible reading |
|---|---|---|---|
| AI Search page impressions | 120 | 210 | Reported Search exposure increased for these canonical pages. |
| AI Discover page impressions | 30 | 24 | Discover exposure moved in the opposite direction under its separate counting rules. |
| Ordinary Search impressions | 8,000 | 8,100 | Conventional exposure was broadly similar. |
| Ordinary Search clicks | 320 | 316 | The ordinary Search report shows no matching click gain. |
| Analytics landing-page sessions | 390 | 385 | Measured visits were broadly similar; the source does not identify AI-surface clicks. |
| Meaningful reads | 250 | 270 | More read events were recorded, but the table cannot assign them to the AI impressions. |
The narrow conclusion is that reported AI Search visibility rose while the other observed series did not move in the same direction. The page revision may have contributed, but the report cannot identify the triggering queries, competing pages, placement, or clicks. A Google product change, demand shift, or different query mix could produce the same pattern.
The next action depends on the predeclared hypothesis. If the goal was greater AI exposure with no material loss in ordinary discovery or reading, the result may justify keeping the revision and repeating the measurement. If the goal was more visits, the evidence is insufficient. Calling 90 additional impressions “90 more opportunities that drove engagement” would smuggle both an unobserved click and an unobserved cause into the result.
This discipline is the same one required in a model evaluation: a metric is useful only when the version, population, and acceptance rule stay visible. When the underlying artifact can change, the reproducibility budget also needs a stop condition and an immutable record.
A control change needs its own observation window
The Search generative AI control applies to AI Overviews, AI Mode, and generative features in Discover. Inclusion is the default. Excluding a property removes its links and content from those features, so it gives up their traffic and impressions. Google says the choice is not used as a ranking or inclusion signal elsewhere in Search and does not control AI training; Google-Extended is the separate training control.
The setting can also be inherited from a parent property. Google says a change generally takes a few days to go live, followed by exclusion within one or two days, although caching and propagation can take longer. That lag, combined with an impression-only report, makes rapid toggle-on/toggle-off testing a poor experiment.
Keep the control stable through the baseline unless the business has independently decided to change participation. If it changes, record the property, inherited state, request time, observed effective date, and a washout period. The resulting before-and-after series can document the operational effect; it cannot recover the clicks or revenue that might have occurred under the other setting.
Google says it will add metrics over time. A click field, query breakdown, or surface-level segmentation would materially change this worksheet because each would remove a current inference gap. Until then, the defensible unit is modest but useful: a dated record that a fixed page cohort appeared more or less often in a defined generative surface, compared with separate measures of ordinary discovery and on-site behavior. That is enough to choose the next investigation. It is not enough to rename visibility as a visit.
Sources
- Google: New opportunities, control and insights for website owners
- Google Search Console Help: Generative AI performance report (Search)
- Google Search Console Help: Generative AI performance report (Discover)
- Google Search Console Help: Search generative AI control
- Google Search Console Help: Performance report overview and setup
- Google Analytics Help: Traffic acquisition report
- Search Engine Journal: Google Search Console AI Reports Rolled Out Worldwide