Camera, not keyword: making visual search measurable

A search box with no text box

Someone photographs a trainer on a train, circles a lamp in a screenshot or points a phone camera at a spare part. The query is not a neatly phrased sentence. It is an image, a crop, a context and sometimes a few additional words.

Since 24 September 2026, Google has made that journey visible in Search Console. According to Google, the new multimodal search filter covers Lens, Circle to Search on Android, image uploads and Chrome's “Search this image” action. The data appears in the standard Search performance report and in the Generative AI performance report, with a global rollout. Naturally, a site will see the filter's data only when it receives impressions from those searches.

For marketing and product teams, this is more useful than another trend label. For the first time, first-party data can show whether images merely decorate pages or actually bring people from a visual discovery moment to the website.

Not a new ranking spell

The temptation is to sell “visual search optimisation” as a separate discipline immediately. The less dramatic view is more useful: Google points to the same fundamentals that apply to the rest of Search. Content needs to be crawlable, indexable, relevant and embedded in a useful page. Google's guidance says generative search requires neither a special AI file nor secret schema markup.

What is new is measurement of an intent that used to be hard to see. A classic keyword roughly tells us what somebody typed. A camera search shows what they currently have in front of them and want to understand, compare or buy. That does not change the job of the website, but it opens another entrance.

The right first question is not “How do we rank in Lens?” It is “For which visual problems can our real images provide a better answer than interchangeable stock photography?”

Build a defensible baseline first

Open the Search performance report, choose multimodal as the search type and inspect at least four dimensions: pages, countries, devices and dates. The Generative AI report also breaks visibility down by impressions, pages, countries, devices and time. Export the data when you need to connect it with analytics and business outcomes.

We would not invent a complex metric at the start. A simple table is enough:

• Which landing pages receive multimodal impressions?
• Which images on those pages carry the actual information?
• Do impressions, clicks and qualified next actions rise together?
• Which countries and devices behave differently?
• Which pages are seen but not visited?

Keep visibility and impact separate. An impression means a URL appeared in a visual search context. It does not prove revenue or a good experience. Add website signals such as a product view, enquiry, basket or download, but do not misrepresent the Search Console impression as a conversion.

Sort the image inventory by job

Many websites own hundreds of images and very few visual answers. A hero full of smiling people does not explain a product. A generic laptop mock-up does not help anyone recognise a component, material or outcome.

For a useful inventory, we classify images by their job:

Recognise: Show a product, model, shape, colour, material or condition unambiguously.

Compare: Make variants, sizes, before-and-after states or differences visible.

Understand: Explain use, installation, process or spatial context.

Trust: Prove real work, people, place, detail and results.

Act: Connect the pictured object to price, availability, booking or the next step.

A page does not need every image type. It needs the image that matches intent. For an online shop, that could be a clean side view and a detailed close-up. A trades business benefits from real faults, materials and finished installations. For software, meaningful states and workflows teach more than a decorative device in perspective.

Images must be genuinely discoverable on the web

Google continues to recommend ordinary HTML image elements. An image in the src of an img element can be found and processed; the documentation says CSS background images are not indexed. Picture and srcset remain appropriate for responsive delivery as long as there is a real img fallback with src.

This sounds basic, yet it is a common failure in modern frontends. A visually central product photograph lives as a component background, loads only after an interaction or exists solely inside a canvas. The page looks complete to a person while its most important object is absent from image search.

Check every priority page:

1. Does the core image exist as an actual image in the rendered HTML?
2. Does its URL reliably return a supported file type?
3. Can Google fetch both page and image without a login or blocking rule?
4. Do responsive variants always retain a working default source?
5. Should hard-to-discover assets be added to an image sitemap?

Technical discovery is not glamorous. Without it, the best creative never enters the room.

Alt text describes; advertising copy sells

Google uses alt text together with computer vision and page content to understand an image. Alt text is also an accessibility feature. That is exactly why it should describe what the image contributes in this context, rather than list hoped-for search terms.

“Hamburg web agency best cheap app development” is a poor description for both a screen reader and a search system. “Three seagulls review mobile app designs at a workbench in Hamburg harbour” describes the subject and its context.

Not every image needs a long explanation. Purely decorative images should have empty alt text. Informative images need a concise, concrete description. A chart needs additional accessible page copy when its message cannot fit into one sentence.

The filename and image title provide only supporting clues. The complete package matters: image, alt text, visible heading, caption and landing-page content should all mean the same thing.

Quality and speed belong together

Visual search does not reward an enormous original file. Google explicitly notes that sharp, high-quality imagery is attractive while images are often the largest contributor to page weight. The answer is not “small or beautiful”; it is a responsive delivery pipeline.

In practice, that means:

• several useful widths instead of one 4,000-pixel file for every device;
• modern formats such as WebP or AVIF with a reliable fallback where needed;
• explicit width and height to prevent layout shifts;
• load the most important image early and defer images below the fold;
• never lock essential information exclusively inside artwork;
• preserve genuine detail where visitors zoom or compare features.

The landing page remains part of the answer. A visitor arriving from visual search wants to find the recognised object immediately, understand it and see the next step. A strong picture on a slow, confusing page is simply an attractive reason to leave.

Metadata helps; it does not replace substance

For appropriate page types, structured data can connect images with products, articles or other entities and enable rich results. Google also identifies primaryImageOfPage, the main entity's image property and og:image as ways to indicate a preferred preview.

That image should be relevant, representative, high-resolution and not extremely cropped. A logo reused as every page's default says very little. Nor should schema claim anything that visible page content cannot support.

For merchants and local businesses, product feeds, Merchant Center and a maintained Business Profile belong in the same workflow. They keep facts such as price, availability, place and offer consistent. Structured data is a clean translation of information that exists — not a substitute for missing information.

Run a four-week test, not an image marathon

Do not start with the entire media library. Choose five to ten pages that already have demand or commercial value. Keep the test manageable:

Week 1: Export the baseline; document primary images and technical problems.

Week 2: Improve one to three visuals per page: real perspective, clear detail, appropriate crop, descriptive alt text and visible context.

Week 3: Check responsive delivery, load performance, metadata and indexability; publish the changes cleanly.

Week 4: Compare Search Console with web analytics. Do not declare winners too early; record the first patterns and continue observing.

Search data takes time and fluctuates. Record publication date, changed images and page type. Without a change log, every later chart becomes a guessing game.

What we would deliberately avoid

We would not manufacture a thousand near-identical pages for possible camera subjects. We would not stuff alt text with keywords, replace truthful product photography with imaginary generated versions or sell a third-party number as an “AI rank” when it does not come from the site's measurement system.

The new Search Console view is not permission for false precision either. Small samples, seasonality, assortment changes and concurrent campaigns can all distort the result. Good analysis states those limits.

The goal is not to trick a machine. It is to show and explain real things so clearly that a person and a search system both recognise the page as the helpful continuation.

Visual discoverability is team work

Multimodal search connects disciplines that companies often separate. Design owns subject and hierarchy. Content creates context. Development provides discoverable HTML, performance and metadata. SEO checks accessibility to search systems. Analytics connects visibility with an actual outcome.

The new report gives those teams a shared starting point. The conversation can move from “we need nicer pictures” to specific pages, impressions and journeys.

Camera instead of keyword does not end classic search. It means strong websites must make their products, places, processes and results understandable visually as well as verbally. Teams that measure carefully and begin with a few strong pages will learn faster than teams that immediately open another optimisation circus.

Sources

• Google Search Central: Web multimodal Search performance reporting
• Google Search Central: Generative AI performance reports in Search Console
• Google Search Central: Optimizing for generative AI features
• Google Search Central: Image SEO best practices