The mid-September 2026 search landscape delivered another significant algorithmic adjustment. On September 10, Google updated its documentation on “Multi-Modal Intent Resolution” within AI Overviews, confirming a major architectural shift. Search models now process visual, text, and contextual location signals simultaneously in a single retrieval layer for complex queries. This update coincided with sharp, unconfirmed ranking fluctuations between September 11 and 13, heavily impacting local e-commerce, multi-location service brands, and technical SaaS hubs.
Â
Traditional search treated images, text, and local signals as separate algorithmic silos. Multi-modal retrieval evaluates them as a single combined entity vector. Multimodal search accepts combinations of text, image, voice, or video to return answers blending media. This post breaks down how Google’s mid-September update merges these signals into a unified retrieval layer and provides a blueprint for optimizing your visual assets alongside structured markup to capture multi-modal AI Overviews.
Â
Deconstructing Multi-Modal Intent Resolution
Vision-language models, specifically Gemini 3.5, now analyze image content alongside surrounding HTML text to determine entity relevance. By leveraging Gemini’s multimodal capabilities, AI Mode can understand an entire scene in an image, including the unique materials, colors, shapes, and relationships of objects within it.
Â
This creates a critical visual-text disconnect for sites relying on generic stock photos. Google Lens precisely identifies objects and uses query fan-out to issue multiple deep searches about an image. A mismatched stock photo actively dilutes a page’s topical authority score in AI Overviews. Cross-entity mapping is now standard. Google expects the visual attributes of a product schematic or interface screenshot to perfectly align with the JSON-LD ImageObject and ItemPage properties in the code.
Â
Advanced Strategy: Visual-Semantic Optimization (VSO)
To capture this new multi-modal search real estate, you must implement a 3-step Visual-Semantic Optimization (VSO) strategy:
- Descriptive Contextual Embedding: Do not orphan your images. Place unique first-party diagrams directly next to relevant H2 or H3 text blocks containing explicit entity definitions. The physical proximity of the text to the image in the Document Object Model (DOM) reinforces the multi-modal vector.
- Advanced ImageObject Schema Nesting: Stop using standalone image tags in your schema. Link your images directly to primary page entities by using the about and mentions properties within nested ImageObject JSON-LD. This explicitly maps the visual asset to the semantic topic.
- Vector-Friendly Image Assets: Replace decorative graphics with high-information visual assets. Data charts, annotated technical diagrams, and real-world proof photos provide Large Language Models (LLMs) with actionable information they can parse, verify, and cite.
Â
Mid-September Diagnostic Protocol
If your domain experienced turbulence between September 11 and 13, run a multi-modal diagnostic check:
- Visual Search Audit: Use Google Lens and Search Console’s Generative AI performance reports to identify which of your pages have visual assets being extracted into dynamic AI carousels. Look for impression spikes on image-heavy pages.
- The Multi-Modal SERP Test: Run your core brand and commercial queries across both text and image search. Verify that your visual entity citations (the images Google associates with your brand) match your textual landing page rankings.
- Correcting Mismatched Signals: Identify legacy stock imagery or outdated screenshots that create conflicting topical signals and replace them with verified, first-party visuals.
Â
Execution Plan
Search optimization is no longer a text-exclusive discipline. Visual and semantic signals must operate as a single, cohesive unit to survive in an AI-first retrieval environment.
Â
Immediate Action Checklist:
- Audit your top 20 high-value pages and replace generic stock images with first-party visual data.
- Implement deeply nested ImageObject schema across key product and informational templates.
- Review GSC performance data to isolate visual query impression changes following the September 10 update.
Expert Value-Add Tip
Image filename and surrounding DOM structure now carry more weight in multi-modal retrieval than standard alt text alone. Ensure image file names follow strict descriptive conventions (e.g., enterprise-saas-architecture-diagram-2026.webp instead of IMG_4092.webp) and are placed within the same parent <figure> or <div> container as the corresponding textual explanation.