Every reporting change arrives twice. First as a feature announcement, which is how the vendor experiences it, and then weeks later as an unexplained step in somebody's chart, which is how the client experiences it. Google's new multimodal search type in Search Console is currently between those two arrivals, and the gap has a genuine factual dispute sitting in it.
The feature itself is small and welcome. Google began rolling out a Search Console multimodal report globally on September 24, 2026, as a new search type in the Search results performance report and in the generative AI report alongside it. Select it and you see how your pages perform when somebody searches with an image rather than with words.
The dispute is about where those numbers came from. Google says they are new. A working practitioner's charts say they look reallocated. Nobody has reconciled the two, and until somebody does, any September impression delta on your account has at least two candidate explanations that point in opposite directions.
Four numbers, and the interesting one is the zero. A search type with no query dimension is a channel you can size but cannot interrogate, which is a different kind of reporting object from everything else in the performance report.
What the Search Console multimodal report actually covers
The Search Console multimodal report covers web results reached through Google Lens, Circle to Search on Android, images uploaded directly to Google Search, and the Chrome right-click option to search an image. Four entry points, one row in the search type filter.
Those four have almost nothing in common from a user's point of view. Circle to Search is a gesture inside another app. Lens is a camera pointed at a physical object. A Chrome right-click is a desktop researcher checking an image they already found. Google has collapsed them because they share a technical property, which is that the query is an image, and that property is exactly what removes the query dimension.
| WHAT THE REPORT GIVES YOU | WHAT IT WITHHOLDS | WHY THE GAP EXISTS |
|---|---|---|
| Clicks and impressions for multimodal searches, in the Search results report | The text of the query, because there mostly is not one | Google's documentation states that multimodal searches use images rather than text, so text query data is not available for this traffic |
| Pages, countries, devices and dates, filterable as usual | Which of the four surfaces produced a given row | The search type is a single value. Lens, Circle to Search, uploads and Chrome right-clicks are not separable from one another |
| Impressions in the generative AI report, covering AI Overviews and AI Mode | Whether a visual search resolved into an answer or a link | The generative AI report has always measured impressions inside AI surfaces rather than the mechanism that got the user there |
| A size for a channel most teams have never measured at all | Any way to optimize toward a keyword | Visual search has no keyword. The optimizable unit is the image and the page it sits on, which is a different workflow from the one your reporting is built around |
That last row is the practical content of the release. For years the honest answer to how much traffic Lens sends you was that nobody outside Google knew. Now there is a number. The number arrives without the dimension that would let you act on it, which is frustrating but not dishonest: Google cannot report a text query that was never typed.
Google says the data is new, and one practitioner's charts disagree
Google's position, stated by John Mueller after he checked with the team, is that multimodal traffic was never counted in Search Console before and the new report therefore adds volume rather than moving it. His exact words were that the data was not previously in the counts and the report has new data.
Worth noting how he got there. Mueller's first reply hedged, saying he would double-check and that he thought these were not reported at all before. The follow-up, after asking the team, was definite. That is a good-faith answer from a named person at the company, and it is the most authoritative statement on the record.
It also does not fit what Dave Smart reported. Smart asked publicly whether multimodal impressions had previously sat inside the web search type, and said he saw corresponding declines in impressions and clicks across several properties starting on September 10, declines that lined up with the arrival of the new data. Addition cannot cause a decline anywhere. Reallocation can.
“A vendor telling you a number is new, and your chart telling you a number left, are not opinions in tension. One of them is checkable on your own account in about ten minutes.”
The honest summary is that Google has stated the data is additive, at least one careful observer has evidence it is not, and the disagreement is unresolved as of today. Treat the additive claim as the leading explanation, because it is on the record from the source, and verify it against your own series before you report a September number to anybody who makes decisions with it.
Additive or reallocated changes what your September chart means
Whether multimodal impressions were added to Search Console or carved out of the existing web totals decides whether a mid-September dip in a performance chart is a reporting artifact or a real loss of visibility. Same chart, two opposite readings, and only one of them warrants action.
If the data is additive, your web series is untouched, your totals went up, and any dip you see in September belongs to something else entirely. That is the comfortable reading and it is what Google has said. If the data is reallocated, your web series lost volume it never actually lost, your year-over-year comparison is broken at a specific date, and a team that does not know about the break will spend a quarter chasing a decline that is an accounting entry.
The second failure mode is the expensive one, and it is not hypothetical. We have watched teams rebuild content calendars around Search Console movements that turned out to be reporting changes, which is the same category of error as the generative AI logging bug that undercounted impressions in August. A metric that moves because its definition moved will not respond to any amount of good work, and a team that keeps working the metric anyway burns real budget on it.
Google's stated monthly Google Lens query volume over time, indexed so that 25 billion equals 100
The growth curve is why this matters beyond a single month's reporting hygiene. Visual search roughly doubled in three years by Google's own accounting, and Google has said roughly a fifth of Lens queries carry commercial intent. A channel at that scale with no measurement was a real blind spot, and it is the reason the release is worth the attention even while the provenance question is open.
For most sites the multimodal share will be small enough to make the whole dispute academic. For product-led catalogs it will not. If somebody can point a camera at your product, or at a competitor's product, and land on a page of yours, then this report is describing acquisition you have never optimized, which is a live question for anybody running ecommerce and retail search programs.
The num=100 precedent, when a reporting change looked like a collapse
Search Console's impression baseline already moved once for reasons that had nothing to do with performance, in September 2025, when Google disabled the num=100 search parameter that let rank-tracking tools pull a hundred results in a single request.
The effect was industry-wide and severe on paper. Search Engine Land's analysis of the aftermath found that 87.7% of sites lost impressions in Search Console and 77.6% saw their count of unique ranking keywords fall. Nothing had happened to any of those sites. Scraper traffic had been generating impressions for pages sitting on result pages two through ten, and when the scrapers lost their cheap request, those impressions stopped existing.
| REPORTING EVENT | WHAT MOVED | WHAT DID NOT MOVE | THE LESSON THAT TRANSFERRED |
|---|---|---|---|
| num=100 removal, September 2025 | Impressions and unique keyword counts fell across most of the web. Average position improved because low-position scraper impressions left the denominator | Clicks held steady and click-through rate improved, because no human behavior had changed | When impressions fall and clicks do not, suspect the definition of impression before you suspect your rankings |
| Generative AI report logging bug, August 2026 | Impressions inside AI surfaces were undercounted for a period, then corrected | The underlying appearances in AI answers, which were happening the whole time | A metric can be wrong in the direction that flatters nobody, and a correction looks exactly like growth |
| Multimodal search type, September 2026 | A new search type appeared with its own clicks and impressions. Whether the web series lost matching volume is disputed | How visual searches select your pages, which this release does not touch at all | Verify against your own totals rather than waiting for the general answer, because the general answer may not arrive |
Three events in thirteen months, all of which shifted numbers that a lot of people report upward to executives as performance. That frequency is the argument for building the habit rather than handling each one as news. Reporting changes are now a standing category of variance in search analytics, and a measurement practice that cannot distinguish them from performance is not a measurement practice. Separating the two is ordinary scope in reporting and analytics work and it is the first thing we look for when a client brings us a chart that dropped.
How to read the Search Console multimodal report without fooling yourself
Reading the Search Console multimodal report honestly means treating late September 2026 as a baseline break rather than as a data point, and reconstructing the comparison you actually want out of the pieces the report does give you.
One further thing worth doing while the export is open: check whether your images are actually eligible to be the thing a camera matches. Blocked image directories, thumbnails too small to index and pages that render their images only after a script runs all remove you from this surface silently, and they are standard findings in a technical SEO and GEO audit. The report can only show you traffic that was possible in the first place.
Questions teams are asking about the Search Console multimodal report
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Tyler leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.