Optimisation, , 13 min read

Google Lens processes 20 billion visual searches monthly in 2026

Images with spatial alt text improve answer accuracy. Multimodal models semantically analyze images beyond keywords.

Citead

In brief

Multimodal search engines now analyze images with sophistication comparable to text processing. Google Lens processes billions of visual searches each month, a significant share of them related to shopping. Multimodal language models use Visual Transformers to understand semantic relationships between objects in an image, going beyond simple keyword reading in alt attributes. This evolution requires content creators to optimize their images according to precise technical criteria. Alt descriptions containing spatial information help multimodal systems match images more accurately. Compressed images with visible artifacts reduce model confidence, and images below 800x800 pixels are more prone to classification errors. Optimizing images for multimodal AI citations requires particular attention to resolution, compression quality, descriptive metadata, and semantic file naming. Brands that anticipate this multimodal transition significantly increase their visibility in AI-generated responses.

Key points

Visual search has experienced exponential growth since the introduction of multimodal capabilities in major language models. Google Lens, Google's visual search tool, now processes nearly 20 billion visual queries each month according to official company data. This massive adoption reflects a fundamental change in user search behavior, with many preferring to photograph an object rather than attempt to describe it textually. The implications for online visibility strategies are considerable. Google's AI Overviews now appear on a large share of queries, and ChatGPT reaches hundreds of millions of weekly active users. In this context, optimizing images for AI system comprehension becomes a major visibility lever. Multimodal models no longer simply read keywords in alt attributes. They perform semantic analysis of the image itself to understand its content and context. This capability requires rethinking the entire image optimization strategy.

Visual Transformers semantically analyze images

The year 2026 marks the era of multimodal models where large language models are no longer limited to text but also analyze images according to Passionfruit. Generative engines use Visual Transformers to understand relationships between objects within a visual frame. This technology allows search systems to go beyond simple keyword matching in alt attributes to perform semantic analysis of visual content. The technical process relies on visual tokenization that breaks an image into a grid of patches called visual tokens. The system converts raw pixels into sequences of vectors, allowing AI to process content like "an image of [visual token] on a table" as a single coherent sentence. This unified approach to text-image processing enables models to understand contextual relationships between visual and textual elements on a page.

For large language models, images, audio, and video constitute sources of structured data just like text. Visual Transformers analyze not only individual objects in an image but also their spatial relationships, relative proportions, and surrounding context. An image showing a laptop on a desk with a coffee cup will be understood as representing a workspace, regardless of keywords in the alt attribute. This contextual understanding allows models to cite relevant images even when associated textual descriptors don't exactly match the user's query. Content creators must therefore ensure their images visually convey the message they want to communicate, beyond traditional textual optimization.

Spatial alt attributes help multimodal models

Alt attributes containing spatial descriptors help multimodal search systems match an image to the right product. This comes from models' ability to better anchor their visual understanding when associated text describes spatial relationships between elements. An alt attribute like "blue shirt with white vertical stripes worn with beige pants" provides multimodal models with spatial and relational information that reinforces their visual content interpretation. This practice differs from traditional SEO approaches that prioritized inserting target keywords in alt attributes without consideration for accurate visual content description.

Writing alt attributes optimized for multimodal systems follows specific principles. Descriptions should start with the main element of the image, then detail secondary elements and their spatial relationships. Including colors, textures, relative positions, and usage contexts enriches model understanding. Generic alt attributes like "product image" or "illustration photo" provide no value to multimodal systems and may even reduce citation chances. High-performing editorial teams develop style guides for alt attribute writing that integrate these multimodal requirements while maintaining accessibility for screen reader users.

Quality and resolution directly impact AI citations

Heavily compressed images with lossy artifacts are harder for models to analyze. This degradation is explained by the visual tokenization process. When an image has compression artifacts, resulting visual tokens become noisy and less reliable. The model may then misinterpret these tokens, generating hallucinations where AI describes objects or text that don't actually exist in the image.

The balance between image quality and loading performance remains a technical challenge. Modern formats like WebP and AVIF achieve higher compression rates while preserving the visual quality necessary for multimodal models. The recommended strategy is to maintain a minimum resolution of 1200x1200 pixels for main images with WebP format at quality 85 or higher. Secondary images can go down to 800x800 pixels but should not go below. Using lazy loading and responsive formats maintains loading performance without compromising image quality served to AI engine crawlers. Modern CDNs offer automatic image transformations that optimize quality based on request context.

Metadata enriches context for AI

EXIF metadata and captions provide AI engines with additional contextual information that reinforces their understanding of visual content according to Passionfruit recommendations. Geolocation data, capture date, device settings, and copyright information constitute context signals that multimodal models can leverage. A landscape image with GPS coordinates corresponding to a known tourist location will have better chances of being cited for queries related to that destination. Visible captions on the web page provide immediate textual context that models can directly associate with visual content during their analysis.

Image metadata enrichment follows a systematic process. IPTC information like title, description, keywords, and credits should be completed for each important image. HTML captions using the figcaption element associated with figure provide explicit semantic context to crawlers. Schema.org ImageObject structured data allows formally declaring image properties including the subject represented (about), creator, creation date (dateCreated), license, and identifiable entities (mentions). This metadata layer structures the information that multimodal models extract and integrate into their overall content understanding. E-commerce sites particularly benefit from these enrichments that connect product images to cataloged entities.

Semantic file naming influences indexing

Image file names constitute an indexing signal that search engine crawlers analyze before even loading the visual content according to Passionfruit. A file named "repotting-fiddle-leaf-fig.jpg" immediately communicates the image subject to indexing systems, while "IMG_3476.jpg" provides no exploitable information. This semantic naming practice is part of a global approach where every element associated with the image reinforces the signal sent to search engines. The file name, URL, alt attribute, caption, and surrounding textual context should form a coherent whole that confirms the image subject.

Optimal naming conventions use descriptive keywords separated by hyphens, in lowercase, without special characters or accents. The recommended structure places the main subject first, followed by qualifiers and context. For a product image, the format "category-product-feature-color.jpg" allows precise indexing while remaining readable. Sites with large image catalogs should implement automated naming rules during upload to ensure consistency. Modern CMS platforms offer automatic renaming features based on metadata or alt attributes that facilitate this standardization.

Image structured data reinforces understanding

Implementing Schema.org structured data for images provides search engines with a formal framework for interpreting visual content according to Search Engine Land. The ImageObject type allows explicitly declaring image properties including the subject represented (about), creator, creation date (dateCreated), license, and identifiable entities (mentions). This explicit structuring eliminates potential ambiguity from automatic interpretation and guides multimodal models toward correct understanding. E-commerce sites use the Product type with enriched image properties to connect visuals to structured product listings.

Technical implementation of image structured data is done via JSON-LD integrated in the page head or via Microdata itemprop attributes. JSON-LD format offers better readability and simplified maintenance for technical teams. Essential properties to complete include @type, contentUrl, caption, description, author, and datePublished. For images representing products, association with Product markup via the image property allows engines to understand the relationship between the visual and product entity. Editorial images benefit from about and mentions properties that connect them to Knowledge Graph entities. This connection to structured entities significantly increases citation chances in AI responses that rely on knowledge graphs.

Schema.org propertyDescriptionGEO impact
contentUrlDirect image URLCorrect indexing
captionDescriptive captionTextual context
aboutEntity representedKnowledge Graph connection
author/creatorImage creatorE-E-A-T signal
licenseUsage licenseSource reliability

The impact of visual search on AI citations

Google Lens with its 20 billion monthly visual searches represents a massive discovery channel that GEO strategies cannot ignore according to Lumar. Twenty percent of these searches concern shopping, creating a direct commercial visibility opportunity. Users photograph products they want to identify, compare, or purchase. Catalog images optimized for multimodal understanding have significantly better chances of appearing in these visual search results. Presence in Google Lens results generates qualified visits from users in active product research phase.

Projections indicate that AI Overviews will approach 70 to 80% of informational queries by end of 2026 according to Goodie. In this context, appearing as a cited source in AI Overviews becomes a new visibility benchmark. Correctly optimized images participate in this visibility by being referenced in generated visual responses. AI engines increasingly cite visual sources in addition to textual sources, particularly for product queries, tutorials, and explanatory content. Brands investing in multimodal optimization now are gaining competitive advantage on a rapidly expanding visibility channel.

Practical recommendations for multimodal image optimization

Optimizing images for multimodal systems requires a rigorous technical approach combined with editorial attention to descriptions. Teams should audit their existing image catalog to identify strategic visuals requiring priority optimization. Flagship product images, high search potential editorial visuals, and structural content illustrations constitute priorities. Each priority image should be evaluated according to minimum resolution criteria, compression quality, presence and quality of alt attributes, EXIF metadata, file naming, and associated structured data.

The image optimization workflow integrates these checks at each creation and publication stage. Photographers and graphic designers receive resolution and format guidelines. Writers have alt attribute templates including spatial information. Developers implement ImageObject structured data systematically. Publishing tools automatically verify image compliance before going live. This systematization ensures each new image meets multimodal system requirements from publication.

Sources

  1. 1.Lumar - Multimodal Search in 2025: Image, Video, & Voice Search
  2. 2.Passionfruit - Multimodal AI Search Optimization: Text, Image & Video Strategies
  3. 3.Search Engine Land - Image SEO for multimodal AI
  4. 4.Goodie - Future of AI Search: Citation Power & Visibility 2026
  5. 5.Semrush - AI Search Trends for 2026 & How You Can Adapt

FAQ

Frequently asked questions

What minimum resolution is recommended for AI-optimized images?

The recommended resolution for main images is minimum 1200x1200 pixels, with an absolute threshold of 800x800 pixels for secondary images.

How to write an alt attribute optimized for multimodal systems?

Optimized alt attributes include spatial descriptors describing relationships between visual elements. This approach helps improve matching accuracy. Describe the main element, then secondary elements and their relative positions.

What image format to prioritize for AI models?

WebP and AVIF formats offer the best balance between quality and compression. WebP quality of 85 or higher preserves details necessary for multimodal models. Images compressed with visible lossy artifacts reduce model confidence.

Does EXIF metadata influence AI citations?

EXIF metadata like geolocation, date, and creator information provide additional context to multimodal models. This information reinforces contextual understanding and may favor citations for geolocated or temporal queries.

How to implement structured data for images?

The Schema.org ImageObject type in JSON-LD format allows declaring essential properties. Include contentUrl, caption, description, author, and datePublished. For product images, associate them with Product markup via the image property.

Does image file naming impact AI indexing?

Semantic file names like "repotting-fiddle-leaf-fig.jpg" communicate the subject to crawlers before content loads. Use descriptive keywords separated by hyphens, in lowercase, without special characters.

How does Google Lens influence GEO strategy?

Google Lens processes billions of monthly visual searches, a notable share concerning shopping. Catalog images optimized for multimodal understanding appear more frequently in these visual search results.

Do HTML captions improve image citations?

Captions using the figcaption element associated with figure provide explicit semantic context to crawlers. This direct text-image association reinforces multimodal model understanding.

What audit frequency is recommended for image catalogs?

A quarterly audit of strategic images identifies visuals requiring optimization. Rapid evolution of AI engine multimodal capabilities justifies regular monitoring and practice updates.

Are AI-generated images well indexed by multimodal models?

AI-generated images are processed the same way as photographs by Visual Transformers. Image quality, resolution, and associated metadata determine citation potential, regardless of origin.

Read next

Raise your chances of being cited by ChatGPT and Google.

Free. Then a 3-day trial. Then Starter at €29 a month excl. VAT.

Start my 3 days, no card needed