AI-powered configuration is an experimental Search Console feature that converts a natural-language request into proposed filters, comparisons, metrics, and table selections for the Search results Performance report.
It configures the report. It does not interpret the result, find a cause, rank opportunities, export data, or make changes to a website.
The feature is available to only a small percentage of Search Console users during testing. Each user is currently limited to 20 requests per day.
This guide reflects Google's documentation on July 18, 2026.
What the feature actually does
Instead of opening several filter dialogs manually, a user can describe the desired report state. Search Console proposes the corresponding configuration and asks the user to confirm it.
The supported configuration categories are:
| Category | Examples |
|---|---|
| Filters | Query, page, country, device, search appearance, date range |
| Comparisons | Date versus date, mobile versus desktop, one country versus another |
| Metrics | Clicks, impressions, average CTR, average position |
| Table selection | Queries, pages, countries, devices, dates, or supported search appearances |
For example, a request for “mobile queries containing sports in Spain during the last six months” can produce a mobile device filter, a Spain country filter, a query-contains filter, and a six-month date range.
The generated state uses the same Performance report controls available to the user. It does not create a new data source or unlock hidden rows.
Current availability
The feature is:
- Experimental.
- Visible to only a small percentage of users.
- Limited to selected websites during testing.
- Being expanded gradually.
- Restricted to the Performance report for Search results.
It is not available for the Discover or Google News Performance reports.
If the configuration icon is absent, Google has not documented a manual opt-in, markup requirement, or request form that enables it. A missing control does not indicate a property error.
Open and apply the configurator
- Open the property's Performance > Search results report.
- Select the AI configuration icon in the report header.
- Enter the request in the side panel.
- Review the proposed filters, comparison, metrics, and table.
- Confirm the suggestion to apply it.
- Inspect the final report controls again before interpreting the data.
The proposal is not applied silently. The confirmation step is the user's opportunity to catch an incorrect date, country, match type, dimension, or comparison.
Existing filters remain active
The configurator is aware of existing report filters and keeps them while applying new customizations.
Suppose the current report is already set to:
- Last 28 days.
- Mobile devices.
- Web search type.
If the user asks, “Show pages containing /blog/,” Search Console preserves the existing date and device state and adds the page filter.
This behavior is useful for iterative analysis, but it creates an important risk: a result can be narrower than the prompt alone suggests.
Before applying a proposal, inspect:
- Date range.
- Search type.
- Query and page filters.
- Country and device filters.
- Search appearance.
- Existing comparison.
- Selected metrics and table dimension.
When sharing a result, record the complete final state—not only the latest prompt.
Supported prompt patterns
Good requests describe a report configuration using explicit Search Console dimensions and periods.
Filter a query or page pattern
Show clicks and impressions for pages containing
/blog/case-studyduring the last six months.
Expected configuration:
- Page URL contains
/blog/case-study. - Date is the last six months.
- Clicks and impressions are selected.
Compare equivalent dates
Compare clicks for pages containing
/blog/in Q2 this year with Q2 last year.
Expected configuration:
- Page filter for
/blog/. - A custom date comparison using the two specified quarters.
- Clicks selected.
Check the calendar dates. “Quarter,” “week,” and “year” can be ambiguous when an organization uses a non-calendar fiscal year or a different definition of week.
Compare devices
Compare mobile and desktop CTR in Spain during the last 28 days.
Expected configuration:
- Spain country filter.
- Last 28 days.
- Mobile versus desktop comparison.
- Average CTR selected.
Select a query table
Show average position and clicks for queries containing common animal names during the last six months.
Google's documentation shows that the tool can translate a request like this into a query regex, select the period and metrics, and display the query table.
Review the generated regular expression. Search Console uses RE2 syntax and partial, case-insensitive matching by default unless the expression changes that behavior.
Write prompts that are easy to verify
A reliable request names:
- The metric.
- The row dimension.
- The required filters.
- The exact period.
- The comparison, if any.
For example:
Show clicks and impressions by page for URLs containing
/guides/, for January 1–June 30, 2026, compared with January 1–June 30, 2025, on mobile in the United States.
This is easier to audit than:
How are our guides doing?
The second request asks for an interpretation without defining “doing,” a period, metric, device, country, or comparison. Configuration tools work best when the intended controls are explicit.
The daily request limit
Each Search Console user is currently limited to 20 AI configuration requests per day.
Treat prompts as a limited convenience, not as a batch-query interface. To conserve requests:
- Decide the desired report state before prompting.
- Modify a nearly correct proposal manually instead of regenerating it repeatedly.
- Use ordinary filters for simple one-control changes.
- Save complex repeatable analysis in documented procedures or external queries.
- Do not use the feature as an automation endpoint.
The documentation describes the limit per user, not per property. Do not assume switching properties resets it.
It cannot analyze the data
AI-powered configuration does not answer analytical questions such as:
- Why did clicks fall?
- Which pages should be updated first?
- Did a migration cause the decline?
- Is CTR unusually low for the current position?
- Which query represents the best opportunity?
- Did an algorithm update affect the site?
It can configure a useful view for investigating those questions, but the user must evaluate the resulting metrics and context.
A prompt such as “Why did traffic fall?” lacks a supported report action. A better configuration request would be:
Compare clicks and impressions by page for the last 28 days versus the previous 28 days.
After the report is configured, analysis is still required to separate demand, visibility, CTR, page mix, device, country, and data-quality effects.
Unsupported actions
Google explicitly documents that the tool cannot perform data manipulation or other actions beyond initial report configuration.
Sorting
It cannot sort the table to show the highest or lowest values. “Show pages with the most clicks” requires the user to select the page table and then sort it manually.
Metric-threshold filters
It cannot apply direct table conditions such as:
- Clicks below 100.
- CTR above 5%.
- Position between 5 and 15.
- Impressions greater than 1,000.
Search Console's filter controls operate on dimensions, not arbitrary metric predicates. Export or query the data elsewhere for threshold analysis.
Pagination
It cannot page through the table or retrieve every row.
Exporting
It cannot export the configured report. Apply the configuration, inspect it, and use Search Console's normal export controls manually.
Annotations
It cannot add a custom annotation to the Performance chart.
External event reasoning
It cannot translate an external event into dates, such as “one month after the World Cup” or “before the election.” Supply explicit calendar dates after verifying the event window yourself.
Discover and Google News
It cannot configure the Discover or Google News Performance reports. Its documented scope is Search results Performance only.
Report limits still apply
AI configuration does not change any underlying Search Console limitation.
The resulting report can still be affected by:
- Omission of anonymized queries for privacy.
- UI table row limits.
- Data truncation when grouping by query or page.
- Canonical URL attribution.
- Property-level versus page-level aggregation.
- Rare values in comparisons.
- Preliminary recent data.
- Metric-definition differences across search appearances.
- The rule that only one comparison can be active at a time.
Changing or adding a query or page filter can change totals because anonymized-query and aggregation behavior changes. A generated filter is not evidence that the table contains every matching event.
Review generated regex carefully
The configurator can generate a regular expression for requests involving several terms or patterns.
Verify:
- Whether the expression is a partial or anchored match.
- Whether special characters are escaped.
- Whether alternation is grouped correctly.
- Whether uppercase and lowercase should differ.
- Whether paths, parameters, and hostnames are included unintentionally.
- Whether the pattern includes singular, plural, or spelling variants.
Test a generated expression against known matching and non-matching examples. A syntactically valid regex can still describe the wrong business segment.
Validate every proposal
Google warns that AI can misinterpret a request. Use a repeatable validation pass.
1. Translate the request yourself
Before accepting the proposal, write down the controls you expect.
2. Compare every generated control
Check expected versus proposed filters, dates, comparison, metrics, and table.
3. Look for inherited state
Identify filters that existed before the prompt and remain active.
4. Confirm ambiguous language
Resolve terms such as “traffic,” “recent,” “best,” “this quarter,” “UK,” “phone,” or “brand.” They can map to different metrics or dimensions.
5. Apply and inspect the final chips
After confirmation, verify the report header and filter chips rather than trusting the conversational summary.
6. Sanity-check the result
Compare with a broader report, a manually configured equivalent, or a known page/query example.
If the proposal is wrong, cancel it or correct the report manually. Do not base an analysis on a configuration that merely looks plausible.
Privacy and prompt handling
Google states that text entered into the tool is not used to train its AI models. Google may temporarily log the input for feature improvement after scrubbing personal information.
That statement does not make the prompt field an appropriate place for secrets or sensitive data. Avoid entering:
- Credentials or tokens.
- Personal information.
- Confidential customer or employee data.
- Unannounced business events.
- Private incident details.
- Information unnecessary to configure a Search Console report.
Most useful requests need only report dimensions, URL patterns, query patterns, countries, devices, and dates.
When manual configuration is better
Use normal controls when:
- Only one simple filter is needed.
- The exact match type matters.
- The team already has a documented report recipe.
- A prompt would consume more time to verify than the manual setup.
- The daily limit is being conserved.
- The feature is unavailable.
Use AI-powered configuration when a valid request combines several supported controls and the generated proposal is faster to verify than building the state manually.
The value is reduced setup effort, not superior data or analysis.
A safe operating workflow
- Open the correct property and Search results report.
- Clear unrelated filters or deliberately retain and record them.
- State the metric, dimension, filters, exact dates, and comparison.
- Generate one proposal.
- Compare it with the expected configuration.
- Confirm only after every control matches.
- Inspect the applied report state.
- Perform sorting, export, thresholding, and analysis separately.
- Record the final controls with any finding.
For recurring reporting, save the explicit configuration specification. A natural-language prompt is not a reproducible data definition unless the resulting controls are also preserved.
Common mistakes
- Treating the feature as an AI analyst instead of a configurator.
- Assuming access is universal rather than a limited experiment.
- Trying to use it in Discover or Google News reports.
- Ignoring filters already active before the prompt.
- Accepting an incorrect date interpretation.
- Trusting a generated regex without examples.
- Asking it to sort, paginate, export, annotate, or filter by metric thresholds.
- Using an external event name instead of verified dates.
- Assuming generated views bypass privacy or row limits.
- Calling Web Search clicks sessions or conversions.
- Repeatedly regenerating a nearly correct view and exhausting the daily limit.
- Putting confidential or personal information into a prompt.
What the feature can and cannot conclude
It can:
- Translate supported natural language into proposed report controls.
- Combine several dimensions and filters.
- Configure one supported comparison.
- Select clicks, impressions, CTR, and position.
- Reduce manual setup for a complex Search results view.
It cannot:
- Explain why performance changed.
- Decide which opportunity matters most.
- Manipulate, sort, paginate, or export the table.
- Add annotations.
- Configure Discover or Google News.
- Retrieve hidden or unlimited data.
- Guarantee that its interpretation matches the request.
- Change rankings, indexing, or website content.
Review checklist
- The feature is labeled experimental and limited-rollout.
- The report is Performance for Search results.
- The 20-request per-user daily limit is understood.
- Existing filters have been reviewed.
- Dates and comparisons are explicit.
- Proposed filters match the request.
- Generated regex has positive and negative test examples.
- Metrics and table dimension are correct.
- Unsupported sorting, thresholds, pagination, export, and annotations are handled separately.
- Search Console privacy, aggregation, and row limits remain documented.
- The final report state is recorded with the finding.
- Prompts contain no unnecessary sensitive information.
Related content
- The Performance report: complete guide
- Filters, comparisons & regex
- Search Console data limits
- Find performance winners & losers
- Search Console regex cookbook
- GSC UI vs API vs BigQuery