Search Console now separates image-based search. What the multimodal filter actually tells you

On 24 September 2026 Google quietly split one row in your Search Console Performance report into two. The Web search type is now Web: text-based and Web: multimodal, and the second one contains every visit that began with a camera rather than a keyboard. It is the first time most site owners have been able to see that number at all, and the first thing almost everyone does with it is read it wrong.

What Google actually changed

Google added a new option to the Search type filter in the Search Console Performance report and began rolling it out globally on 24 September 2026. Where the filter previously offered Web, Image, Video and News, Web is now divided in two: text-based for the searches people typed or spoke, and multimodal for the searches that started with a picture.

Multimodal covers four entry points Google has named: Google Lens, Circle to Search on Android, an image uploaded directly into Google Search, and the "Search this image" action in Chrome. In each case the person pointed at something rather than describing it, and Google returned ordinary web results, which is why this data was already sitting inside your Web totals.

Settle that point before you look at the numbers. No traffic arrived on 24 September, and your September clicks did not go up because of this change: text-based plus multimodal equals the Web figure you have been reporting all year. What you gained is the ability to ask how much of it came from images, which nobody could answer with their own data until about ten days ago.

Multimodal is not the Image search type

This is the distinction that most of the coverage has glossed over, and getting it backwards leads to the wrong work.

The Image search type has existed for years and reports your images appearing as results inside Google Images. The subject is the picture: it ranked, someone saw it, maybe they clicked through to the page hosting it.

Web: multimodal reports your pages appearing as results when the search itself was an image. The subject is the page. Somebody photographed a chair, a label, a part number or a building, and Google decided your page answered it.

A page can perform strongly in multimodal while none of its images rank in Google Images at all, because what earned the match may have been the product schema and the body copy rather than the photograph. The reverse happens too. Treating the two reports as one number is how teams end up compressing images to chase a metric that was never about image rankings.

Two rows in that table have no Queries column, and both for the same underlying reason: Google is reporting on surfaces where the thing the user supplied was not a search string it is willing or able to hand back to you. The generative AI reports withhold queries by policy. Multimodal withholds them because there was nothing to withhold.

Why the missing queries are less of a loss than they look

Every SEO's first reaction to an impressions-only report is to write it off, and that reflex is usually right. It is wrong here, because in multimodal the page is the query. When someone photographs a product and Google serves your page, the match itself is the signal: you do not need the query string to know what the search was about, because the page that satisfied it tells you. A spike on four product pages and nowhere else points at an asset you control, which is more actionable than a list of long-tail variants.

What you genuinely lose is intent granularity. You cannot tell whether a Lens search was "what is this", "where do I buy this" or "is this the right replacement part", and those three warrant different pages. Until Google reports something resembling intent here, you are inferring it from the page that matched. That works for product and parts catalogues and is close to useless for editorial content.

How to read impressions-only data: three ratios

Raw multimodal impressions are nearly meaningless in isolation. A number with no benchmark and no history invites both overreaction and dismissal. Three comparisons turn it into something you can make a decision from. Run all three on the same date range, which cannot start before 24 September 2026.

1. Multimodal share of impressions. Multimodal impressions divided by text-based plus multimodal impressions, sitewide and then for your revenue pages alone. This is the only question that decides whether visual search deserves a budget line. The sitewide figure will usually be small; the figure for product, parts or property pages is often several times higher, and that gap is the finding.
2. Multimodal CTR against text-based CTR, same pages. If multimodal impressions convert to clicks at a materially lower rate than text-based impressions on the same URLs, your image won the match but your result did not win the click. That is a title, snippet and page-intent problem, not an image problem, and fixing images will do nothing for it.
3. Page concentration. What share of multimodal impressions comes from your ten strongest pages. High concentration means those pages share a template, an image treatment or a schema pattern, so one fix propagates. Low concentration means the exposure is diffuse and a dedicated project will not pay back yet.

What a visual match is actually built on

If the three ratios say the exposure is real, the work divides into three layers, and they are not equally hard or equally effective. A visual match draws on the image itself, the structured data describing what the page sells, and the text context around the image. Pixel similarity alone does not decide it.

The ordering is the useful part. Most teams reach for the last row because it is the one their tooling measures, then conclude visual search did not work for them. Compression is worth doing and will not make a chair recognisable. Meanwhile product schema is the item on the list that costs a day of template work and does the most per hour spent.

Who should care, and who honestly should not

Visual search exposure is not evenly distributed, and pretending otherwise is how agencies sell projects nobody needed. Lens handles a reported 25 billion visual searches a month, with roughly one in five said to carry commercial intent, which sounds like it should matter to everyone. It does not, because the searches cluster around things people can point a phone at.

It matters most for physical products with visual identity: furniture, fashion, lighting, tiles and finishes, plants, tools, components and spare parts. It matters a great deal for anything identified by a label, a nameplate or a part number, where photographing the thing is genuinely faster than describing it. It matters moderately for food, property and vehicles. It matters very little for professional services, B2B software and most editorial publishing, where the useful move is to look at the report once a quarter and spend the time elsewhere.

If you fall in the first group and your product pages were photographed for a lookbook rather than for recognition, that is now a measurable gap rather than a theory.

What to do in the first month

A short sequence, in order, with no step that depends on a tool you do not already have.

1. Open the Performance report, set the date range to start 24 September 2026, and add Web: multimodal as a filter. Export pages for text-based and multimodal separately.
2. Calculate the three ratios above. Write the numbers down somewhere you will find them in January, because the trend will be worth more than the level.
3. Run the same exercise inside the Generative AI report if your property has it, remembering that it carries impressions only and begins on 18 May 2026 with no backfill.
4. Take your top ten multimodal pages and check them against the layer table. Most sites find the same two gaps: no product schema, and alt text written as a filename.
5. Fix the template-level items first, then re-measure after six to eight weeks. One date range tells you nothing about a surface this new.

None of this is a project. It is an hour of reporting and a day of template work, and it belongs in the same quarterly routine as the rest of your Search Console housekeeping.

What it signals about where Search Console is going

Three reporting changes landed in roughly four months: generative AI performance reports from 3 June 2026, those reports and the associated AI usage control reaching every property by 31 August, and the multimodal split on 24 September. All three share a shape. Google is giving site owners visibility into newer surfaces while withholding the query data that made the old Performance report useful for building keyword plans.

So more of your diagnosis in website and SEO work will come from page-level patterns and less from query lists, because page-level patterns are all the newer surfaces return. Teams already working in pages and templates will barely notice. Teams whose entire reporting model is a keyword spreadsheet will.

Frequently asked questions

What is the multimodal search type in Google Search Console?

Multimodal is a search type filter Google added to the Search Console Performance report on 24 September 2026. It isolates web results that a person reached by searching with an image rather than words: Google Lens, Circle to Search on Android, an image uploaded to Google Search, and Chrome's "Search this image" action. Google split the existing Web search type into "text-based" and "multimodal", so the two together equal what Web reported before.

Why does the multimodal filter have no query data?

Because the person did not type a query. The search was an image, so there is no text string for Search Console to attribute the impression to. When you select the multimodal search type the Queries tab is not offered at all. You still get pages, countries, devices and dates, which is enough to see which pages are being found visually but not what the searcher was holding up to their camera.

Is multimodal the same as the Image search type in Search Console?

No, and the difference matters. The Image search type reports your images appearing as results in Google Images. Multimodal reports your web pages appearing as results when the search itself started from an image. One is about your pictures being found; the other is about your pages being found by a picture. A page can rank well in multimodal while none of its images rank in Google Images.

Does Google Lens traffic count as organic search traffic?

Yes. Lens and Circle to Search sessions that land on your site were already counted in your Search Console Web totals and in your analytics organic channel before this filter existed. The 24 September change is a breakout, not new data, so your aggregate clicks and impressions do not change. What changed is that you can now see how much of the total came from image-led searches.

How do I improve visibility in Google Lens and visual search?

Work the three layers a visual match depends on, in this order: the image itself (the product as the clear subject, uncluttered background, at least 800px on the longest side, several angles), the structured data on the page (Product schema with name, brand, price, availability and SKU, matched to a clean Merchant Center feed if you sell), and the page context around the image (descriptive filename, specific alt text, and body copy that names the thing in the picture). Image quality gets you the match; structured data and page context get you the result.

How much multimodal traffic should a site expect?

Google has published no benchmark, so the only honest answer is to measure your own share and watch it move. Compare multimodal impressions against text-based impressions for the same date range and the same pages. As a working threshold, a share under about 2% that is spread thinly across many pages does not justify a dedicated project; a share above about 5% concentrated on pages that make money does. Those are judgement lines, not published figures.

Does the multimodal filter work in the Generative AI report?

Yes, with less data. In the Search results report, multimodal gives you clicks and impressions. In the Generative AI performance report, which covers AI Overviews, AI Mode and AI features in Discover, it gives impressions only. Google rolled the generative AI reports and the related AI usage control out globally by 31 August 2026, with data starting 18 May 2026 and no historical backfill.

The takeaway

The multimodal filter is a small change that answers a question people have been guessing at for two years: how much of my search visibility is images doing the work. For most sites the answer will be "not much, yet", and that is a perfectly good outcome for an hour of reporting. For catalogues of physical things, the answer will be large enough to change what you brief your photographer and your developer to do. Either way, the figure you record this month is the baseline you will be comparing against for the rest of the decade, so pull it before the date range gets long enough to hide the trend.