Track visual search referrals in Search Console

To track visual search referrals in Search Console, use Google’s multimodal reporting to identify exposure from image-led searches, then distinguish that visibility from actual visits. The key is knowing which report you are reading: the Generative AI performance report tracks impressions, so its numbers should not be presented as referral clicks.

Google’s September 24, 2026 announcement introduced web multimodal reporting across Search results performance and Generative AI features. This guide explains what the reporting covers, how to investigate your pages, and how to build a measurement workflow without overstating what the data proves.

What visual search referrals in Search Console actually cover

Google describes multimodal search traffic as activity from searches that use images. The reported signal includes several ways someone can begin a search, rather than one standalone visual-search product.

• Google Lens: searches initiated using Google’s visual search experience.
• Circle to Search: image-led searches using the Android feature.
• Uploaded images: searches made by uploading an image to Google Search.
• Chrome image search: searches initiated through Chrome’s right-click image-search option.

This scope matters because a visitor’s discovery journey may start with an object or a scene rather than a typed phrase. For example, someone might search from a photo of a chair, a screenshot of a jacket, or an image containing several household objects. Those are illustrative scenarios, not evidence that any particular page received traffic through them.

Google announced that the multimodal reporting integration was rolling out globally starting September 24, 2026. Its guidance places these metrics alongside performance data for Search results and generative AI features, making visual-search analysis part of the broader Search Console workflow.

Multimodal reporting is not a separate label for each entry point

The supplied Google guidance identifies the experiences included in multimodal reporting. It does not establish that you can isolate Lens, Circle to Search, uploaded images, and Chrome image search individually within the report.

Keep your reporting language at the level the interface supports. If you select a combined multimodal view, describe the findings as multimodal or image-led search performance rather than assigning all activity to Lens.

Use multimodal reporting to understand image-led discovery. Use the metric definitions to decide whether you are describing visibility or visits.

That distinction is especially important when stakeholders ask for “visual search referrals.” The phrase often implies people arriving on the website, while some of the available reporting measures only whether a link was shown.

How to find and track visual search referrals in Search Console

Start with a specific question, such as which pages appear in image-led searches or whether visibility changed after a content update. A clear question makes it easier to select the right report and avoid treating every available number as a traffic metric.

Direct answer: Open your Search Console property and review multimodal data in Search results performance or Generative AI performance. In the Generative AI report, select the “Web: multimodal” search type, choose a date range, and analyze Pages, Countries, Dates, and Devices. Treat the generative AI figures as impressions, not referral clicks.

1. Choose the property you want to investigate. Keep the property consistent throughout the analysis so comparisons refer to the same website scope.
2. Open the relevant performance report. Google says multimodal reporting is available in Search results performance and Generative AI features. Choose the context that matches your measurement question.
3. Select the multimodal view. The Help Center explicitly identifies “Web: multimodal” as the search type in the Generative AI performance report.
4. Set your analysis period. Use a date range that gives you enough activity to examine, rather than drawing conclusions from one isolated observation.
5. Inspect the available dimensions. Review Pages, Countries, Dates, and Devices to see where exposure is concentrated.
6. Record the metric definition. Note whether the view measures impressions or another explicitly labeled metric before sharing results.
7. Export the data if you need deeper analysis. Google provides an Export button for downloading chart and table data.

Keep the two reporting contexts separate

Search results performance and Generative AI performance are related reporting contexts, but that does not make every metric interchangeable. Record the report name with each result so a later comparison does not mix unlike measurements.

The exact “Web: multimodal” label is established for the Generative AI performance report. For Search results performance, follow the multimodal controls actually shown in your property rather than assuming every menu label and metric is identical.

A useful first deliverable is a short baseline: the property, report, selected search type, date range, metric, and leading pages. This gives you a reproducible starting point without pretending that one screenshot explains the entire visual-search journey.

Separate visual-search impressions from referral traffic

In the Generative AI performance report, Google says the default view shows impression data. An impression is counted when a link to your site is shown to a user in a generative AI feature.

That is a visibility event, not proof of a website visit. A page can therefore earn reported exposure without the report establishing that anyone clicked through, read the page, or completed a business action.

Match each business question to the evidence

• “Are our pages appearing?” Generative AI impression data can help answer this within the report’s coverage.
• “Which pages receive that exposure?” Use the Pages dimension to identify the URLs associated with reported impressions.
• “Are people visiting from these searches?” Use a metric explicitly defined as clicks or visits where available; generative AI impressions do not answer this.
• “Does visual discovery produce valuable outcomes?” Investigate website behavior and business outcomes separately, while acknowledging any attribution gaps.

For example, suppose an instructional page gains multimodal impressions during your comparison period. You can report that the page’s recorded visibility increased in that view. You cannot conclude from impressions alone that the page received more referrals or generated more sales.

This is not a reason to dismiss impression data. It can identify pages worth investigating, show where exposure is concentrated, and establish whether a content area participates in image-led discovery at all.

Use precise names in dashboards and updates

Label a chart “Generative AI multimodal impressions” rather than “Visual search traffic” when impressions are the underlying measurement. Clear naming prevents a visibility metric from becoming an assumed traffic result as the chart moves between teams.

If your analysis also uses website analytics, keep its visit and outcome measures distinct from Search Console exposure. Do not assume that a downstream session can be assigned to a specific visual-search entry point unless your measurement setup provides evidence for that assignment.

Likewise, do not manufacture a click-through rate from an impression-only view. A rate needs a compatible numerator and denominator; an unrelated traffic total divided by multimodal impressions would not establish a reliable visual-search click-through rate.

The practical goal is a measurement chain with clear boundaries: observed exposure, separately evidenced arrivals, and separately measured outcomes. Even when the chain is incomplete, naming the missing link is more useful than filling it with an assumption.

Segment multimodal performance by page, country, date, and device

Google supports Pages, Countries, Dates, and Devices as dimensions in the report. Each answers a different question, so use them sequentially rather than collecting every breakdown without a purpose.

Pages: find the content behind the exposure

Start with Pages to identify which URLs earn recorded visual-search exposure. Then inspect those pages directly to understand their subject, images, and the information available to a visitor.

Group pages into categories that make sense for your site, such as product pages, buying guides, tutorials, or reference content. These are your analytical categories, not Google-provided classifications, so document how you assign them.

A page-level result does not tell you which particular image triggered discovery. If a page contains multiple images or covers several related objects, treat the URL as the observed unit and any image-level explanation as a hypothesis.

Countries: understand where visibility appears

The Countries dimension lets you investigate the geographic distribution of reported activity. This can help prioritize a closer review of pages serving particular markets or languages.

If one market stands out, examine whether the destination content is useful for that audience. For a product page, that might mean checking whether availability, specifications, and purchase information are clear; it does not mean the country breakdown itself proves purchase intent.

Dates: distinguish a pattern from a single observation

Use Dates to see whether exposure is sustained, intermittent, or concentrated around a particular period. Compare periods of the same length where practical, and keep the selected report and search type consistent.

Record relevant site changes alongside the timeline, including image replacements, page rewrites, or product updates. A shared timeline can suggest what to investigate, but a change occurring near an increase does not establish that it caused the increase.

Devices: check the destination experience

The Devices dimension adds another view of where reported exposure occurs. Use it to decide which page experiences deserve closer inspection rather than assuming that a device category identifies a particular Google product.

For instance, a device breakdown alone does not prove that activity came from Circle to Search. The report’s combined multimodal scope still applies unless a more specific breakdown is explicitly available.

• Begin broadly: identify the pages with meaningful activity in the selected view.
• Add context: examine country and device patterns for those pages.
• Check persistence: review the date pattern before choosing an action.
• Write a bounded conclusion: describe what the dimensions show without inventing the user’s exact search journey.

Export Search Console data into a repeatable analysis workflow

Google’s Export button lets you download chart and table data for external analysis. This is useful when you want to group pages, maintain a history of observations, or combine reporting notes with a content-change log.

The export is a way to analyze the available data, not a way to recover information the report does not supply. Keep the same metric definitions and scope when moving the numbers into a spreadsheet or another reporting tool.

Save the context with every export

A file containing page URLs and numbers is not sufficient documentation by itself. Someone reviewing it later needs to know what was selected in Search Console and what those numbers represent.

• Property: the website property used for the analysis.
• Report: Search results performance or Generative AI performance.
• Search type: the multimodal selection used, including “Web: multimodal” where applicable.
• Date range: the period covered by the export.
• Metric: the exact measurement being analyzed.
• Dimensions and filters: the breakdowns and restrictions applied.
• Export timing: when you downloaded the data, especially if the period includes the newest available figures.

Build a baseline before creating targets

Use an initial export to establish what your property actually shows. Identify the pages present, the distribution of exposure, and whether the data is substantial enough to support detailed segmentation.

Then choose a review cadence suited to the volume you observe. A property with sparse activity may need a broader period for useful comparisons, while more frequent checks can serve as monitoring rather than evidence of a lasting trend.

For recurring reviews, preserve an unedited export and perform grouping or calculations in a separate working copy. This practical safeguard makes it easier to trace a conclusion back to the original data.

Compare like with like

Keep the property, report, search type, metric, and period length aligned when comparing results. If you change one of those choices, note the change explicitly rather than presenting the result as a continuous series.

Do not add totals from Search results and Generative AI views merely because both include multimodal reporting. The supplied guidance does not establish that those totals are mutually exclusive or suitable for a combined unique-traffic figure.

A useful export analysis ends with a decision, not just a larger spreadsheet. It might identify a group of pages for editorial review, a market requiring better destination content, or a finding that needs more observation before action.

Why the multimodal report may be missing or incomplete

A missing report does not automatically mean your site is excluded from visual search. Google’s Help Center says the Generative AI performance report usually does not appear until a property has received enough impressions in generative AI features.

Google also states that Generative AI performance insights were available to all websites worldwide by August 31, 2026. That availability statement is distinct from the September 24, 2026 global rollout announcement for multimodal reporting.

Separate availability from observed activity

A feature being available globally does not guarantee that every property will display meaningful data. Eligibility for reporting, sufficient impressions, and activity within your selected period are separate considerations.

1. Check the property and reporting context. Confirm that you are reviewing the intended site and the relevant performance report.
2. Review the selected period. A narrow window may not be useful for a property with limited eligible exposure.
3. Check your filters. Restrictive page, country, or device selections can limit what you see.
4. Use Google’s stated threshold explanation. For the Generative AI report, insufficient eligible impressions are a documented reason it may not appear.

Allow recent figures to settle

Google notes that the newest data can be preliminary and may change over the next few hours. If you are investigating a recent spike or drop, record it as an initial observation and recheck before treating it as final.

This is particularly important when a report is being used for a same-day status update. Avoid turning an unsettled figure into a firm performance verdict or an urgent content-change request.

Recognize the experimental coverage limit

Google explicitly excludes Search Labs experiments from this report. Some experimental visual-search experiences may therefore fall outside the reported data.

Keep that boundary in mind when reconciling a test someone performed with the figures in Search Console. A personally observed experimental result and the report are not necessarily measuring the same covered experience.

Once these checks are complete, document what remains unknown. “No sufficient reported exposure in this view” is a more defensible conclusion than “Google never uses our content for visual search.”

Use visual-search findings to choose better content investigations

Reporting becomes useful when it changes what you examine next. The data can identify pages for review, but it does not provide a guaranteed formula for increasing visual-search visibility.

Google’s product announcements explain why image-led discovery deserves attention. In March 2026, Google highlighted a Lens and Circle to Search update capable of breaking down and searching multiple objects within one image simultaneously.

Google also says Circle to Search is available on more than 580 million Android devices worldwide. That figure describes product availability, not the number of users, searches, or referrals your website should expect.

Consider the broader visual context

In its September 2025 AI Mode announcement, Google described a visual search fan-out approach that analyzes subtle details and secondary objects, then runs multiple background queries to better understand visual intent. Google said the experience builds on Lens and Image Search, combined with Gemini 2.5 and that fan-out technique.

These capabilities make it sensible to review the whole subject of an image and its destination page, rather than assuming that only the most prominent object matters. They do not reveal which detail caused your URL to appear in a particular reported impression.

For an illustrative example, a room photograph might contain seating, lighting, and decorative objects. If a related page earns multimodal impressions, inspect whether its content helps readers understand the relevant items instead of claiming that one specific object generated the exposure.

Turn page findings into testable editorial questions

• Is the subject clear? Review whether the page explains the objects, products, or processes represented in its images.
• Does the destination answer a plausible need? Consider identification, comparison, instructions, specifications, or purchasing information where appropriate.
• Are important distinctions explained? Check whether readers can understand differences in model, material, size, version, or use case when those details matter.
• Does the page work for the observed audience? Use country and device findings to prioritize practical usability checks.

These are content-quality investigations, not stated Google ranking factors or promises of improved placement. Their purpose is to make a discovered page more useful and to give your team a concrete, reviewable hypothesis.

Log changes and evaluate cautiously

When you make a change, record the page, what changed, why you changed it, and when it went live. Revisit the same reporting view after a suitable observation period and compare the available evidence.

If impressions increase, report the increase and the timing of the change separately. If nothing changes, use that result to reconsider the hypothesis rather than assuming the report is broken or repeating the same edits across the site.

Visual search has a long product history: Google’s July 2026 anniversary post traces Google Images to July 2001, Search by Image to 2011, Lens to 2018, and Circle to Search to 2024. That history provides context for the reporting, but your own page-level evidence should determine the next action.

Build a visual-search report stakeholders can act on

A useful stakeholder report explains what happened, what the data can support, and what you will do next. It should not require readers to understand every Search Console setting before they can interpret the conclusion.

Lead with the most decision-relevant observation, such as a particular content category receiving recorded exposure or a previously observed pattern persisting across another period. Keep the metric name visible in the line or first sentence.

Use a compact reporting structure

1. Scope: identify the property, performance report, multimodal selection, and date range.
2. Observation: describe the measured pattern using the correct metric.
3. Breakdown: name the pages, countries, or devices that explain the finding.
4. Interpretation: separate the supported conclusion from any hypothesis about user intent.
5. Action: identify the review, content change, or further analysis the finding warrants.
6. Validation: state what evidence you will check next to assess the action.

For example, an update could say that instructional pages are receiving multimodal impressions in the Generative AI view, and that the team will review the leading destination pages for completeness. This communicates a real observation and a reasonable next step without claiming referral growth.

Choose actions according to the strength of the evidence

When the pattern is sparse or inconsistent, continued observation may be the best decision. When a group of pages repeatedly appears, a focused editorial or usability review may be justified.

If stakeholders need revenue attribution, be explicit that the generative AI impression report cannot supply it by itself. The next task is to assess what separate arrival and outcome evidence your measurement setup can provide, not to rename visibility as revenue contribution.

Assign an owner to each follow-up and keep a short record of the resulting decision. This prevents recurring reports from becoming a collection of screenshots with no connection to site improvements or business questions.

The strongest reporting habit is to preserve the boundary between observation and explanation. Search Console can show you where to investigate; your analysis should make clear which findings are measured, which are inferred, and which remain unresolved.

Track visual search referrals in Search Console by starting with the multimodal reporting context, preserving its metric definitions, and analyzing pages before drawing broader conclusions. In Generative AI performance, “Web: multimodal” identifies image-led search exposure, but the reported impressions are not proof of visits.

Open the relevant report, establish a documented baseline, and choose one page group to investigate first. Export the evidence, record your next action, and return to the same view for a consistent follow-up rather than chasing isolated changes.